| { |
| "schema_version": "run-index-v1", |
| "historical_scored_before": "2026-01-01T00:00:00Z", |
| "policy": "Unindexed shards at or after the cutoff are experimental. Explicit entries override the historical default. Only indexed shards are scored.", |
| "runs": [ |
| {"path":"data/run-rescue-older-2026-03-23-and-2026-03-30.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-rescue-sweep-results-2026-04-11.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-04-19T20-59-19-b11881ba-reasoning-gemini-25-flash.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-04-19T20-59-19-b11881ba-reasoning-gemini-3-flash.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-04-19T20-59-19-b11881ba-reasoning-gpt-5-mini.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-04-19T20-59-19-b11881ba-reasoning-gpt-54-mini.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-04-19T20-59-19-b11881ba-reasoning-qwen3-coder-30b-4bit.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-03T00-34-49-3f73990-throughput-mlx-community-qwen3-coder-30b-a3b-instruct-4bit-af05f318.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-03T00-37-49-3f73990-throughput-mlx-community-qwen3-coder-30b-a3b-instruct-4bit-5d5ffd21.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-03T00-43-38-3f73990-throughput-mlx-community-qwen3-coder-30b-a3b-instruct-4bit-ae951779.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-03T00-49-34-3f73990-throughput-mlx-community-gpt-oss-120b-mxfp4-q8-6828d4cb.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-03T00-53-59-3f73990-throughput-mlx-community-gpt-oss-120b-mxfp4-q8-1327c588.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-03T00-56-59-3f73990-throughput-mlx-community-gpt-oss-120b-mxfp4-q8-080a50fb.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-08T08-46-40-99eb1ab-math-hard-mlx-community-qwen3-coder-30b-a3b-instruct-4bit-61e0336b.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-08T12-55-14-99eb1ab-reasoning-gemini-gemini-2-5-flash-70ccbeb7.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-08T15-09-30-35fa038-tool-calling-mlx-community-qwen3-6-35b-a3b-optiq-4bit-aa6ba94f.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-08T15-39-15-35fa038-tool-calling-mlx-community-qwen3-6-35b-a3b-4bit-512c7e6e.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-08T15-43-57-35fa038-tool-calling-mlx-community-glm-4-7-flash-4bit-a9809ab4.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-08T16-02-06-35fa038-tool-calling-mlx-community-qwen3-6-35b-a3b-8bit-8a25d0b6.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-08T16-20-19-35fa038-tool-calling-lmstudio-community-qwen3-6-35b-a3b-mlx-8bit-ea45b51e.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-08T17-14-59-35fa038-tool-calling-mlx-community-qwen3-next-80b-a3b-thinking-4bit-44914f4c.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-08T17-17-21-35fa038-tool-calling-mlx-community-qwen3-coder-30b-a3b-instruct-4bit-4fb7a7fb.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-08T17-39-45-35fa038-tool-calling-mlx-community-gpt-oss-120b-mxfp4-q8-fdecb5b7.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-09T03-18-50-8cf985a-tool-calling-openyourmind-qwopus3-5-122b-a10b-kimi-k2-6-destill-ba11c673.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-09T03-43-23-8cf985a-tool-calling-lmstudio-community-hermes-4-70b-mlx-4bit-f0a76c82.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-13T11-48-49-dd41038-tool-calling-mlx-community-qwen3-next-80b-a3b-thinking-4bit-c301d8cf.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-13T11-54-24-dd41038-tool-calling-mlx-community-qwen3-next-80b-a3b-thinking-4bit-a9f79c4c.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-13T12-00-45-dd41038-tool-calling-mlx-community-qwen3-next-80b-a3b-thinking-4bit-c9a9c418.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-13T12-10-13-dd41038-tool-calling-mlx-community-qwen3-next-80b-a3b-thinking-4bit-4ff835f7.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-17T22-34-07-a48eb3b-tool-calling-mlx-community-gpt-oss-120b-mxfp4-q8-3b6fa543.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-17T22-34-07-a48eb3b-tool-calling-mlx-community-gpt-oss-120b-mxfp4-q8-685c3d56.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-17T23-46-13-a48eb3b-tool-calling-mlx-community-qwen3-next-80b-a3b-instruct-4bit-105bb937.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-17T23-46-13-a48eb3b-tool-calling-mlx-community-qwen3-next-80b-a3b-instruct-4bit-2652dcd3.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-18T00-31-43-a48eb3b-tool-calling-mlx-community-qwen3-next-80b-a3b-thinking-4bit-15c6cf50.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-18T00-31-43-a48eb3b-tool-calling-mlx-community-qwen3-next-80b-a3b-thinking-4bit-48a505a7.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-18T00-31-46-a48eb3b-tool-calling-mlx-community-gemma-4-31b-it-optiq-4bit-3ec5a98d.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-18T00-31-46-a48eb3b-tool-calling-mlx-community-gemma-4-31b-it-optiq-4bit-7aa020ca.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-19T00-41-19-565049c-throughput-mlx-community-qwen3-coder-30b-a3b-instruct-4bit-a54ee076.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-19T00-41-27-565049c-throughput-mlx-community-qwen3-next-80b-a3b-instruct-4bit-4b2f7c4a.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-19T00-51-45-94c0407-throughput-mlx-community-qwen3-6-35b-a3b-4bit-1f0b18ee.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-19T19-18-51-f73629e-throughput-mlx-community-qwen3-6-35b-a3b-optiq-4bit-80519dd5.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-19T19-26-27-f73629e-throughput-mlx-community-qwen3-6-35b-a3b-4bit-db8be763.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-26T16-16-21-f73629e-throughput-mlx-community-glm-4-7-reap-50-mxfp4-8c4b79ee.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-26T16-16-21-f73629e-throughput-mlx-community-qwen3-coder-30b-a3b-instruct-4bit-87cc436a.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-26T16-16-21-f73629e-throughput-mlx-community-qwen3-coder-30b-a3b-instruct-4bit-a342f021.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-27T06-51-50-f73629e-throughput-mlx-community-qwen3-next-80b-a3b-thinking-4bit-af07d295.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-07-27T06-51-58-f73629e-throughput-mlx-community-gpt-oss-120b-mxfp4-q8-b0b878c7.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-16T06-07-13-0129bab-reasoning-mlx-community-qwen3-8-27b-4bit-ecfc3471.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-16T06-20-57-0129bab-reasoning-mlx-community-qwen3-8-27b-4bit-8d9010ad.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-16T06-44-26-0129bab-reasoning-mlx-community-qwen3-8-27b-8bit-9f0fb689.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-16T07-07-38-0129bab-reasoning-mlx-community-qwen3-8-27b-8bit-c4299ff9.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-16T07-32-00-0129bab-math-hard-mlx-community-qwen3-8-27b-4bit-2f8c6edb.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-16T08-58-55-0129bab-math-hard-mlx-community-qwen3-8-27b-4bit-1c9422a7.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-23T22-33-18-73c7054-tool-calling-mlx-community-qwen3-4b-instruct-2507-4bit-9a95b66a.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-23T22-37-16-73c7054-tool-calling-mlx-community-qwen3-5-9b-mlx-4bit-721eebd9.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-23T22-40-12-73c7054-reasoning-mlx-community-qwen3-4b-instruct-2507-4bit-0191fce3.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-23T22-40-45-73c7054-reasoning-mlx-community-qwen3-5-9b-mlx-4bit-8b6c5c1e.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-29T01-04-16-b80570a-throughput-mlx-community-qwen3-8-27b-4bit-c4142e15.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-29T01-24-35-b80570a-throughput-mlx-community-qwen3-5-9b-mlx-4bit-ed88d0c1.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-29T01-29-11-b80570a-throughput-mlx-community-qwen3-6-35b-a3b-4bit-c21c9caf.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-29T01-32-13-b80570a-throughput-mlx-community-qwen3-6-35b-a3b-4bit-7b3677e7.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-29T01-40-46-b80570a-throughput-mlx-community-qwen3-8-27b-4bit-e78c7b3f.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-29T10-51-22-b80570a-math-hard-mlx-community-qwen3-5-9b-mlx-4bit-b429e658.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-29T10-52-57-b80570a-math-hard-mlx-community-qwen3-6-35b-a3b-4bit-40d31b6b.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-29T10-54-00-b80570a-math-hard-mlx-community-qwen3-8-27b-4bit-1fa0b6e6.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-29T11-10-32-5699697-throughput-mlx-community-qwen3-8-27b-4bit-5843a3df.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-29T11-24-57-5699697-throughput-mlx-community-qwen3-8-27b-4bit-e58ee844.parquet","status":"experimental","variant":"Unindexed historical run","comparison_group":"unclassified","context_band":"unspecified","max_output_tokens":null,"completion_state":"unknown","caveat":"Original run classification was not recorded."}, |
| {"path":"data/run-2026-08-24T04-19-37-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-dd4201dc.parquet","status":"experimental","variant":"MTP exploration","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T04-30-11-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-a9be338b.parquet","status":"experimental","variant":"MTP exploration","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T04-34-49-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-5fe7c88c.parquet","status":"experimental","variant":"MTP exploration","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T11-32-35-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-cd514ed8.parquet","status":"experimental","variant":"MTP exploration","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T11-33-46-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-0d5e747d.parquet","status":"experimental","variant":"MTP exploration","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T11-41-57-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-8220887a.parquet","status":"experimental","variant":"MTP default","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T11-42-51-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-e600135b.parquet","status":"experimental","variant":"MTP default","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T11-44-55-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-be59c6b6.parquet","status":"experimental","variant":"baseline","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T11-46-39-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-2b397c93.parquet","status":"experimental","variant":"baseline","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T11-49-22-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-f29a6154.parquet","status":"experimental","variant":"MTP block size 1","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T11-50-28-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-3c2cac66.parquet","status":"experimental","variant":"MTP block size 3","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T11-51-42-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-562a1954.parquet","status":"experimental","variant":"MTP block size 3","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T11-54-28-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-b39c41a5.parquet","status":"experimental","variant":"MTP 8-bit","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T11-56-07-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-12c145f5.parquet","status":"experimental","variant":"MTP 8-bit","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T11-59-03-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-f8c65ec1.parquet","status":"experimental","variant":"MTP MXFP8","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T12-00-42-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-40556739.parquet","status":"experimental","variant":"MTP MXFP8","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T12-03-37-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-bd10debf.parquet","status":"experimental","variant":"MTP winner concurrency 2","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T12-04-56-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-b16f7660.parquet","status":"experimental","variant":"MTP winner concurrency 2","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T12-06-35-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-86d864dc.parquet","status":"experimental","variant":"MTP winner concurrency 4","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T12-08-33-7ef5a4d-throughput-mlx-community-qwen3-8-27b-4bit-7229ac53.parquet","status":"experimental","variant":"MTP winner concurrency 4","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"capped","caveat":"Capped microbenchmark; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T12-13-53-7ef5a4d-math-hard-mlx-community-qwen3-8-27b-4bit-f891e5c3.parquet","status":"experimental","variant":"MTP math smoke","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":1024,"completion_state":"sampled","caveat":"Small-sample experiment; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T12-45-29-7ef5a4d-math-hard-mlx-community-qwen3-8-27b-4bit-7eac3112.parquet","status":"experimental","variant":"MTP math pair","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":1024,"completion_state":"sampled","caveat":"Small-sample experiment; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T12-59-55-7ef5a4d-math-hard-mlx-community-qwen3-8-27b-4bit-d791721a.parquet","status":"experimental","variant":"MTP math pair","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":1024,"completion_state":"sampled","caveat":"Small-sample experiment; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T13-14-02-7ef5a4d-math-hard-mlx-community-qwen3-8-27b-4bit-8c4dd06f.parquet","status":"experimental","variant":"baseline math pair","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":1024,"completion_state":"sampled","caveat":"Small-sample experiment; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T13-34-11-7ef5a4d-math-hard-mlx-community-qwen3-8-27b-4bit-3571feab.parquet","status":"experimental","variant":"baseline math pair","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":1024,"completion_state":"sampled","caveat":"Small-sample experiment; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T14-32-37-7ef5a4d-tool-calling-mlx-community-qwen3-8-27b-4bit-8cd7c4d8.parquet","status":"experimental","variant":"MTP agentic pair","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"sampled","caveat":"Small-sample experiment; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T18-29-03-7ef5a4d-tool-calling-mlx-community-qwen3-8-27b-4bit-00d1605a.parquet","status":"recovered","variant":"MTP agentic partial","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"partial","caveat":"Recovered partial output; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T18-29-03-7ef5a4d-tool-calling-mlx-community-qwen3-8-27b-4bit-12c1f86c.parquet","status":"experimental","variant":"MTP agentic thinking off","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"sampled","caveat":"Small-sample experiment; excluded from rankings."}, |
| {"path":"data/run-2026-08-24T18-47-14-7ef5a4d-tool-calling-mlx-community-qwen3-8-27b-4bit-7c074988.parquet","status":"experimental","variant":"MTP agentic timeout smoke","comparison_group":"qwen3.8-mtp-2026-08","context_band":"short-prompt","max_output_tokens":256,"completion_state":"timeout","caveat":"Timeout smoke; excluded from rankings."}, |
| {"path":"data/run-2026-08-27T15-18-41-249cf89e6f8936ef146afa2d4119a29c82866288-throughput-mlx-community-qwen3-8-27b-4bit-2ca83011.parquet","status":"experimental","variant":"MTP upstream e82d557","comparison_group":"qwen3.8-mtp-long-context-2026-08","context_band":"64k","max_output_tokens":512,"completion_state":"complete","caveat":"Managed window began under swap pressure; no prompt-variant cache busting, but each measured request reported zero cached tokens."}, |
| {"path":"data/run-2026-08-27T15-42-44-249cf89e6f8936ef146afa2d4119a29c82866288-throughput-mlx-community-qwen3-8-27b-4bit-cb9bc797.parquet","status":"experimental","variant":"MTP upstream e82d557","comparison_group":"qwen3.8-mtp-long-context-2026-08","context_band":"128k","max_output_tokens":512,"completion_state":"complete","caveat":"Managed window began under swap pressure; no prompt-variant cache busting, but each measured request reported zero cached tokens."}, |
| {"path":"data/run-2026-08-30T20-06-14-8cd852a4a06770510b3daf3fcab991a510510671-throughput-mlx-community-qwen3-8-27b-4bit-ef9e8438.parquet","status":"experimental","variant":"base c1 clean reset","comparison_group":"qwen3.8-clean-reset-128k-2026-08","context_band":"128k","max_output_tokens":64,"completion_state":"complete","caveat":"Four resident samples after a cold readiness request; isolated zero-swap clean-reset evidence."}, |
| {"path":"data/run-2026-08-30T21-26-15-8cd852a4a06770510b3daf3fcab991a510510671-throughput-mlx-community-qwen3-8-27b-4bit-149aa815.parquet","status":"experimental","variant":"native MTP c1 clean reset","comparison_group":"qwen3.8-clean-reset-128k-2026-08","context_band":"128k","max_output_tokens":64,"completion_state":"complete","caveat":"Four resident samples after a cold readiness request; isolated zero-swap clean-reset evidence."}, |
| {"path":"data/run-2026-08-30T22-32-45-8cd852a4a06770510b3daf3fcab991a510510671-throughput-mlx-community-qwen3-8-27b-4bit-2e31a861.parquet","status":"experimental","variant":"native MTP c2 clean reset","comparison_group":"qwen3.8-clean-reset-128k-2026-08","context_band":"128k","max_output_tokens":64,"completion_state":"complete","caveat":"Four resident samples plus one successful two-request probe; aggregate throughput remained at the c1 rate without swap."}, |
| {"path":"data/run-2026-09-07T02-23-30-789655f9-coding-mlx-community-qwen3-6-35b-a3b-4bit-cc816416.parquet","status":"invalid","variant":"legacy coding; task identities unavailable · reasoning unspecified · unreported","comparison_group":"legacy-coding-unresolved-20260907T022330","context_band":"unreported","max_output_tokens":null,"completion_state":"all published pass_at_1 item rows present; benchmark task identity unresolved","caveat":"two-task-resolution; task-set-saturated-two-task-resolution; shared-endpoint-not-a-dedicated-measurement; raw item ids absent","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-07T13-36-52-789655f9-coding-mlx-community-kimi-linear-48b-a3b-instruct-6bit-60ad661d.parquet","status":"invalid","variant":"legacy coding; task identities unavailable · reasoning unspecified · unreported","comparison_group":"legacy-coding-unresolved-20260907T133652","context_band":"unreported","max_output_tokens":null,"completion_state":"all published pass_at_1 item rows present; benchmark task identity unresolved","caveat":"two-task-resolution; task-set-saturated-two-task-resolution; raw item ids absent","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-07T14-55-25-789655f9-coding-mlx-community-kimi-linear-48b-a3b-instruct-6bit-ee8baffa.parquet","status":"invalid","variant":"legacy coding; task identities unavailable · reasoning unspecified · unreported","comparison_group":"legacy-coding-unresolved-20260907T145525","context_band":"unreported","max_output_tokens":null,"completion_state":"all published pass_at_1 item rows present; benchmark task identity unresolved","caveat":"two-task-resolution; task-set-saturated-two-task-resolution; raw item ids absent","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-07T16-03-06-789655f9-coding-mlx-community-kimi-linear-48b-a3b-instruct-6bit-eee0ebce.parquet","status":"invalid","variant":"legacy coding; task identities unavailable · reasoning unspecified · unreported","comparison_group":"legacy-coding-unresolved-20260907T160306","context_band":"unreported","max_output_tokens":null,"completion_state":"all published pass_at_1 item rows present; benchmark task identity unresolved","caveat":"two-task-resolution; task-set-saturated-two-task-resolution; raw item ids absent","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-07T17-08-27-789655f9-coding-mlx-community-qwen3-6-35b-a3b-4bit-7945cf8e.parquet","status":"invalid","variant":"legacy coding; task identities unavailable · reasoning unspecified · unreported","comparison_group":"legacy-coding-unresolved-20260907T170827","context_band":"unreported","max_output_tokens":null,"completion_state":"all published pass_at_1 item rows present; benchmark task identity unresolved","caveat":"two-task-resolution; task-set-saturated-two-task-resolution; shared-endpoint-terminated-early-partial-run; raw item ids absent","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-09T13-39-37-789655f9-coding-mlx-community-qwen3-6-35b-a3b-4bit-9758dc59.parquet","status":"invalid","variant":"legacy coding; task identities unavailable · reasoning unspecified · unreported","comparison_group":"legacy-coding-unresolved-20260909T133937","context_band":"unreported","max_output_tokens":null,"completion_state":"all published pass_at_1 item rows present; benchmark task identity unresolved","caveat":"two-task-resolution; task-set-saturated-two-task-resolution; raw item ids absent","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-09T14-31-38-789655f9-coding-mlx-community-qwen3-6-35b-a3b-4bit-784a13df.parquet","status":"invalid","variant":"legacy coding; task identities unavailable · reasoning unspecified · unreported","comparison_group":"legacy-coding-unresolved-20260909T143138","context_band":"unreported","max_output_tokens":null,"completion_state":"all published pass_at_1 item rows present; benchmark task identity unresolved","caveat":"two-task-resolution; task-set-saturated-two-task-resolution; raw item ids absent","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-09T16-49-38-789655f9-coding-mlx-community-qwen3-6-35b-a3b-4bit-649146e2.parquet","status":"invalid","variant":"legacy coding; task identities unavailable · reasoning unspecified · unreported","comparison_group":"legacy-coding-unresolved-20260909T164938","context_band":"unreported","max_output_tokens":null,"completion_state":"all published pass_at_1 item rows present; benchmark task identity unresolved","caveat":"two-task-resolution; task-set-saturated-two-task-resolution; raw item ids absent","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-09T17-46-21-789655f9-coding-mlx-community-qwen3-6-35b-a3b-4bit-8bfce38e.parquet","status":"invalid","variant":"legacy coding; task identities unavailable · reasoning unspecified · unreported","comparison_group":"legacy-coding-unresolved-20260909T174621","context_band":"unreported","max_output_tokens":null,"completion_state":"all published pass_at_1 item rows present; benchmark task identity unresolved","caveat":"two-task-resolution; task-set-saturated-two-task-resolution; raw item ids absent","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-09T17-48-26-789655f9-coding-mlx-community-qwen3-6-35b-a3b-4bit-0ada45cb.parquet","status":"invalid","variant":"legacy coding; task identities unavailable · reasoning unspecified · unreported","comparison_group":"legacy-coding-unresolved-20260909T174826","context_band":"unreported","max_output_tokens":null,"completion_state":"all published pass_at_1 item rows present; benchmark task identity unresolved","caveat":"two-task-resolution; task-set-saturated-two-task-resolution; raw item ids absent","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-09T22-09-24-ca3ae9c4605eb44450fbe1c579409ff1c18d1aa2-throughput-mlx-community-qwen3-6-35b-a3b-4bit-aa34c1ed.parquet","status":"invalid","variant":"engine unreported profile-unreported unknown","comparison_group":"throughput-unassigned","context_band":"n/a","max_output_tokens":null,"completion_state":"unverified","caveat":"no eligible raw source in supplied directories","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-09T22-10-15-ca3ae9c4605eb44450fbe1c579409ff1c18d1aa2-throughput-mlx-community-qwen3-6-35b-a3b-4bit-a292eb7d.parquet","status":"invalid","variant":"engine unreported profile-unreported unknown","comparison_group":"throughput-unassigned","context_band":"n/a","max_output_tokens":null,"completion_state":"unverified","caveat":"no eligible raw source in supplied directories","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-09T22-11-07-ca3ae9c4605eb44450fbe1c579409ff1c18d1aa2-throughput-mlx-community-qwen3-6-35b-a3b-4bit-918de949.parquet","status":"invalid","variant":"engine unreported profile-unreported unknown","comparison_group":"throughput-unassigned","context_band":"n/a","max_output_tokens":null,"completion_state":"unverified","caveat":"no eligible raw source in supplied directories","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-09T22-11-59-ca3ae9c4605eb44450fbe1c579409ff1c18d1aa2-throughput-mlx-community-qwen3-6-35b-a3b-4bit-84d688b6.parquet","status":"invalid","variant":"engine unreported profile-unreported unknown","comparison_group":"throughput-unassigned","context_band":"n/a","max_output_tokens":null,"completion_state":"unverified","caveat":"no eligible raw source in supplied directories","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-09T22-12-58-ca3ae9c4605eb44450fbe1c579409ff1c18d1aa2-throughput-mlx-community-qwen3-6-35b-a3b-4bit-fb83d358.parquet","status":"invalid","variant":"engine unreported profile-unreported unknown","comparison_group":"throughput-unassigned","context_band":"n/a","max_output_tokens":null,"completion_state":"unverified","caveat":"no eligible raw source in supplied directories","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-09T22-13-58-ca3ae9c4605eb44450fbe1c579409ff1c18d1aa2-throughput-mlx-community-qwen3-6-35b-a3b-4bit-df1bddc0.parquet","status":"invalid","variant":"engine unreported profile-unreported unknown","comparison_group":"throughput-unassigned","context_band":"n/a","max_output_tokens":null,"completion_state":"unverified","caveat":"no eligible raw source in supplied directories","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-09T23-48-38-2e0e9d1-throughput-mlx-community-nvidia-nemotron-3-super-120b-a12b-4b-68c168c5.parquet","status":"invalid","variant":"engine unreported profile-unreported unknown","comparison_group":"throughput-unassigned","context_band":"n/a","max_output_tokens":null,"completion_state":"unverified","caveat":"no eligible raw source in supplied directories","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-09-09T23-53-15-2e0e9d1-throughput-mlx-community-nvidia-nemotron-3-super-120b-a12b-4b-10a83319.parquet","status":"invalid","variant":"engine unreported profile-unreported unknown","comparison_group":"throughput-unassigned","context_band":"n/a","max_output_tokens":null,"completion_state":"unverified","caveat":"no eligible raw source in supplied directories","machine":"Mac Studio M4 Max 128GB"}, |
| {"path":"data/run-2026-10-07T00-21-59-5465c0c-throughput-unsloth-qwen3-8-27b-gguf-c4224986.parquet","status":"invalid","variant":"llama.cpp b11457 64k-context unreported","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"unverified","caveat":"server summary only; no raw result metrics source","machine":"RTX PRO 6000 Max-Q 96GB"}, |
| {"path":"data/run-2026-10-07T00-25-08-5465c0c-throughput-unsloth-qwen3-8-27b-gguf-adad7a1e.parquet","status":"invalid","variant":"llama.cpp b11457 64k-context unreported","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"unverified","caveat":"server summary only; no raw result metrics source","machine":"RTX PRO 6000 Max-Q 96GB"}, |
| {"path":"data/run-2026-10-07T00-27-09-5465c0c-throughput-unsloth-qwen3-8-27b-gguf-1c8eb49e.parquet","status":"invalid","variant":"llama.cpp b11457 64k-context unreported","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"unverified","caveat":"server summary only; no raw result metrics source","machine":"RTX PRO 6000 Max-Q 96GB"}, |
| {"path":"data/run-2026-10-07T00-41-31-5465c0c-throughput-unsloth-qwen3-8-27b-gguf-1cd01acb.parquet","status":"invalid","variant":"llama.cpp b11457 64k-context unreported","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"unverified","caveat":"server summary only; no raw result metrics source","machine":"RTX PRO 6000 Max-Q 96GB"}, |
| {"path":"data/run-2026-10-07T00-48-41-5465c0c-throughput-unsloth-qwen3-8-27b-gguf-8fc003ef.parquet","status":"invalid","variant":"llama.cpp b11457 64k-context unreported","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"unverified","caveat":"server summary only; no raw result metrics source","machine":"RTX PRO 6000 Max-Q 96GB"}, |
| {"path":"data/run-2026-10-16T17-36-43-4649920-throughput-unsloth-qwen3-8-27b-gguf-ee738b4a.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T17-37-12-4649920-throughput-unsloth-qwen3-8-27b-gguf-1893ba7b.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"sequential 3 ok/0 err; concurrent 4 ok/0 err","caveat":"prompt-cache hit inflation; prefill throughput >7k tok/s","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T17-37-45-4649920-throughput-unsloth-qwen3-8-27b-gguf-e54ec885.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"sequential 3 ok/0 err; concurrent 8 ok/0 err","caveat":"prompt-cache hit inflation; prefill throughput >7k tok/s","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T17-38-18-4649920-throughput-unsloth-qwen3-8-27b-gguf-1feb6c0e.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T17-38-58-4649920-throughput-unsloth-qwen3-8-27b-gguf-743ba41a.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"sequential 3 ok/0 err; concurrent 4 ok/0 err","caveat":"prompt-cache hit inflation; prefill throughput >7k tok/s","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T17-39-35-4649920-throughput-unsloth-qwen3-8-27b-gguf-0925d113.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"sequential 3 ok/0 err; concurrent 8 ok/0 err","caveat":"prompt-cache hit inflation; prefill throughput >7k tok/s","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T17-40-28-4649920-throughput-unsloth-qwen3-8-27b-gguf-bc51ddd2.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T17-41-16-4649920-throughput-unsloth-qwen3-8-27b-gguf-971545c9.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"sequential 3 ok/0 err; concurrent 4 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T17-42-06-4649920-throughput-unsloth-qwen3-8-27b-gguf-f3439bbf.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"sequential 3 ok/0 err; concurrent 8 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-08-43-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-825a24d1.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-09-28-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-7bda6629.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"sequential 3 ok/0 err; concurrent 4 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-10-18-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-bd2b8ca6.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"sequential 3 ok/0 err; concurrent 8 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-11-16-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-9f1033ef.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-11-39-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-67c6e4fe.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"sequential 3 ok/0 err; concurrent 4 ok/0 err","caveat":"prompt-cache hit inflation; prefill throughput >7k tok/s","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-12-00-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-b0b779d2.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"sequential 3 ok/0 err; concurrent 8 ok/0 err","caveat":"prompt-cache hit inflation; prefill throughput >7k tok/s","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-12-25-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-ded2fed3.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-12-51-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-5c11afa2.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"sequential 3 ok/0 err; concurrent 4 ok/0 err","caveat":"prompt-cache hit inflation; prefill throughput >7k tok/s","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-13-20-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-15cda1b9.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"sequential 3 ok/0 err; concurrent 8 ok/0 err","caveat":"prompt-cache hit inflation; prefill throughput >7k tok/s","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-33-40-4649920-throughput-mlx-community-qwen3-8-27b-8bit-04f747cd.parquet","status":"experimental","variant":"MLX 0.31.3 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-36-39-4649920-throughput-mlx-community-qwen3-8-27b-8bit-bbfde9ff.parquet","status":"experimental","variant":"MLX 0.31.3 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"sequential 3 ok/0 err; concurrent 4 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-39-39-4649920-throughput-mlx-community-qwen3-8-27b-8bit-d9cc0b5d.parquet","status":"experimental","variant":"MLX 0.31.3 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"sequential 3 ok/0 err; concurrent 8 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-43-10-4649920-throughput-mlx-community-qwen3-8-27b-8bit-f4376039.parquet","status":"experimental","variant":"MLX 0.31.3 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-46-05-4649920-throughput-mlx-community-qwen3-8-27b-8bit-6f32fd6d.parquet","status":"experimental","variant":"MLX 0.31.3 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"sequential 3 ok/0 err; concurrent 4 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-48-07-4649920-throughput-mlx-community-qwen3-8-27b-8bit-2fa93afa.parquet","status":"experimental","variant":"MLX 0.31.3 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"sequential 3 ok/0 err; concurrent 8 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-50-35-4649920-throughput-mlx-community-qwen3-8-27b-8bit-155fb939.parquet","status":"experimental","variant":"MLX 0.31.3 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-53-59-4649920-throughput-mlx-community-qwen3-8-27b-8bit-96a08b20.parquet","status":"experimental","variant":"MLX 0.31.3 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"sequential 3 ok/0 err; concurrent 4 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T20-58-06-4649920-throughput-mlx-community-qwen3-8-27b-8bit-b8a4640c.parquet","status":"experimental","variant":"MLX 0.31.3 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"sequential 3 ok/0 err; concurrent 8 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-05-29-4649920-throughput-unsloth-qwen3-8-27b-gguf-941de558.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"RTX 4080 SUPER 16GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-07-25-4649920-throughput-unsloth-qwen3-8-27b-gguf-cd3fe262.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/3 err","caveat":"concurrent HTTP/request errors=3; raw metrics retained","machine":"RTX 4080 SUPER 16GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-08-39-4649920-throughput-unsloth-qwen3-8-27b-gguf-436455cd.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/7 err","caveat":"concurrent HTTP/request errors=7; raw metrics retained","machine":"RTX 4080 SUPER 16GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-09-53-4649920-throughput-unsloth-qwen3-8-27b-gguf-31ec2b2b.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"RTX 4080 SUPER 16GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-11-08-4649920-throughput-unsloth-qwen3-8-27b-gguf-a3df3474.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/3 err","caveat":"concurrent HTTP/request errors=3; raw metrics retained","machine":"RTX 4080 SUPER 16GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-11-36-4649920-throughput-unsloth-qwen3-8-27b-gguf-fc99bc2f.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/7 err","caveat":"concurrent HTTP/request errors=7; raw metrics retained","machine":"RTX 4080 SUPER 16GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-12-04-4649920-throughput-unsloth-qwen3-8-27b-gguf-fae6d0e9.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/0 err","caveat":"valid one-off sequential+concurrent probe; not final comparison","machine":"RTX 4080 SUPER 16GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-12-53-4649920-throughput-unsloth-qwen3-8-27b-gguf-32d22754.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/3 err","caveat":"concurrent HTTP/request errors=3; raw metrics retained","machine":"RTX 4080 SUPER 16GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-13-35-4649920-throughput-unsloth-qwen3-8-27b-gguf-8cc4e5c3.parquet","status":"invalid","variant":"llama.cpp b11459 32k-context sequential+concurrent probe","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"sequential 3 ok/0 err; concurrent 1 ok/7 err","caveat":"concurrent HTTP/request errors=7; raw metrics retained","machine":"RTX 4080 SUPER 16GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-31-00-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-589e6e4d.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-32-03-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-178c3f46.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-33-36-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-107f1baa.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-35-58-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-b07f2e3c.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-39-19-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-00b8fed0.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-43-58-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-62c8d06e.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-45-25-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-1e686850.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-48-53-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-85c31b5f.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T21-52-24-4649920-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-9bfb0e5a.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T22-10-07-4649920-throughput-nvidia-qwen3-8-27b-nvfp4-4e87adc6.parquet","status":"invalid","variant":"vLLM 0.31.0 192k-context MTP-off batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"wrong vLLM profile; converge used medium-a without MTP","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.88}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T22-10-57-4649920-throughput-nvidia-qwen3-8-27b-nvfp4-f2975339.parquet","status":"invalid","variant":"vLLM 0.31.0 192k-context MTP-off batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"wrong vLLM profile; converge used medium-a without MTP","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.88}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T22-11-44-4649920-throughput-nvidia-qwen3-8-27b-nvfp4-bb181025.parquet","status":"invalid","variant":"vLLM 0.31.0 192k-context MTP-off batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"wrong vLLM profile; converge used medium-a without MTP","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.88}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T22-13-41-4649920-throughput-nvidia-qwen3-8-27b-nvfp4-51c7a200.parquet","status":"invalid","variant":"vLLM 0.31.0 192k-context MTP-off batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"wrong vLLM profile; converge used medium-a without MTP","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.88}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T22-15-51-4649920-throughput-nvidia-qwen3-8-27b-nvfp4-954bad38.parquet","status":"invalid","variant":"vLLM 0.31.0 192k-context MTP-off batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"wrong vLLM profile; converge used medium-a without MTP","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.88}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T22-18-56-4649920-throughput-nvidia-qwen3-8-27b-nvfp4-29212ed5.parquet","status":"invalid","variant":"vLLM 0.31.0 192k-context MTP-off batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"wrong vLLM profile; converge used medium-a without MTP","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.88}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T22-20-06-4649920-throughput-nvidia-qwen3-8-27b-nvfp4-98b87efe.parquet","status":"invalid","variant":"vLLM 0.31.0 192k-context MTP-off batched","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"wrong vLLM profile; converge used medium-a without MTP","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.88}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T22-21-27-4649920-throughput-nvidia-qwen3-8-27b-nvfp4-6d58a736.parquet","status":"invalid","variant":"vLLM 0.31.0 192k-context MTP-off batched","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"wrong vLLM profile; converge used medium-a without MTP","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.88}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T22-22-50-4649920-throughput-nvidia-qwen3-8-27b-nvfp4-bd6740b6.parquet","status":"invalid","variant":"vLLM 0.31.0 192k-context MTP-off batched","comparison_group":"throughput-unassigned","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"wrong vLLM profile; converge used medium-a without MTP","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.88}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T22-29-26-4649920-knowledge-nvidia-qwen3-8-27b-nvfp4-33440b7a.parquet","status":"scored","variant":"MMLU-Pro; stratified 5-shot loglik · reasoning off · 196608","comparison_group":"mmlu-pro-stratified-420-five-shot","context_band":"196608","max_output_tokens":null,"completion_state":"complete","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.88}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T23-13-48-4649920-throughput-unsloth-qwen3-8-27b-gguf-4b501908.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T23-14-41-4649920-throughput-unsloth-qwen3-8-27b-gguf-8e107431.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T23-15-59-4649920-throughput-unsloth-qwen3-8-27b-gguf-3f860dde.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T23-18-13-4649920-throughput-unsloth-qwen3-8-27b-gguf-a511e1c9.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T23-20-57-4649920-throughput-unsloth-qwen3-8-27b-gguf-815c49db.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T23-24-28-4649920-throughput-unsloth-qwen3-8-27b-gguf-857b333d.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T23-25-44-4649920-throughput-unsloth-qwen3-8-27b-gguf-ad691520.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T23-28-34-4649920-throughput-unsloth-qwen3-8-27b-gguf-3a92fcc6.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T23-31-01-4649920-throughput-unsloth-qwen3-8-27b-gguf-2302806c.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T23-39-59-4649920-throughput-unsloth-qwen3-8-27b-gguf-8063c12e.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context Poisson 0.5 rps","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T23-43-48-4649920-throughput-unsloth-qwen3-8-27b-gguf-1c0abb55.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context Poisson 1 rps","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T23-47-35-4649920-throughput-unsloth-qwen3-8-27b-gguf-44e7465d.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context Poisson 2 rps","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-16T23-51-25-4649920-throughput-unsloth-qwen3-8-27b-gguf-baf5f416.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T00-02-16-4649920-throughput-unsloth-qwen3-8-27b-gguf-17c170d5.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context Poisson 0.03 rps","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T00-07-58-4649920-throughput-unsloth-qwen3-8-27b-gguf-ef4c615a.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context Poisson 0.06 rps","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T00-11-18-4649920-throughput-unsloth-qwen3-8-27b-gguf-109fe150.parquet","status":"experimental","variant":"llama.cpp b11459 32k-context Poisson 0.12 rps","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T01-27-34-4649920-throughput-mlx-community-qwen3-8-27b-8bit-66f83981.parquet","status":"experimental","variant":"MLX 0.32.2 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T01-35-00-4649920-throughput-mlx-community-qwen3-8-27b-8bit-6a8dd904.parquet","status":"experimental","variant":"MLX 0.32.2 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T01-52-55-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-45eeff19.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p77-o512-batched-seed-unspecified","context_band":"4k","max_output_tokens":512,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun 0e5ef9f2","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T01-53-38-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-9d03cb1c.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p77-o512-batched-seed-unspecified","context_band":"4k","max_output_tokens":512,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun e91ae156","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T01-54-23-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-b0d7ad40.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p77-o512-batched-seed-unspecified","context_band":"4k","max_output_tokens":512,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun 3b63a928","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T01-56-18-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-1713f2a7.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun 5242ebe7","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T01-59-28-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-801b68bd.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun 3c12b342","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-02-21-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-e2b62210.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun 392f56ae","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-03-26-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-662fe353.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p4148-o1024-batched-seed-unspecified","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun ab4033e1","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-04-46-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-d38fa8e4.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p4148-o1024-batched-seed-unspecified","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun 6715c260","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-06-05-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-f3aefac9.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p4148-o1024-batched-seed-unspecified","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun d7b21add","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-11-23-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-90cf49c9.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 Poisson 0.06 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-poisson0p06rps-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun e1539e16","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-11-56-4649920-throughput-mlx-community-qwen3-8-27b-8bit-5c586c67.parquet","status":"experimental","variant":"MLX 0.32.2 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-13-37-74a21ee-framework-eval-unsloth-qwen3-8-27b-gguf-ca51c36f.parquet","status":"scored","variant":"IFBench prompt_level_strict · reasoning off · 32k","comparison_group":"ifbench-prompt-level-strict-n100","context_band":"32k","max_output_tokens":null,"completion_state":"complete","caveat":"","machine":"RTX 4080 SUPER 16GB","engine":"llama.cpp","engine_version":"b11459","engine_build":"b11459","engine_commit":"f498f864f","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-14-30-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-dafe0076.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 Poisson 0.12 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-poisson0p12rps-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun 6c81cf72","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-16-35-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-8f707c65.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 Poisson 0.25 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-poisson0p25rps-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun 0451a8db","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-18-19-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-2b0b0114.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun 392f56ae","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-21-15-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-65c67ef2.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p251-o1024-batched-seed-unspecified","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 8/8; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun 2ebd6f42","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-24-07-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-6e62da80.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p242-o1024-batched-seed-unspecified","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 64/64; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun a93984c5","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-26-11-74a21ee-throughput-mlx-community-qwen3-8-27b-8bit-fe30cd47.parquet","status":"experimental","variant":"MLX 0.32.2 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-33-46-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-0e5ef9f2.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p77-o512-batched-seed-unspecified","context_band":"4k","max_output_tokens":512,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-34-33-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-e91ae156.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p77-o512-batched-seed-unspecified","context_band":"4k","max_output_tokens":512,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-35-16-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-3b63a928.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p77-o512-batched-seed-unspecified","context_band":"4k","max_output_tokens":512,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-37-17-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-5242ebe7.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-39-31-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-3c12b342.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-41-17-74a21ee-throughput-mlx-community-qwen3-8-27b-8bit-40513908.parquet","status":"experimental","variant":"MLX 0.32.2 256k-context Poisson 0.06 rps","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-41-39-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-d4fabad8.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun 392f56ae","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-42-44-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-ab4033e1.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p4148-o1024-batched-seed-unspecified","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-43-57-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-6715c260.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p4148-o1024-batched-seed-unspecified","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-45-12-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-d7b21add.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p4148-o1024-batched-seed-unspecified","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-48-15-74a21ee-framework-eval-unsloth-qwen3-8-27b-gguf-17a6dd22.parquet","status":"scored","variant":"IFBench prompt_level_strict · reasoning low · 32k","comparison_group":"ifbench-prompt-level-strict-n100","context_band":"32k","max_output_tokens":null,"completion_state":"complete","caveat":"","machine":"RTX 4080 SUPER 16GB","engine":"llama.cpp","engine_version":"b11459","engine_build":"b11459","engine_commit":"f498f864f","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-55-20-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-6c81cf72.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 Poisson 0.12 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-poisson0p12rps-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-55-40-74a21ee-throughput-mlx-community-qwen3-8-27b-8bit-a75ed666.parquet","status":"experimental","variant":"MLX 0.32.2 256k-context Poisson 0.12 rps","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-57-15-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-0451a8db.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 Poisson 0.25 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-poisson0p25rps-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-58-50-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-392f56ae.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T02-58-58-74a21ee-throughput-mlx-community-qwen3-8-27b-8bit-1d248755.parquet","status":"experimental","variant":"MLX 0.32.2 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-01-18-74a21ee-throughput-mlx-community-qwen3-8-27b-8bit-18b839d9.parquet","status":"experimental","variant":"MLX 0.32.2 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-01-23-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-2ebd6f42.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p251-o1024-batched-seed-unspecified","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 8/8; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-04-11-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-a93984c5.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p242-o1024-batched-seed-unspecified","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 64/64; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-04-41-74a21ee-throughput-mlx-community-qwen3-8-27b-8bit-45619dda.parquet","status":"experimental","variant":"MLX 0.32.2 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-10-35-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-e1539e16.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP0 Poisson 0.06 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-poisson0p06rps-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-10-42-74a21ee-throughput-mlx-community-qwen3-8-27b-8bit-e26ddd60.parquet","status":"experimental","variant":"MLX 0.32.2 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-17-50-74a21ee-throughput-mlx-community-qwen3-8-27b-8bit-d7145aef.parquet","status":"experimental","variant":"MLX 0.32.2 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-21-34-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-fa2409eb.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p77-o512-batched-seed-unspecified","context_band":"4k","max_output_tokens":512,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-22-13-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-c7b29ccf.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p77-o512-batched-seed-unspecified","context_band":"4k","max_output_tokens":512,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-22-46-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-c43e9c83.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p77-o512-batched-seed-unspecified","context_band":"4k","max_output_tokens":512,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-23-57-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-0c971c46.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-25-10-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-59da6572.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-26-28-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-31993221.parquet","status":"invalid","variant":"vLLM 0.31.0 64k×8 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"superseded by later exact-cell rerun 79e6ace3","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-27-13-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-a836ea74.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p4148-o1024-batched-seed-unspecified","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-28-05-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-64c0a715.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p4148-o1024-batched-seed-unspecified","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-29-05-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-6db8ab16.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p4148-o1024-batched-seed-unspecified","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-30-23-74a21ee-throughput-mlx-community-qwen3-8-27b-8bit-33ef4839.parquet","status":"experimental","variant":"MLX 0.32.2 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":false,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-39-02-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-606ea2ca.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on Poisson 0.12 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-poisson0p12rps-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-40-37-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-af8e6c23.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on Poisson 0.25 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-poisson0p25rps-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-41-40-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-79e6ace3.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-43-09-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-bf20e127.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p251-o1024-batched-seed-unspecified","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 8/8; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-44-59-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-5c9ede73.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p242-o1024-batched-seed-unspecified","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 64/64; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-56-05-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-f8753bf6.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP-on Poisson 0.06 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-poisson0p06rps-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-58-09-74a21ee-framework-eval-nvidia-qwen3-8-27b-nvfp4-7dd1ade8.parquet","status":"scored","variant":"IFBench prompt_level_strict · reasoning off · 64k","comparison_group":"ifbench-prompt-level-strict-n100","context_band":"64k","max_output_tokens":null,"completion_state":"complete","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-58-20-74a21ee-framework-eval-nvidia-qwen3-8-27b-nvfp4-f62f4349.parquet","status":"scored","variant":"IFBench prompt_level_strict · reasoning low · 64k","comparison_group":"ifbench-prompt-level-strict-n100","context_band":"64k","max_output_tokens":null,"completion_state":"complete","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T03-58-22-74a21ee-framework-eval-nvidia-qwen3-8-27b-nvfp4-c9d7de1e.parquet","status":"scored","variant":"IFBench prompt_level_strict · reasoning xhigh · 64k","comparison_group":"ifbench-prompt-level-strict-n100","context_band":"64k","max_output_tokens":null,"completion_state":"complete","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T04-19-32-74a21ee-throughput-unsloth-qwen3-8-27b-gguf-8efd3a3f.parquet","status":"experimental","variant":"llama.cpp 10063 (7d56da7) 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T04-23-58-74a21ee-throughput-unsloth-qwen3-8-27b-gguf-93cd0635.parquet","status":"experimental","variant":"llama.cpp 10063 (7d56da7) 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T04-32-08-74a21ee-throughput-unsloth-qwen3-8-27b-gguf-f0b0d34e.parquet","status":"experimental","variant":"llama.cpp 10063 (7d56da7) 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T05-16-56-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-a36ed64a.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context batched","comparison_group":"throughput-unassigned","context_band":"128k","max_output_tokens":1024,"completion_state":"complete 2/2; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T05-17-59-74a21ee-framework-eval-unsloth-qwen3-8-flash-next-gguf-c545b1fa.parquet","status":"scored","variant":"IFBench prompt_level_strict · reasoning off · 128k","comparison_group":"ifbench-prompt-level-strict-n100","context_band":"128k","max_output_tokens":null,"completion_state":"complete","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T05-18-16-74a21ee-framework-eval-unsloth-qwen3-8-flash-next-gguf-3214fcaa.parquet","status":"scored","variant":"IFBench prompt_level_strict · reasoning low · 128k","comparison_group":"ifbench-prompt-level-strict-n100","context_band":"128k","max_output_tokens":null,"completion_state":"complete","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T05-18-37-74a21ee-framework-eval-unsloth-qwen3-8-flash-next-gguf-32566dc9.parquet","status":"scored","variant":"IFBench prompt_level_strict · reasoning xhigh · 128k","comparison_group":"ifbench-prompt-level-strict-n100","context_band":"128k","max_output_tokens":null,"completion_state":"complete","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T05-58-37-74a21ee-throughput-unsloth-qwen3-8-27b-gguf-ee0f77d5.parquet","status":"experimental","variant":"llama.cpp 10063 (7d56da7, not latest) 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T06-02-45-74a21ee-throughput-unsloth-qwen3-8-27b-gguf-062d950a.parquet","status":"experimental","variant":"llama.cpp 10063 (7d56da7, not latest) 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T06-14-00-74a21ee-throughput-unsloth-qwen3-8-27b-gguf-bcf73cfc.parquet","status":"experimental","variant":"llama.cpp 10063 (7d56da7, not latest) 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T06-34-55-74a21ee-throughput-unsloth-qwen3-8-27b-gguf-edcf8b44.parquet","status":"experimental","variant":"llama.cpp 10063 (7d56da7, not latest) 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T07-52-47-74a21ee-throughput-unsloth-qwen3-8-27b-gguf-bf633124.parquet","status":"experimental","variant":"llama.cpp v0.7.0 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T08-01-39-74a21ee-throughput-unsloth-qwen3-8-27b-gguf-451f5f3c.parquet","status":"experimental","variant":"llama.cpp v0.7.0 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T08-17-02-74a21ee-throughput-unsloth-qwen3-8-27b-gguf-eeb13d16.parquet","status":"experimental","variant":"llama.cpp v0.7.0 256k-context Poisson 0.12 rps","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T08-47-06-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-7d7cdc4d.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context batched","comparison_group":"throughput-unassigned","context_band":"128k","max_output_tokens":1024,"completion_state":"complete 2/2; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-01-45-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-f631b6f6.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-02-30-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-482a18b9.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-03-34-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-2465816e.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-04-49-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-3cd50a42.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-07-31-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-a10e4057.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-10-58-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-9d585f7a.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-11-45-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-354ab2d4.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-13-36-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-45539174.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-16-24-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-1f35bb1b.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-21-35-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-6808a97c.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context Poisson 0.06 rps","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-44-39-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-5f12e1c8.parquet","status":"scored","variant":"vLLM 0.31.0 128k×2 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen38128k2fp8-p127028-o1024-batched-seed201","context_band":"128k","max_output_tokens":1024,"completion_state":"complete 2/2; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-47-19-74a21ee-framework-eval-unsloth-qwen3-8-27b-gguf-ae9100e6.parquet","status":"experimental","variant":"IFBench prompt_level_strict · reasoning xhigh · 32k","comparison_group":"ifbench-prompt-level-strict-n30-easy-slice","context_band":"32k","max_output_tokens":null,"completion_state":"complete","caveat":"n=30 easy slice; distinct item set from n=100 IFBench rows","machine":"RTX 4080 SUPER 16GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-47-32-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-aef2795c.parquet","status":"scored","variant":"vLLM 0.31.0 128k×2 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen38128k2fp8-p127028-o1024-batched-seed202","context_band":"128k","max_output_tokens":1024,"completion_state":"complete 4/4; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T09-48-30-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-df1df822.parquet","status":"scored","variant":"vLLM 0.31.0 128k×2 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen38128k2fp8-p1076-o2048-batched-seed203","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 6/6; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-05-48-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-b5e43f85.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on Poisson 0.25 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p4148-o1024-poisson0p25rps-seed301","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-08-58-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-f424bdcc.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on Poisson 0.1 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p31796-o1024-poisson0p1rps-seed302","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-10-51-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-c2546cbf.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p31796-o1024-batched-seed303","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-12-15-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-010eb746.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p1076-o2048-batched-seed304","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 48/48; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-16-25-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-241d130c.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on Poisson 0.15 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p31796-o1024-poisson0p15rps-seed305","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 32/32; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-17-56-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-699883e4.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p77-o512-batched-seed311","context_band":"4k","max_output_tokens":512,"completion_state":"complete 36/36; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-18-31-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-7322df55.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p77-o512-batched-seed312","context_band":"4k","max_output_tokens":512,"completion_state":"complete 48/48; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-19-52-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-0d474562.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p1076-o2048-batched-seed313","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 36/36; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-20-52-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-89640b0b.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p4148-o1024-batched-seed314","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 36/36; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-29-31-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-9c00d830.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on Poisson 0.5 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p4148-o1024-poisson0p5rps-seed316","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 48/48; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-33-21-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-faa6a9fd.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on Poisson 0.12 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p31796-o1024-poisson0p12rps-seed317","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-35-07-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-7eebe41d.parquet","status":"experimental","variant":"llama.cpp v0.7.0 6k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-36-27-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-adb42e57.parquet","status":"experimental","variant":"llama.cpp v0.7.0 6k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-36-45-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-0c7fc647.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p31796-o1024-batched-seed318","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 32/32; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-38-36-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-743c6b5c.parquet","status":"experimental","variant":"llama.cpp v0.7.0 6k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":512,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-39-31-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-861696c2.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p4148-o1024-batched-seed321","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 48/48; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-43-04-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-ca7f6ed0.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p31796-o1024-batched-seed322","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 34/34; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-43-59-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-8757f5be.parquet","status":"experimental","variant":"llama.cpp v0.7.0 6k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T10-44-42-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-4cfaefa1.parquet","status":"scored","variant":"vLLM 0.31.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p1076-o2048-batched-seed323","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 51/51; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T11-01-41-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-ad8be854.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP2 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed401","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T11-02-48-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-7998fd87.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP2 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed402","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T11-03-57-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-65747a20.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP2 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed403","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T11-05-49-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-f5e10f42.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP2 Poisson 0.25 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p4148-o1024-poisson0p25rps-seed404","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T11-08-54-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-9ec89997.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP2 Poisson 0.1 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p31796-o1024-poisson0p1rps-seed405","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T11-23-32-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-993c62f3.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP4 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed401","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T11-24-47-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-3d60eb79.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP4 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed402","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T11-25-54-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-410e65fc.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP4 batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p1076-o2048-batched-seed403","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T11-27-46-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-bb32a0f8.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP4 Poisson 0.25 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p4148-o1024-poisson0p25rps-seed404","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T11-30-51-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-33a30639.parquet","status":"scored","variant":"vLLM 0.31.0 64k×8 MTP4 Poisson 0.1 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k8fp8-p31796-o1024-poisson0p1rps-seed405","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T12-26-32-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-d3fd1d71.parquet","status":"scored","variant":"vLLM 0.31.0 128k×11 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen38128k11fp8-p127028-o1024-batched-seed501","context_band":"128k","max_output_tokens":1024,"completion_state":"complete 4/4; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T12-32-49-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-29347aa2.parquet","status":"scored","variant":"llama.cpp 0.7.0 8×6144 batched","comparison_group":"macstudiom4max128gb-q38flashnextgsqrco-llamacpp-studiofn8x6144-p3976-o1024-batched-seed-unspecified","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"Mac Studio M4 Max 128GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d222","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T12-34-33-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-b64c99ef.parquet","status":"scored","variant":"vLLM 0.31.0 128k×11 MTP-on Poisson 0.02 rps","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen38128k11fp8-p127028-o1024-poisson0p02rps-seed502","context_band":"128k","max_output_tokens":1024,"completion_state":"complete 8/8; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T12-39-40-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-6f3a9964.parquet","status":"scored","variant":"vLLM 0.31.0 128k×11 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen38128k11fp8-p127028-o1024-batched-seed503","context_band":"128k","max_output_tokens":1024,"completion_state":"complete 8/8; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T12-40-26-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-d520d519.parquet","status":"scored","variant":"llama.cpp 0.7.0 8×6144 batched","comparison_group":"macstudiom4max128gb-q38flashnextgsqrco-llamacpp-studiofn8x6144-p4035-o1024-batched-seed-unspecified","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"Mac Studio M4 Max 128GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d222","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T12-46-32-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-670c8500.parquet","status":"scored","variant":"vLLM 0.31.0 128k×11 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen38128k11fp8-p127028-o1024-batched-seed504","context_band":"128k","max_output_tokens":1024,"completion_state":"complete 11/11; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T12-54-13-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-0b98fea7.parquet","status":"scored","variant":"vLLM 0.31.0 128k×11 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen38128k11fp8-p127028-o1024-batched-seed505","context_band":"128k","max_output_tokens":1024,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T12-54-26-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-9e878713.parquet","status":"scored","variant":"llama.cpp 0.7.0 8×6144 batched","comparison_group":"macstudiom4max128gb-q38flashnextgsqrco-llamacpp-studiofn8x6144-p4030-o1024-batched-seed-unspecified","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 27/27; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"Mac Studio M4 Max 128GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d222","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T13-04-39-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-763b87b9.parquet","status":"scored","variant":"llama.cpp 0.7.0 4×40960 Poisson 0.25 rps","comparison_group":"macstudiom4max128gb-q38flashnextgsqrco-llamacpp-studiofn4x40960-p4044-o1024-poisson0p25rps-seed-unspecified","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"Mac Studio M4 Max 128GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d222","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T13-06-07-74a21ee3b6ecca3ba8c03ce7d87f8a62fa57426b-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-40e5b59f.parquet","status":"scored","variant":"llama.cpp 0.7.0 4×40960 Poisson 0.05 rps","comparison_group":"macstudiom4max128gb-q38flashnextgsqrco-llamacpp-studiofn4x40960-p30533-o1024-poisson0p05rps-seed-unspecified","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"Mac Studio M4 Max 128GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d222","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T13-12-54-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-3a78a1f6.parquet","status":"scored","variant":"vLLM 0.31.0 256k×5 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen38256k5fp8-p260149-o1024-batched-seed601","context_band":"256k","max_output_tokens":1024,"completion_state":"complete 1/1; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T13-17-05-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-82fae1b9.parquet","status":"scored","variant":"vLLM 0.31.0 256k×5 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen38256k5fp8-p260148-o1024-batched-seed602","context_band":"256k","max_output_tokens":1024,"completion_state":"complete 2/2; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T13-25-15-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-0014b2a2.parquet","status":"scored","variant":"vLLM 0.31.0 256k×5 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen38256k5fp8-p260148-o1024-batched-seed603","context_band":"256k","max_output_tokens":1024,"completion_state":"complete 4/4; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T13-35-22-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-2af3e1d3.parquet","status":"scored","variant":"vLLM 0.31.0 256k×5 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen38256k5fp8-p260148-o1024-batched-seed604","context_band":"256k","max_output_tokens":1024,"completion_state":"complete 5/5; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T13-47-32-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-2b4f182c.parquet","status":"scored","variant":"vLLM 0.31.0 256k×5 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen38256k5fp8-p260148-o1024-batched-seed605","context_band":"256k","max_output_tokens":1024,"completion_state":"complete 6/6; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T14-04-05-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-230e7da5.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context Poisson 0.25 rps","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":1024,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T14-41-24-74a21ee-throughput-unsloth-qwen3-8-flash-next-gguf-499c6ff6.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 128k-context Poisson 0.01 rps","comparison_group":"throughput-unassigned","context_band":"128k","max_output_tokens":1024,"completion_state":"complete 6/6; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T15-02-27-74a21ee-framework-eval-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-3efcba2c.parquet","status":"scored","variant":"IFBench prompt_level_strict · reasoning off · 16k","comparison_group":"ifbench-prompt-level-strict-n100","context_band":"16k","max_output_tokens":null,"completion_state":"complete","caveat":"","machine":"Mac Studio M4 Max 128GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d222","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T15-21-43-74a21ee-framework-eval-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-4b65721e.parquet","status":"scored","variant":"IFBench prompt_level_strict · reasoning low · 16k","comparison_group":"ifbench-prompt-level-strict-n100","context_band":"16k","max_output_tokens":null,"completion_state":"complete","caveat":"","machine":"Mac Studio M4 Max 128GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d222","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T18-51-50-74a21ee-coding-nvidia-qwen3-8-27b-nvfp4-e7324bca.parquet","status":"invalid","variant":"BigCodeBench-Hard Complete; first-30 preliminary · reasoning xhigh","comparison_group":"bigcodebench-hard-first30-32768-cap","context_band":"n/a","max_output_tokens":32768,"completion_state":"complete","caveat":"superseded by complete 122880-token-cap official rerun","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"HISTORICAL_PROFILE_NOT_RECORDED"}}, |
| {"path":"data/run-2026-10-17T19-19-51-74a21ee-coding-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-390f1d69.parquet","status":"invalid","variant":"BigCodeBench-Hard Complete; first-30 preliminary · reasoning xhigh","comparison_group":"bigcodebench-hard-first30-32768-cap","context_band":"n/a","max_output_tokens":32768,"completion_state":"complete","caveat":"superseded by complete 122880-token-cap official rerun","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T19-26-12.734-74a21ee-tool-queries-unsloth-qwen3-8-flash-next-gguf-34aaf35e.parquet","status":"invalid","variant":"tool-queries v2 · reasoning xhigh · 160k slot context","comparison_group":"promptfoo-v2-tool-queries-20","context_band":"160k slot context","max_output_tokens":null,"completion_state":"all prompts completed; score grader invalid","caveat":"v2 grading later found invalid; full-correct counts 0/20, 0/20, and 1/14 respectively","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T19-28-26.729-74a21ee-agentic-unsloth-qwen3-8-flash-next-gguf-7beb4baf.parquet","status":"invalid","variant":"agentic v2 · reasoning xhigh · 160k slot context","comparison_group":"promptfoo-v2-agentic-20","context_band":"160k slot context","max_output_tokens":null,"completion_state":"all prompts completed; score grader invalid","caveat":"v2 grading later found invalid; full-correct counts 0/20, 0/20, and 1/14 respectively","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T19-30-24.236-74a21ee-agentic-unsloth-qwen3-8-flash-next-gguf-5b22bc03.parquet","status":"invalid","variant":"agentic-mcp v2 · reasoning xhigh · 160k slot context","comparison_group":"promptfoo-v2-agentic-mcp-14","context_band":"160k slot context","max_output_tokens":null,"completion_state":"all prompts completed; score grader invalid","caveat":"v2 grading later found invalid; full-correct counts 0/20, 0/20, and 1/14 respectively","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T21-02-53-74a21ee-coding-unsloth-qwen3-8-flash-next-gguf-7a477f2a.parquet","status":"invalid","variant":"BigCodeBench-Hard Complete; first-30 preliminary · reasoning xhigh","comparison_group":"bigcodebench-hard-first30-32768-cap","context_band":"n/a","max_output_tokens":32768,"completion_state":"complete","caveat":"superseded by complete 122880-token-cap official rerun","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T21-11-42.771-74a21ee3b6ecca3ba8c03ce7d87f8a62fa57426b-framework-eval-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-274270ef.parquet","status":"invalid","variant":"tool-queries v2 · reasoning xhigh · 160k slot context","comparison_group":"promptfoo-v2-tool-queries-20","context_band":"160k slot context","max_output_tokens":null,"completion_state":"all prompts completed; score grader invalid","caveat":"v2 grading later found invalid; full-correct counts 0/20, 0/20, and 1/14 respectively","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T21-16-39.571-74a21ee3b6ecca3ba8c03ce7d87f8a62fa57426b-framework-eval-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-f31a8738.parquet","status":"invalid","variant":"agentic v2 · reasoning xhigh · 160k slot context","comparison_group":"promptfoo-v2-agentic-20","context_band":"160k slot context","max_output_tokens":null,"completion_state":"all prompts completed; score grader invalid","caveat":"v2 grading later found invalid; full-correct counts 0/20, 0/20, and 1/14 respectively","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T21-18-26.440-74a21ee3b6ecca3ba8c03ce7d87f8a62fa57426b-framework-eval-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-6c3ebc3e.parquet","status":"invalid","variant":"agentic-mcp v2 · reasoning xhigh · 160k slot context","comparison_group":"promptfoo-v2-agentic-mcp-14","context_band":"160k slot context","max_output_tokens":null,"completion_state":"all prompts completed; score grader invalid","caveat":"v2 grading later found invalid; full-correct counts 0/20, 0/20, and 1/14 respectively","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T23-33-24-ccadbce-coding-unsloth-qwen3-8-flash-next-gguf-eaffd69f.parquet","status":"scored","variant":"BigCodeBench-Hard Complete; official scorer · reasoning xhigh","comparison_group":"bigcodebench-hard-complete-v0.1.4","context_band":"n/a","max_output_tokens":122880,"completion_state":"complete","caveat":"published machine field missing; canonical machine assigned from official scorer report","machine":"RTX PRO 6000 Max-Q 96GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T23-39-35-ccadbce5283d75e4c2057e1af1a56eb4208d6f9a-coding-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-1e850802.parquet","status":"scored","variant":"BigCodeBench-Hard Complete; official scorer · reasoning xhigh","comparison_group":"bigcodebench-hard-complete-v0.1.4","context_band":"n/a","max_output_tokens":122880,"completion_state":"complete","caveat":"published machine field missing; canonical machine assigned from official scorer report","machine":"Mac Studio M4 Max 128GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d222","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T23-46-53.229-dc5d1efd24ec029b2ce5176aa1626d77d31296fe-repo-qa-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-7d73f546.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning off · 30k","comparison_group":"repo-qa-deterministic-60","context_band":"30k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"Mac Studio M4 Max 128GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d222","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-17T23-55-49.132-74a21ee-framework-eval-unsloth-qwen3-8-flash-next-gguf-c5ff2be6.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning off · 30k","comparison_group":"repo-qa-deterministic-60","context_band":"30k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T01-17-45.889-74a21ee-framework-eval-unsloth-qwen3-8-flash-next-gguf-487998b6.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning xhigh · 30k","comparison_group":"repo-qa-deterministic-60","context_band":"30k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T02-33-01.748-74a21ee-framework-eval-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-4b00d360.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning xhigh · 30k","comparison_group":"repo-qa-deterministic-60","context_band":"30k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"Mac Studio M4 Max 128GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d222","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T02-33-01.748-74a21ee-framework-eval-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-ccb4b6ce.parquet","status":"invalid","variant":"repo-qa; deterministic 60 questions · reasoning xhigh · 30k","comparison_group":"repo-qa-deterministic-60","context_band":"30k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"stale duplicate of complete 4b00d360 shard; retain 4b00d360","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T03-43-00-74a21ee3b6ecca3ba8c03ce7d87f8a62fa57426b-framework-eval-nvidia-qwen3-8-27b-nvfp4-fcee70e3.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning off · 30k","comparison_group":"repo-qa-deterministic-60","context_band":"30k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T03-50-36-74a21ee3b6ecca3ba8c03ce7d87f8a62fa57426b-framework-eval-nvidia-qwen3-8-27b-nvfp4-426cb5b3.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning off · 120k","comparison_group":"repo-qa-deterministic-60","context_band":"120k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T04-05-30-74a21ee3b6ecca3ba8c03ce7d87f8a62fa57426b-framework-eval-nvidia-qwen3-8-27b-nvfp4-87799e4f.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning xhigh · 30k","comparison_group":"repo-qa-deterministic-60","context_band":"30k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T04-07-58-74a21ee3b6ecca3ba8c03ce7d87f8a62fa57426b-framework-eval-nvidia-qwen3-8-27b-nvfp4-a2b72c22.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning xhigh · 120k","comparison_group":"repo-qa-deterministic-60","context_band":"120k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T04-15-07.292-74a21ee3b6ecca3ba8c03ce7d87f8a62fa57426b-framework-eval-unsloth-qwen3-8-27b-gguf-5453d6b3.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning off · 30k","comparison_group":"repo-qa-deterministic-60","context_band":"30k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"RTX 4080 SUPER 16GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T04-24-26.574-74a21ee3b6ecca3ba8c03ce7d87f8a62fa57426b-framework-eval-unsloth-qwen3-8-27b-gguf-13923485.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning xhigh · 30k","comparison_group":"repo-qa-deterministic-60","context_band":"30k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"RTX 4080 SUPER 16GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T04-40-47-74a21ee-coding-unsloth-qwen3-8-27b-gguf-075af03f.parquet","status":"scored","variant":"BigCodeBench-Hard Complete; official scorer · reasoning xhigh","comparison_group":"bigcodebench-hard-complete-v0.1.4","context_band":"n/a","max_output_tokens":122880,"completion_state":"complete","caveat":"published machine field missing; canonical machine assigned from official scorer report; engine build from the same serving profile (converged before both runs, unchanged until scoring), as recorded on the 04:24Z repo-qa run","machine":"RTX 4080 SUPER 16GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":null},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"HOST_RAM_LIMIT_NOT_RECORDED","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T05-02-52-ccadbce5283d75e4c2057e1af1a56eb4208d6f9a-coding-nvidia-qwen3-8-27b-nvfp4-9cc2bd80.parquet","status":"scored","variant":"BigCodeBench-Hard Complete; official scorer · reasoning xhigh","comparison_group":"bigcodebench-hard-complete-v0.1.4","context_band":"n/a","max_output_tokens":122880,"completion_state":"complete","caveat":"published machine field missing; canonical machine assigned from official scorer report","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"HISTORICAL_PROFILE_NOT_RECORDED"}}, |
| {"path":"data/run-2026-10-18T05-22-21-ccadbce-coding-nvidia-qwen3-8-27b-nvfp4-723dd669.parquet","status":"scored","variant":"BigCodeBench-Hard Complete; official scorer · reasoning off","comparison_group":"bigcodebench-hard-complete-v0.1.4","context_band":"n/a","max_output_tokens":122880,"completion_state":"complete","caveat":"published machine field missing; canonical machine assigned from official scorer report","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"HISTORICAL_PROFILE_NOT_RECORDED"}}, |
| {"path":"data/run-2026-10-18T05-24-01-ccadbce-coding-nvidia-qwen3-8-27b-nvfp4-7b0c0790.parquet","status":"scored","variant":"BigCodeBench-Hard Complete; official scorer · reasoning low","comparison_group":"bigcodebench-hard-complete-v0.1.4","context_band":"n/a","max_output_tokens":122880,"completion_state":"complete","caveat":"published machine field missing; canonical machine assigned from official scorer report","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.31.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"HISTORICAL_PROFILE_NOT_RECORDED"}}, |
| {"path":"data/run-2026-10-18T07-11-24.562-74a21ee-framework-eval-nvidia-nvidia-nemotron-3-5-bolt-74b-a9b-24gb-gguf-a2e1781f.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning off · 30k","comparison_group":"repo-qa-deterministic-60","context_band":"30k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T07-21-05.118-74a21ee-framework-eval-nvidia-nvidia-nemotron-3-5-bolt-74b-a9b-24gb-gguf-1cf98ace.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning xhigh · 30k","comparison_group":"repo-qa-deterministic-60","context_band":"30k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T07-44-14.978-74a21ee-framework-eval-nvidia-nvidia-nemotron-3-5-bolt-74b-a9b-24gb-gguf-92a02d2a.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning off · 120k","comparison_group":"repo-qa-deterministic-60","context_band":"120k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T08-16-46.827-74a21ee-framework-eval-nvidia-nvidia-nemotron-3-5-bolt-74b-a9b-24gb-gguf-7c0dee21.parquet","status":"scored","variant":"repo-qa; deterministic 60 questions · reasoning xhigh · 120k","comparison_group":"repo-qa-deterministic-60","context_band":"120k","max_output_tokens":null,"completion_state":"complete 60-question set","caveat":"","machine":"RTX PRO 6000 Max-Q 96GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T09-49-58-74a21ee-throughput-nvidia-nvidia-nemotron-3-5-bolt-74b-a9b-24gb-gguf-7ec27269.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 4/4; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T09-51-44-74a21ee-throughput-nvidia-nvidia-nemotron-3-5-bolt-74b-a9b-24gb-gguf-f78c7580.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T09-57-01-74a21ee-throughput-nvidia-nvidia-nemotron-3-5-bolt-74b-a9b-24gb-gguf-7028c69d.parquet","status":"experimental","variant":"llama.cpp b11669-620d2225c 256k-context batched","comparison_group":"throughput-unassigned","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"RTX PRO 6000 Max-Q 96GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T10-54-33-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-d9e6556d.parquet","status":"scored","variant":"llama.cpp 0.7.0 8×163840 batched","comparison_group":"macstudiom4max128gb-q38flashnextgsqrco-llamacpp-studiofn8x163840-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 4/4; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"Mac Studio M4 Max 128GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d222","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T11-03-04-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-88e9916e.parquet","status":"scored","variant":"llama.cpp 0.7.0 8×163840 Poisson 0.05 rps","comparison_group":"macstudiom4max128gb-q38flashnextgsqrco-llamacpp-studiofn8x163840-p1076-o2048-poisson0p05rps-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 4/4; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"Mac Studio M4 Max 128GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d222","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T11-08-05-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-04c945cf.parquet","status":"scored","variant":"llama.cpp 0.7.0 8×163840 Poisson 0.1 rps","comparison_group":"macstudiom4max128gb-q38flashnextgsqrco-llamacpp-studiofn8x163840-p31744-o1024-poisson0p1rps-seed-unspecified","context_band":"32k","max_output_tokens":1024,"completion_state":"complete 16/16; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"Mac Studio M4 Max 128GB","engine":"llama.cpp","engine_version":"b11669","engine_build":"b11669","engine_commit":"620d2225c","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T11-32-56-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-011e6aee.parquet","status":"experimental","variant":"llama.cpp 0.7.0 8×163840 batched","comparison_group":"throughput-unassigned","context_band":"128k","max_output_tokens":1024,"completion_state":"complete 1/1; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T11-39-30-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-26dc3d24.parquet","status":"experimental","variant":"llama.cpp 0.7.0 8×163840 batched","comparison_group":"throughput-unassigned","context_band":"128k","max_output_tokens":1024,"completion_state":"complete 2/2; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T11-51-55-74a21ee-throughput-ista-daslab-qwen3-8-flash-next-gsq-rco-gguf-10c156e5.parquet","status":"experimental","variant":"llama.cpp 0.7.0 1×262144 batched","comparison_group":"throughput-unassigned","context_band":"256k","max_output_tokens":1024,"completion_state":"complete 1/1; failed=0; ignore_eos=true","caveat":"valid raw cell; outside final kept comparison","machine":"Mac Studio M4 Max 128GB","campaign_dimensions":{"hardware":{"host_ram_limit_gib":128.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"llamacpp_cache_ram_mib":8192,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T11-54-58-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-9b161102.parquet","status":"scored","variant":"vLLM 0.32.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.32.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T11-58-13-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-37701fd1.parquet","status":"scored","variant":"vLLM 0.32.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 12/12; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.32.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T11-59-38-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-41ad7884.parquet","status":"scored","variant":"vLLM 0.32.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.32.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T12-01-15-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-162b7dc5.parquet","status":"scored","variant":"vLLM 0.32.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p1076-o2048-batched-seed-unspecified","context_band":"4k","max_output_tokens":2048,"completion_state":"complete 48/48; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.32.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T12-01-49-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-8538c55a.parquet","status":"scored","variant":"vLLM 0.32.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p77-o512-batched-seed-unspecified","context_band":"4k","max_output_tokens":512,"completion_state":"complete 3/3; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.32.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T12-02-28-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-d48049ff.parquet","status":"scored","variant":"vLLM 0.32.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p77-o512-batched-seed-unspecified","context_band":"4k","max_output_tokens":512,"completion_state":"complete 24/24; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.32.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"path":"data/run-2026-10-18T12-03-11-74a21ee-throughput-nvidia-qwen3-8-27b-nvfp4-53b96c7e.parquet","status":"scored","variant":"vLLM 0.32.0 64k×16 MTP-on batched","comparison_group":"rtxpro6000maxq96gb-q3827b-vllm-qwen3864k16fp8-p77-o512-batched-seed-unspecified","context_band":"4k","max_output_tokens":512,"completion_state":"complete 48/48; failed=0; ignore_eos=true","caveat":"fixed-length ignore_eos; raw request_rate recorded","machine":"RTX PRO 6000 Max-Q 96GB","engine":"vLLM","engine_version":"0.32.0","campaign_dimensions":{"hardware":{"host_ram_limit_gib":48.0},"run":{"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"cache_workload":"none","prefix_cache":true,"max_num_batched_tokens":8192,"kv_pool_tokens":null,"llamacpp_cache_ram_mib":null,"gpu_memory_utilization":0.85}},"dimension_null_reasons":{"run.kv_pool_tokens":"HISTORICAL_STARTUP_LOG_NOT_CAPTURED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE"}}, |
| {"campaign_dimensions":{"hardware":{"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","interconnect":"tb5-rdma","world_size":2,"host_ram_limit_gib":null},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.8","id":"mlx-community/Qwen3.8-27B-4bit","hf_repo":"mlx-community/Qwen3.8-27B-4bit","revision_sha":"10c35caafbb80f7dc6a7a432cdd11af10a6d4818","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none","llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"MULTI_HOST_RAM_AGGREGATION_UNDEFINED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"},"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T06-34-25-8111301-throughput-mlx-community-qwen3-8-27b-4bit-4bfeb60f.parquet","variant":"MLX 0.32.2 tensor parallel; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-tb5-rdma-qwen3-8-27b-4bit-tensor-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err (1 of 3 reached the 128-token cap); concurrent 1 ok/0 err; answered_rate 1.0","caveat":"valid one-off c=1 sequential+concurrent probe; not final comparison; system wired peak node A 24548 MB, node B 17708 MB (pre-load baseline 10183/3074 MB, includes OS wired); RDMA protection domains 0/0 before load, 1/1 while serving, 0/0 after detach; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","interconnect":"tb5-rdma","world_size":2,"host_ram_limit_gib":null},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.8","id":"mlx-community/Qwen3.8-27B-4bit","hf_repo":"mlx-community/Qwen3.8-27B-4bit","revision_sha":"10c35caafbb80f7dc6a7a432cdd11af10a6d4818","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none","llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"MULTI_HOST_RAM_AGGREGATION_UNDEFINED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"},"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T06-37-54-8111301-reasoning-mlx-community-qwen3-8-27b-4bit-0cbb85a6.parquet","variant":"MLX 0.32.2 tensor parallel; GSM8K 5-shot multiturn, limit 50, thinking off, max_gen_toks 1024","comparison_group":"mlx-tb5-rdma-qwen3-8-27b-4bit-tensor-gsm8k","context_band":"8k","max_output_tokens":1024,"completion_state":"gsm8k 50 samples scored; exact_match 0.98 +/- 0.02 (flexible-extract and strict-match)","caveat":"limit-50 subset of GSM8K; not a full-benchmark score; not final comparison; system wired peak node A 24548 MB, node B 17708 MB (pre-load baseline 10183/3074 MB, includes OS wired); RDMA protection domains 0/0 before load, 1/1 while serving, 0/0 after detach; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","interconnect":"tb5-rdma","world_size":2,"host_ram_limit_gib":null},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"GLM-4.7-Flash","id":"mlx-community/GLM-4.7-Flash-4bit","hf_repo":"mlx-community/GLM-4.7-Flash-4bit","revision_sha":"1454cffb1a21737e162f508e5bc70be9def89276","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none","llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"MULTI_HOST_RAM_AGGREGATION_UNDEFINED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"},"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T07-13-10-8111301-throughput-mlx-community-glm-4-7-flash-4bit-94041f1c.parquet","variant":"MLX 0.32.2 tensor parallel; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-tb5-rdma-glm-4-7-flash-4bit-tensor-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err (0 of 3 reached the 128-token cap); concurrent 1 ok/0 err; answered_rate 1.0","caveat":"valid one-off c=1 sequential+concurrent probe; not final comparison; system wired peak node A 20859 MB, node B 13515 MB (pre-load baseline 10078/3943 MB, includes OS wired); RDMA protection domains 0/0 before load, 1/1 while serving, 0/0 after detach; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","interconnect":"tb5-rdma","world_size":2,"host_ram_limit_gib":null},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"GLM-4.7-REAP-50","id":"mlx-community/GLM-4.7-REAP-50-mxfp4","hf_repo":"mlx-community/GLM-4.7-REAP-50-mxfp4","revision_sha":"9ad1984a355f7593fcc2db1c08576f68ec8c2f12","quantization":"mxfp4"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none","llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"MULTI_HOST_RAM_AGGREGATION_UNDEFINED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"},"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T07-17-46-8111301-throughput-mlx-community-glm-4-7-reap-50-mxfp4-eecf3da2.parquet","variant":"MLX 0.32.2 pipeline parallel; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-tb5-rdma-glm-4-7-reap-50-mxfp4-pipeline-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err (0 of 3 reached the 128-token cap); concurrent 1 ok/0 err; answered_rate 1.0","caveat":"valid one-off c=1 sequential+concurrent probe; not final comparison; system wired peak node A 65957 MB, node B 56150 MB (pre-load baseline 10206/2948 MB, includes OS wired); RDMA protection domains 0/0 before load, 1/1 while serving, 0/0 after detach; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","interconnect":"tb5-rdma","world_size":2,"host_ram_limit_gib":null},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"GLM-4.7-REAP-50","id":"mlx-community/GLM-4.7-REAP-50-mxfp4","hf_repo":"mlx-community/GLM-4.7-REAP-50-mxfp4","revision_sha":"9ad1984a355f7593fcc2db1c08576f68ec8c2f12","quantization":"mxfp4"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none","llamacpp_cache_ram_mib":null,"gpu_memory_utilization":null,"max_num_batched_tokens":null,"kv_pool_tokens":null}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"MULTI_HOST_RAM_AGGREGATION_UNDEFINED","run.llamacpp_cache_ram_mib":"NOT_APPLICABLE_ENGINE","run.gpu_memory_utilization":"NOT_APPLICABLE_ENGINE","run.max_num_batched_tokens":"NOT_APPLICABLE_ENGINE","run.kv_pool_tokens":"NOT_APPLICABLE_ENGINE"},"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T07-22-04-8111301-reasoning-mlx-community-glm-4-7-reap-50-mxfp4-2c53412a.parquet","variant":"MLX 0.32.2 pipeline parallel; GSM8K 5-shot multiturn, limit 50, thinking off, max_gen_toks 1024","comparison_group":"mlx-tb5-rdma-glm-4-7-reap-50-mxfp4-pipeline-gsm8k","context_band":"8k","max_output_tokens":1024,"completion_state":"gsm8k 50 samples scored; exact_match 0.86 +/- 0.0496 (flexible-extract and strict-match)","caveat":"limit-50 subset of GSM8K; not a full-benchmark score; not final comparison; system wired peak node A 65957 MB, node B 56150 MB (pre-load baseline 10206/2948 MB, includes OS wired); RDMA protection domains 0/0 before load, 1/1 while serving, 0/0 after detach; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","interconnect":"tb5-rdma","world_size":2,"host_ram_limit_gib":null},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"gpt-oss-120b","id":"mlx-community/gpt-oss-120b-MXFP4-Q8","hf_repo":"mlx-community/gpt-oss-120b-MXFP4-Q8","revision_sha":"81e5ac3ad0af6efb1298a8e8c7a10ed2990c137b","quantization":"MXFP4 experts, 8-bit other layers"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"MULTI_HOST_RAM_AGGREGATION_UNDEFINED"},"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T07-53-27-8111301-throughput-mlx-community-gpt-oss-120b-mxfp4-q8-1eec5543.parquet","variant":"MLX 0.32.2 tensor parallel; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-tb5-rdma-gpt-oss-120b-mxfp4-q8-tensor-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err (1 of 3 reached the 128-token cap); concurrent 1 ok/0 err; answered_rate 1.0","caveat":"valid one-off c=1 sequential+concurrent probe; not final comparison; reasoning_effort low (no thinking-off switch); system wired peak node A 44809 MB, node B 37275 MB (pre-load baseline 10121/2813 MB, includes OS wired); RDMA protection domains 0/0 before load, 1/1 while serving, 0/0 after detach; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","interconnect":"tb5-rdma","world_size":2,"host_ram_limit_gib":null},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"gpt-oss-120b","id":"mlx-community/gpt-oss-120b-MXFP4-Q8","hf_repo":"mlx-community/gpt-oss-120b-MXFP4-Q8","revision_sha":"81e5ac3ad0af6efb1298a8e8c7a10ed2990c137b","quantization":"MXFP4 experts, 8-bit other layers"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"MULTI_HOST_RAM_AGGREGATION_UNDEFINED"},"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T07-54-45-8111301-reasoning-mlx-community-gpt-oss-120b-mxfp4-q8-d5b59a91.parquet","variant":"MLX 0.32.2 tensor parallel; GSM8K 5-shot multiturn, limit 50, reasoning_effort low, max_gen_toks 2048","comparison_group":"mlx-tb5-rdma-gpt-oss-120b-mxfp4-q8-tensor-gsm8k","context_band":"8k","max_output_tokens":2048,"completion_state":"gsm8k 50 samples scored; flexible-extract exact_match 0.94 +/- 0.0339; strict-match exact_match 0.08 +/- 0.0388","caveat":"limit-50 subset of GSM8K; not a full-benchmark score; not final comparison; headline value is flexible-extract (0.94); strict-match 0.08 is low because the final channel does not emit the #### N answer format; reasoning_effort low (no thinking-off switch); system wired peak node A 44809 MB, node B 37275 MB (pre-load baseline 10121/2813 MB, includes OS wired); RDMA protection domains 0/0 before load, 1/1 while serving, 0/0 after detach; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","interconnect":"tb5-rdma","world_size":2,"host_ram_limit_gib":null},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"gpt-oss-120b","id":"mlx-community/gpt-oss-120b-MXFP4-Q8","hf_repo":"mlx-community/gpt-oss-120b-MXFP4-Q8","revision_sha":"81e5ac3ad0af6efb1298a8e8c7a10ed2990c137b","quantization":"MXFP4 experts, 8-bit other layers"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"MULTI_HOST_RAM_AGGREGATION_UNDEFINED"},"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T07-57-50-8111301-reasoning-mlx-community-gpt-oss-120b-mxfp4-q8-9b0959ad.parquet","variant":"MLX 0.32.2 tensor parallel; GPQA Diamond CoT zero-shot, limit 50, reasoning_effort low, max_gen_toks 4096","comparison_group":"mlx-tb5-rdma-gpt-oss-120b-mxfp4-q8-tensor-gpqa","context_band":"8k","max_output_tokens":4096,"completion_state":"gpqa_diamond_cot_zeroshot 50 samples scored; flexible-extract exact_match 0.64 +/- 0.0686; strict-match exact_match 0.00 +/- 0.0000","caveat":"limit-50 subset of GPQA Diamond; not a full-benchmark score; not final comparison; headline value is flexible-extract (0.64); strict-match 0.00 is a format miss (the strict filter expects the phrase The answer is, which the final answers do not emit); reasoning_effort low (no thinking-off switch); system wired peak node A 44809 MB, node B 37275 MB (pre-load baseline 10121/2813 MB, includes OS wired); RDMA protection domains 0/0 before load, 1/1 while serving, 0/0 after detach; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","interconnect":"tb5-rdma","world_size":2,"host_ram_limit_gib":null},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.5-122B-A10B","id":"mlx-community/Qwen3.5-122B-A10B-OptiQ-2bit","hf_repo":"mlx-community/Qwen3.5-122B-A10B-OptiQ-2bit","revision_sha":"1081c89e171447884d329abd7232566ebd38737c","quantization":"mixed 2-bit/4-bit (OptiQ)"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"MULTI_HOST_RAM_AGGREGATION_UNDEFINED"},"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T08-14-43-8111301-throughput-mlx-community-qwen3-5-122b-a10b-optiq-2bit-eda66cf7.parquet","variant":"MLX 0.32.2 tensor parallel; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-tb5-rdma-qwen-3-5-122b-a10b-optiq-2bit-tensor-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err (0 of 3 reached the 128-token cap); concurrent 1 ok/0 err; answered_rate 1.0","caveat":"valid one-off c=1 sequential+concurrent probe; not final comparison; system wired peak node A 39525 MB, node B 32448 MB (pre-load baseline 10147/2866 MB, includes OS wired); RDMA protection domains 0/0 before load, 1/1 while serving, 0/0 after detach; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","interconnect":"tb5-rdma","world_size":2,"host_ram_limit_gib":null},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.5-122B-A10B","id":"mlx-community/Qwen3.5-122B-A10B-OptiQ-2bit","hf_repo":"mlx-community/Qwen3.5-122B-A10B-OptiQ-2bit","revision_sha":"1081c89e171447884d329abd7232566ebd38737c","quantization":"mixed 2-bit/4-bit (OptiQ)"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"MULTI_HOST_RAM_AGGREGATION_UNDEFINED"},"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T08-15-52-8111301-reasoning-mlx-community-qwen3-5-122b-a10b-optiq-2bit-171d89db.parquet","variant":"MLX 0.32.2 tensor parallel; GSM8K 5-shot multiturn, limit 50, thinking off, max_gen_toks 1024","comparison_group":"mlx-tb5-rdma-qwen-3-5-122b-a10b-optiq-2bit-tensor-gsm8k","context_band":"8k","max_output_tokens":1024,"completion_state":"gsm8k 50 samples scored; exact_match 0.96 +/- 0.028 (flexible-extract and strict-match)","caveat":"limit-50 subset of GSM8K; not a full-benchmark score; not final comparison; system wired peak node A 39525 MB, node B 32448 MB (pre-load baseline 10147/2866 MB, includes OS wired); RDMA protection domains 0/0 before load, 1/1 while serving, 0/0 after detach; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","interconnect":"tb5-rdma","world_size":2,"host_ram_limit_gib":null},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.5-122B-A10B","id":"mlx-community/Qwen3.5-122B-A10B-OptiQ-2bit","hf_repo":"mlx-community/Qwen3.5-122B-A10B-OptiQ-2bit","revision_sha":"1081c89e171447884d329abd7232566ebd38737c","quantization":"mixed 2-bit/4-bit (OptiQ)"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"MULTI_HOST_RAM_AGGREGATION_UNDEFINED"},"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T08-19-12-8111301-reasoning-mlx-community-qwen3-5-122b-a10b-optiq-2bit-143b2785.parquet","variant":"MLX 0.32.2 tensor parallel; GPQA Diamond CoT zero-shot, limit 50, thinking off, max_gen_toks 4096","comparison_group":"mlx-tb5-rdma-qwen-3-5-122b-a10b-optiq-2bit-tensor-gpqa-diamond-cot-zeroshot","context_band":"8k","max_output_tokens":4096,"completion_state":"gpqa_diamond_cot_zeroshot 50 samples scored; flexible-extract exact_match 0.72 +/- 0.0641; strict-match 0.02 +/- 0.0200","caveat":"limit-50 subset of GPQA Diamond; not a full-benchmark score; not final comparison; headline value is flexible-extract; strict-match 0.02 applies a filter that expects the phrase 'The answer is'; system wired peak node A 39525 MB, node B 32448 MB (pre-load baseline 10147/2866 MB, includes OS wired); RDMA protection domains 0/0 before load, 1/1 while serving, 0/0 after detach; request failures 0"}, |
| {"path":"data/run-canonical-run-2026-10-19T08-56-04-8111301-throughput-mlx-community-minimax-m2-7-3bit-ce489acd.parquet","status":"experimental","variant":"MLX 0.32.2 tensor parallel; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-tb5-rdma-minimax-m2-7-3bit-tensor-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err (3 of 3 reached the 128-token cap); concurrent 1 ok/0 err; answered_rate 0.0","caveat":"valid one-off c=1 sequential+concurrent probe; not final comparison; answered_rate 0.0 is expected at this 128-token cap, not a failure: every token budget was spent inside the always-on think block (thinking on, no off switch, no chat_template_kwargs sent); system wired peak node A 63970 MB, node B 55892 MB (pre-load baseline 10303/2889 MB, includes OS wired); RDMA protection domains 0/0 before load, 1/1 while serving, 0/0 after detach; request failures 0","machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","engine":"MLX","engine_version":"0.32.2","campaign_dimensions":{"hardware":{"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","interconnect":"tb5-rdma","world_size":2,"host_ram_limit_gib":null},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"MiniMax-M2.7","id":"mlx-community/MiniMax-M2.7-3bit","hf_repo":"mlx-community/MiniMax-M2.7-3bit","revision_sha":"93d504c2bb57253c3d209b997ee05acf0db90077","quantization":"3-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"MULTI_HOST_RAM_AGGREGATION_UNDEFINED"}}, |
| {"path":"data/run-canonical-run-2026-10-19T09-54-58-8111301-reasoning-mlx-community-minimax-m2-7-3bit-96ea95b3.parquet","status":"experimental","variant":"MLX 0.32.2 tensor parallel; GSM8K 5-shot multiturn, limit 50, thinking on, temperature 1.0 sampled, max_gen_toks 4096","comparison_group":"mlx-tb5-rdma-minimax-m2-7-3bit-tensor-gsm8k","context_band":"8k","max_output_tokens":4096,"completion_state":"gsm8k 50 samples scored; flexible-extract exact_match 0.92 +/- 0.0388; strict-match exact_match 0.82 +/- 0.0549","caveat":"limit-50 subset of GSM8K; not a full-benchmark score; not final comparison; headline value is flexible-extract (0.92); thinking on with no off switch; greedy decoding is a known failure mode for this model: a temperature 0 run scored 0.72 because 13 of 50 answers looped 'Thus answer: N' inside the think block until the 4096-token cap (that run is not published); system wired peak node A 63970 MB, node B 55892 MB (pre-load baseline 10303/2889 MB, includes OS wired); RDMA protection domains 0/0 before load, 1/1 while serving, 0/0 after detach; request failures 0","machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","engine":"MLX","engine_version":"0.32.2","campaign_dimensions":{"hardware":{"machine":"2x Apple M4 Max 128GB (Thunderbolt RDMA)","interconnect":"tb5-rdma","world_size":2,"host_ram_limit_gib":null},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"MiniMax-M2.7","id":"mlx-community/MiniMax-M2.7-3bit","hf_repo":"mlx-community/MiniMax-M2.7-3bit","revision_sha":"93d504c2bb57253c3d209b997ee05acf0db90077","quantization":"3-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.host_ram_limit_gib":"MULTI_HOST_RAM_AGGREGATION_UNDEFINED"}}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"host_ram_limit_gib":100.0},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.5","id":"mlx-community/Qwen3.5-4B-MLX-4bit","hf_repo":"mlx-community/Qwen3.5-4B-MLX-4bit","revision_sha":"32f3e8ecf65426fc3306969496342d504bfa13f3","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T16-33-41-8111301-throughput-mlx-community-qwen3-5-4b-mlx-4bit-2482bd0b.parquet","variant":"MLX 0.32.2 single node; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-single-qwen35-4b-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err; concurrent width=1 1 ok/0 err; answered_rate 1.0; decode 124.48 tok/s median; TTFT 0.65 s median","caveat":"system wired peak 9250.64 MB (pre-load baseline 5577.53 MB, includes OS wired); single node, no RDMA; request failures 0; same-night baseline for the 2-node cluster rows; harness aggregate-decode figures at c>1 include prefill time"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"host_ram_limit_gib":100.0},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.5","id":"mlx-community/Qwen3.5-4B-MLX-4bit","hf_repo":"mlx-community/Qwen3.5-4B-MLX-4bit","revision_sha":"32f3e8ecf65426fc3306969496342d504bfa13f3","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T16-35-08-8111301-reasoning-mlx-community-qwen3-5-4b-mlx-4bit-50c048d8.parquet","variant":"MLX 0.32.2 single node; GSM8K 5-shot multiturn, limit 50, max_gen_toks 1024, thinking off","comparison_group":"mlx-single-qwen35-4b-gsm8k","context_band":"8k","max_output_tokens":1024,"completion_state":"gsm8k 50 samples scored; flexible-extract exact_match 0.86 +/- 0.0496; strict-match exact_match 0.86","caveat":"limit-50 subset of GSM8K; headline is flexible-extract exact_match 0.86; strict-match 0.86 matches flexible-extract; n=50; reported stderr 0.0496; system wired peak 9250.64 MB (pre-load baseline 5577.53 MB, includes OS wired); single node, no RDMA; request failures 0; peak measured during the throughput campaign"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"host_ram_limit_gib":100.0},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.5","id":"mlx-community/Qwen3.5-4B-MLX-4bit","hf_repo":"mlx-community/Qwen3.5-4B-MLX-4bit","revision_sha":"32f3e8ecf65426fc3306969496342d504bfa13f3","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T16-36-57-8111301-reasoning-mlx-community-qwen3-5-4b-mlx-4bit-80292b12.parquet","variant":"MLX 0.32.2 single node; GPQA Diamond CoT zero-shot, limit 50, max_gen_toks 4096, thinking off","comparison_group":"mlx-single-qwen35-4b-gpqa","context_band":"8k","max_output_tokens":4096,"completion_state":"gpqa_diamond_cot_zeroshot 50 samples scored; flexible-extract exact_match 0.52 +/- 0.0714; strict-match exact_match 0.00","caveat":"limit-50 subset of GPQA Diamond; headline is flexible-extract exact_match 0.52; strict-match 0.00 is near zero because the strict filter expects a literal answer phrase these outputs omit; n=50; reported stderr 0.0714; system wired peak 9250.64 MB (pre-load baseline 5577.53 MB, includes OS wired); single node, no RDMA; request failures 0; peak measured during the throughput campaign"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"host_ram_limit_gib":100.0},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Gemma 4","id":"mlx-community/gemma-4-e4b-it-4bit","hf_repo":"mlx-community/gemma-4-e4b-it-4bit","revision_sha":"475b9088d29754a3379866cf5aeb6b41acd313c2","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T16-14-04-8111301-throughput-mlx-community-gemma-4-e4b-it-4bit-24fa634f.parquet","variant":"MLX 0.32.2 single node; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-single-gemma4-e4b-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err; concurrent width=1 1 ok/0 err; answered_rate 1.0; decode 98.72 tok/s median; TTFT 0.43 s median","caveat":"system wired peak 10687.90 MB (pre-load baseline 4468.89 MB, includes OS wired); single node, no RDMA; request failures 0; same-night baseline for the 2-node cluster rows; harness aggregate-decode figures at c>1 include prefill time"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"host_ram_limit_gib":100.0},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Gemma 4","id":"mlx-community/gemma-4-e4b-it-4bit","hf_repo":"mlx-community/gemma-4-e4b-it-4bit","revision_sha":"475b9088d29754a3379866cf5aeb6b41acd313c2","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T16-15-26-8111301-reasoning-mlx-community-gemma-4-e4b-it-4bit-7f267811.parquet","variant":"MLX 0.32.2 single node; GSM8K 5-shot multiturn, limit 50, max_gen_toks 1024, thinking off","comparison_group":"mlx-single-gemma4-e4b-gsm8k","context_band":"8k","max_output_tokens":1024,"completion_state":"gsm8k 50 samples scored; flexible-extract exact_match 0.82 +/- 0.0549; strict-match exact_match 0.54","caveat":"limit-50 subset of GSM8K; headline is flexible-extract exact_match 0.82; strict-match 0.54 reflects format misses; n=50; reported stderr 0.0549; system wired peak 10687.90 MB (pre-load baseline 4468.89 MB, includes OS wired); single node, no RDMA; request failures 0; peak measured during the throughput campaign"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"host_ram_limit_gib":100.0},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Gemma 4","id":"mlx-community/gemma-4-e4b-it-4bit","hf_repo":"mlx-community/gemma-4-e4b-it-4bit","revision_sha":"475b9088d29754a3379866cf5aeb6b41acd313c2","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T16-17-54-8111301-reasoning-mlx-community-gemma-4-e4b-it-4bit-eeab5373.parquet","variant":"MLX 0.32.2 single node; GPQA Diamond CoT zero-shot, limit 50, max_gen_toks 4096, thinking off","comparison_group":"mlx-single-gemma4-e4b-gpqa","context_band":"8k","max_output_tokens":4096,"completion_state":"gpqa_diamond_cot_zeroshot 50 samples scored; flexible-extract exact_match 0.44 +/- 0.0709; strict-match exact_match 0.00","caveat":"limit-50 subset of GPQA Diamond; headline is flexible-extract exact_match 0.44; strict-match 0.00 is near zero because the strict filter expects a literal answer phrase these outputs omit; n=50; reported stderr 0.0709; system wired peak 10687.90 MB (pre-load baseline 4468.89 MB, includes OS wired); single node, no RDMA; request failures 0; peak measured during the throughput campaign"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"host_ram_limit_gib":100.0},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.5","id":"mlx-community/Qwen3.5-9B-OptiQ-4bit","hf_repo":"mlx-community/Qwen3.5-9B-OptiQ-4bit","revision_sha":"890b4c43f99ff392819d83605f7b1e59fa9688aa","quantization":"OptiQ 4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T15-55-46-8111301-throughput-mlx-community-qwen3-5-9b-optiq-4bit-026e4258.parquet","variant":"MLX 0.32.2 single node; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-single-qwen35-9b-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err; concurrent width=1 1 ok/0 err; answered_rate 1.0; decode 65.24 tok/s median; TTFT 1.16 s median","caveat":"system wired peak 14071.80 MB (pre-load baseline 2773.86 MB, includes OS wired); single node, no RDMA; request failures 0; same-night baseline for the 2-node cluster rows; harness aggregate-decode figures at c>1 include prefill time"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"host_ram_limit_gib":100.0},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.5","id":"mlx-community/Qwen3.5-9B-OptiQ-4bit","hf_repo":"mlx-community/Qwen3.5-9B-OptiQ-4bit","revision_sha":"890b4c43f99ff392819d83605f7b1e59fa9688aa","quantization":"OptiQ 4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T16-53-32-8111301-reasoning-mlx-community-qwen3-5-9b-optiq-4bit-fe411e8f.parquet","variant":"MLX 0.32.2 single node; GSM8K 5-shot multiturn, limit 50, max_gen_toks 1024, thinking off","comparison_group":"mlx-single-qwen35-9b-gsm8k","context_band":"8k","max_output_tokens":1024,"completion_state":"gsm8k 50 samples scored; flexible-extract exact_match 0.98 +/- 0.0200; strict-match exact_match 0.98","caveat":"limit-50 subset of GSM8K; headline is flexible-extract exact_match 0.98; strict-match 0.98 matches flexible-extract; n=50; reported stderr 0.0200; system wired peak 14071.80 MB (pre-load baseline 2773.86 MB, includes OS wired); single node, no RDMA; request failures 0; peak measured during the throughput campaign"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"host_ram_limit_gib":100.0},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.5","id":"mlx-community/Qwen3.5-9B-OptiQ-4bit","hf_repo":"mlx-community/Qwen3.5-9B-OptiQ-4bit","revision_sha":"890b4c43f99ff392819d83605f7b1e59fa9688aa","quantization":"OptiQ 4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T16-56-48-8111301-reasoning-mlx-community-qwen3-5-9b-optiq-4bit-6f3286a4.parquet","variant":"MLX 0.32.2 single node; GPQA Diamond CoT zero-shot, limit 50, max_gen_toks 4096, thinking off","comparison_group":"mlx-single-qwen35-9b-gpqa","context_band":"8k","max_output_tokens":4096,"completion_state":"gpqa_diamond_cot_zeroshot 50 samples scored; flexible-extract exact_match 0.60 +/- 0.0700; strict-match exact_match 0.00","caveat":"limit-50 subset of GPQA Diamond; headline is flexible-extract exact_match 0.60; strict-match 0.00 is near zero because the strict filter expects a literal answer phrase these outputs omit; n=50; reported stderr 0.0700; system wired peak 14071.80 MB (pre-load baseline 2773.86 MB, includes OS wired); single node, no RDMA; request failures 0; peak measured during the throughput campaign"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"host_ram_limit_gib":100.0},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"MiMo","id":"mlx-community/MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit","hf_repo":"mlx-community/MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit","revision_sha":"60061fc7499d25ce238e0687154557e504ea690a","quantization":"OptiQ 4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T15-59-13-8111301-throughput-mlx-community-mimo-v2-6-distill-qwen-9b-optiq-4bit-3b1f461e.parquet","variant":"MLX 0.32.2 single node; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-single-mimo-9b-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err; concurrent width=1 1 ok/0 err; answered_rate 1.0; decode 64.82 tok/s median; TTFT 1.16 s median","caveat":"system wired peak 15381.00 MB (pre-load baseline 2754.61 MB, includes OS wired); single node, no RDMA; request failures 0; same-night baseline for the 2-node cluster rows; harness aggregate-decode figures at c>1 include prefill time"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"host_ram_limit_gib":100.0},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"MiMo","id":"mlx-community/MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit","hf_repo":"mlx-community/MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit","revision_sha":"60061fc7499d25ce238e0687154557e504ea690a","quantization":"OptiQ 4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T17-31-31-8111301-reasoning-mlx-community-mimo-v2-6-distill-qwen-9b-optiq-4bit-88f9287b.parquet","variant":"MLX 0.32.2 single node; GSM8K 5-shot multiturn, limit 50, max_gen_toks 1024, thinking off","comparison_group":"mlx-single-mimo-9b-gsm8k","context_band":"8k","max_output_tokens":1024,"completion_state":"gsm8k 50 samples scored; flexible-extract exact_match 0.78 +/- 0.0592; strict-match exact_match 0.78","caveat":"limit-50 subset of GSM8K; headline is flexible-extract exact_match 0.78; strict-match 0.78 matches flexible-extract; n=50; reported stderr 0.0592; system wired peak 15381.00 MB (pre-load baseline 2754.61 MB, includes OS wired); single node, no RDMA; request failures 0; peak measured during the throughput campaign"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"host_ram_limit_gib":100.0},"software":{"operating_system":"macOS 26.6.2","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"MiMo","id":"mlx-community/MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit","hf_repo":"mlx-community/MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit","revision_sha":"60061fc7499d25ce238e0687154557e504ea690a","quantization":"OptiQ 4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T17-34-12-8111301-reasoning-mlx-community-mimo-v2-6-distill-qwen-9b-optiq-4bit-f0dba9f5.parquet","variant":"MLX 0.32.2 single node; GPQA Diamond CoT zero-shot, limit 50, max_gen_toks 4096, thinking off","comparison_group":"mlx-single-mimo-9b-gpqa","context_band":"8k","max_output_tokens":4096,"completion_state":"gpqa_diamond_cot_zeroshot 50 samples scored; flexible-extract exact_match 0.50 +/- 0.0714; strict-match exact_match 0.04","caveat":"limit-50 subset of GPQA Diamond; headline is flexible-extract exact_match 0.50; strict-match 0.04 is near zero because the strict filter expects a literal answer phrase these outputs omit; n=50; reported stderr 0.0714; system wired peak 15381.00 MB (pre-load baseline 2754.61 MB, includes OS wired); single node, no RDMA; request failures 0; peak measured during the throughput campaign"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"host_ram_limit_gib":100.0,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"chassis_container":"macos"},"software":{"operating_system":"macOS 26.6.2 (build 25G83)","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.6","id":"mlx-community/Qwen3.6-35B-A3B-4bit","hf_repo":"mlx-community/Qwen3.6-35B-A3B-4bit","revision_sha":"38740b847e4cb78f352aba30aa41c76e08e6eb46","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T16-01-30-8111301-throughput-mlx-community-qwen3-6-35b-a3b-4bit-62a03765.parquet","variant":"MLX 0.32.2 single node; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-single-qwen36-35b-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err (0 of 3 reached the 128-token cap); concurrent 1 ok/0 err; answered_rate 1.0","caveat":"valid one-off c=1 probe; not a final comparison; same-night baseline for the 2-node cluster rows; harness aggregate-decode figures at c>1 include prefill time; thinking off (enable_thinking false); system wired peak 23984.5 MB (pre-load baseline 2756.73 MB, includes OS wired); single node, no RDMA; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"host_ram_limit_gib":100.0,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"chassis_container":"macos"},"software":{"operating_system":"macOS 26.6.2 (build 25G83)","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.8","id":"mlx-community/Qwen3.8-27B-4bit","hf_repo":"mlx-community/Qwen3.8-27B-4bit","revision_sha":"10c35caafbb80f7dc6a7a432cdd11af10a6d4818","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T16-02-23-8111301-throughput-mlx-community-qwen3-8-27b-4bit-2fd86efa.parquet","variant":"MLX 0.32.2 single node; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-single-qwen38-27b-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err (0 of 3 reached the 128-token cap); concurrent 1 ok/0 err; answered_rate 1.0","caveat":"valid one-off c=1 probe; not a final comparison; same-night baseline for the 2-node cluster rows; harness aggregate-decode figures at c>1 include prefill time; thinking off (enable_thinking false); system wired peak 31374.8 MB (pre-load baseline 2744.36 MB, includes OS wired); single node, no RDMA; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"host_ram_limit_gib":100.0,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"chassis_container":"macos"},"software":{"operating_system":"macOS 26.6.2 (build 25G83)","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"GLM-4.7","id":"mlx-community/GLM-4.7-Flash-4bit","hf_repo":"mlx-community/GLM-4.7-Flash-4bit","revision_sha":"1454cffb1a21737e162f508e5bc70be9def89276","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T16-06-39-8111301-throughput-mlx-community-glm-4-7-flash-4bit-cc146406.parquet","variant":"MLX 0.32.2 single node; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-single-glm47flash-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err (0 of 3 reached the 128-token cap); concurrent 1 ok/0 err; answered_rate 1.0","caveat":"valid one-off c=1 probe; not a final comparison; same-night baseline for the 2-node cluster rows; harness aggregate-decode figures at c>1 include prefill time; thinking off (enable_thinking false); system wired peak 20080.5 MB (pre-load baseline 2767.06 MB, includes OS wired); single node, no RDMA; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"host_ram_limit_gib":100.0,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"chassis_container":"macos"},"software":{"operating_system":"macOS 26.6.2 (build 25G83)","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"gpt-oss","id":"mlx-community/gpt-oss-120b-MXFP4-Q8","hf_repo":"mlx-community/gpt-oss-120b-MXFP4-Q8","revision_sha":"81e5ac3ad0af6efb1298a8e8c7a10ed2990c137b","quantization":"MXFP4-Q8"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T16-07-35-8111301-throughput-mlx-community-gpt-oss-120b-mxfp4-q8-708d9577.parquet","variant":"MLX 0.32.2 single node; throughput probe c=1; 128 output-token cap; reasoning_effort low","comparison_group":"mlx-single-gptoss120b-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err (0 of 3 reached the 128-token cap); concurrent 1 ok/0 err; answered_rate 1.0","caveat":"valid one-off c=1 probe; not a final comparison; same-night baseline for the 2-node cluster rows; harness aggregate-decode figures at c>1 include prefill time; reasoning_effort low (gpt-oss has no thinking-off switch; recorded as low effort, not off); system wired peak 66499.8 MB (pre-load baseline 2771.83 MB, includes OS wired); single node, no RDMA; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"host_ram_limit_gib":100.0,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"chassis_container":"macos"},"software":{"operating_system":"macOS 26.6.2 (build 25G83)","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.5","id":"mlx-community/Qwen3.5-122B-A10B-OptiQ-2bit","hf_repo":"mlx-community/Qwen3.5-122B-A10B-OptiQ-2bit","revision_sha":"8284ab1029a2b0114824175aa7f04740ff318934","quantization":"OptiQ 2-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T16-09-10-8111301-throughput-mlx-community-qwen3-5-122b-a10b-optiq-2bit-ab28ab0d.parquet","variant":"MLX 0.32.2 single node; throughput probe c=1; 128 output-token cap","comparison_group":"mlx-single-qwen35-122b-throughput","context_band":"32k","max_output_tokens":128,"completion_state":"sequential 3 ok/0 err (0 of 3 reached the 128-token cap); concurrent 1 ok/0 err; answered_rate 1.0","caveat":"valid one-off c=1 probe; not a final comparison; same-night baseline for the 2-node cluster rows; harness aggregate-decode figures at c>1 include prefill time; thinking off (enable_thinking false); system wired peak 60625 MB (pre-load baseline 2830.44 MB, includes OS wired); single node, no RDMA; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"host_ram_limit_gib":100.0,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"chassis_container":"macos"},"software":{"operating_system":"macOS 26.6.2 (build 25G83)","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.6","id":"mlx-community/Qwen3.6-35B-A3B-4bit","hf_repo":"mlx-community/Qwen3.6-35B-A3B-4bit","revision_sha":"38740b847e4cb78f352aba30aa41c76e08e6eb46","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T18-04-54.814280-8111301-reasoning-mlx-community-qwen3-6-35b-a3b-4bit-0364f93d.parquet","variant":"MLX 0.32.2 single node; GSM8K 5-shot multiturn, limit 50, max_gen_toks 1024, enable_thinking false","comparison_group":"mlx-single-qwen36-35b-gsm8k","context_band":"8k","max_output_tokens":1024,"completion_state":"gsm8k 50 samples scored; flexible-extract exact_match 0.96 +/- 0.028; strict-match exact_match 0.96","caveat":"limit-50 subset of GSM8K; not a full-benchmark score; not final comparison; headline value is flexible-extract (0.96); strict-match 0.96 matches the headline; n=50 per task, stderr 0.028; thinking off (enable_thinking false); peak measured during the throughput campaign; system wired peak 23984.5 MB (pre-load baseline 2756.73 MB, includes OS wired); single node, no RDMA; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"host_ram_limit_gib":100.0,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"chassis_container":"macos"},"software":{"operating_system":"macOS 26.6.2 (build 25G83)","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.6","id":"mlx-community/Qwen3.6-35B-A3B-4bit","hf_repo":"mlx-community/Qwen3.6-35B-A3B-4bit","revision_sha":"38740b847e4cb78f352aba30aa41c76e08e6eb46","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T18-06-44.043252-8111301-reasoning-mlx-community-qwen3-6-35b-a3b-4bit-9de314be.parquet","variant":"MLX 0.32.2 single node; GPQA Diamond CoT zero-shot, limit 50, max_gen_toks 4096, enable_thinking false","comparison_group":"mlx-single-qwen36-35b-gpqa","context_band":"8k","max_output_tokens":4096,"completion_state":"gpqa_diamond_cot_zeroshot 50 samples scored; flexible-extract exact_match 0.7 +/- 0.065; strict-match exact_match 0.0","caveat":"limit-50 subset of GPQA Diamond; not a full-benchmark score; not final comparison; headline value is flexible-extract (0.7); strict-match 0.0 is lower because the strict filter expects a literal answer phrase the model does not emit; n=50 per task, stderr 0.065; thinking off (enable_thinking false); peak measured during the throughput campaign; system wired peak 23984.5 MB (pre-load baseline 2756.73 MB, includes OS wired); single node, no RDMA; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"host_ram_limit_gib":100.0,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"chassis_container":"macos"},"software":{"operating_system":"macOS 26.6.2 (build 25G83)","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.8","id":"mlx-community/Qwen3.8-27B-4bit","hf_repo":"mlx-community/Qwen3.8-27B-4bit","revision_sha":"10c35caafbb80f7dc6a7a432cdd11af10a6d4818","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T18-26-14.362070-8111301-reasoning-mlx-community-qwen3-8-27b-4bit-4c457fab.parquet","variant":"MLX 0.32.2 single node; GSM8K 5-shot multiturn, limit 50, max_gen_toks 1024, enable_thinking false","comparison_group":"mlx-single-qwen38-27b-gsm8k","context_band":"8k","max_output_tokens":1024,"completion_state":"gsm8k 50 samples scored; flexible-extract exact_match 0.96 +/- 0.028; strict-match exact_match 0.96","caveat":"limit-50 subset of GSM8K; not a full-benchmark score; not final comparison; headline value is flexible-extract (0.96); strict-match 0.96 matches the headline; n=50 per task, stderr 0.028; thinking off (enable_thinking false); peak measured during the throughput campaign; system wired peak 31374.8 MB (pre-load baseline 2744.36 MB, includes OS wired); single node, no RDMA; request failures 0"}, |
| {"campaign_dimensions":{"hardware":{"machine":"Mac Studio M4 Max 128GB","interconnect":null,"world_size":1,"host_ram_limit_gib":100.0,"accelerator_model":"Apple M4 Max GPU","accelerator_memory_gb":128.0,"host_cpu":"Apple M4 Max","host_ram_gb":128.0,"chassis_container":"macos"},"software":{"operating_system":"macOS 26.6.2 (build 25G83)","macos_version":"26.6.2","engine":"MLX","engine_version":"0.32.2","engine_build":"mlx-lm 0.32.2; mlx 0.32.4; Python 3.14","accelerator_runtime":"MLX 0.32.4","backend":"Metal"},"model":{"family":"Qwen3.8","id":"mlx-community/Qwen3.8-27B-4bit","hf_repo":"mlx-community/Qwen3.8-27B-4bit","revision_sha":"10c35caafbb80f7dc6a7a432cdd11af10a6d4818","quantization":"4-bit"},"run":{"prefix_cache":false,"kv_offload_backend":"none","kv_offload_cpu_gib":0.0,"kv_offload_disk_gib":0.0,"weight_cpu_offload_gib":0.0,"moe_cpu_offload":"none","moe_cpu_layers":0,"cache_workload":"none"}},"dimension_null_reasons":{"hardware.interconnect":"not_applicable"},"machine":"Mac Studio M4 Max 128GB","engine":"MLX","engine_version":"0.32.2","status":"experimental","path":"data/run-canonical-run-2026-10-19T18-35-00.317836-8111301-reasoning-mlx-community-qwen3-8-27b-4bit-ff15e577.parquet","variant":"MLX 0.32.2 single node; GPQA Diamond CoT zero-shot, limit 50, max_gen_toks 4096, enable_thinking false","comparison_group":"mlx-single-qwen38-27b-gpqa","context_band":"8k","max_output_tokens":4096,"completion_state":"gpqa_diamond_cot_zeroshot 50 samples scored; flexible-extract exact_match 0.7 +/- 0.065; strict-match exact_match 0.0","caveat":"limit-50 subset of GPQA Diamond; not a full-benchmark score; not final comparison; headline value is flexible-extract (0.7); strict-match 0.0 is lower because the strict filter expects a literal answer phrase the model does not emit; n=50 per task, stderr 0.065; thinking off (enable_thinking false); peak measured during the throughput campaign; system wired peak 31374.8 MB (pre-load baseline 2744.36 MB, includes OS wired); single node, no RDMA; request failures 0"} |
| ] |
| } |
|
|
