Alexsssu's picture
H200 (Hopper) ablation: vLLM 0.18 -> 0.26 -> 0.29; companion to the B200 study
ebe9d4b verified
|
Raw History Blame Contribute Delete
4.1 kB

Canonical run selection

For every (vLLM version × model) cell, this file states which measurement is authoritative, and what happened to everything else. make_tables.py implements exactly this selection in its LEG_DIRS constant; if you change one, change the other.

The rule

The canonical measurement is the run made under the full scenario table (patches/vllm_only.yaml) in a clean, isolated, serial environment. Where a model could not run there for a fixable dependency reason, the targeted re-run with that dependency pinned is canonical instead.

On H200 this is simpler than on B200: only one model needed a pinned re-run.

Canonical run per cell

Cell contents are run-id suffixes; the full run-id is vllm-<version>-vllm-only-h200<suffix>.

model 0.18.0 0.26.0 0.29.0
Llama-3.1-8B-Instruct -clean -clean -clean
Mixtral-8x7B-Instruct-v0.1 -clean -clean -clean
gpt-oss-120b -clean -clean -clean
mamba-2.8b-hf -clean -clean -clean
Mamba-Codestral-7B-v0.1 -clean -clean -clean
Qwen3-Next-80B-A3B-Instruct -clean -clean -clean
AI21-Jamba-Mini-1.7 -clean -clean -clean
Kimi-Linear-48B-A3B-Instruct -clean -clean -clean
whisper-large-v3 -clean -clean -clean
Qwen2-VL-7B-Instruct -clean -clean -clean
Qwen3-VL-8B-Instruct -clean -clean -clean
Qwen3-VL-235B-A22B-Instruct-FP8 -clean -clean -clean
Qwen2.5-Omni-7B -clean -clean -clean
gemma-4-26B-A4B-it no result — FAIL -tfpin -clean
GLM-5.2-NVFP4 SKIP SKIP SKIP

That is 41 canonical results.json: 13 + 14 + 14.

The two cells without a number

GLM-5.2-NVFP4 — skipped on all three legs, not failed. The harness refuses it on Hopper with skip nvidia/GLM-5.2-NVFP4: NVFP4 is Blackwell-only. It never launched, so it has no run.log and no results.json anywhere. It should not be counted as a failure, and it is excluded from the report's tables entirely — which is why the coverage denominator is 14, not 15.

gemma-4-26B on 0.18 — a genuine failure. vLLM 0.18 contains no Gemma4 implementation and transformers 4.57.6 cannot parse the config. Its run.log is in logs.tar.gz under the 0.18 leg and is the evidence for that attribution.

Why -tfpin for gemma-4 on 0.26

It failed in the clean sweep on an AmbiguousGlobalPerLayerAttributeError from transformers 5.17.0, not because vLLM 0.26 lacks support. The re-run pinned transformers==5.14.1 — the version this project's pyproject.toml declares for 0.26 — and it passed. Everything else in that venv is identical to the clean 0.26 leg.

No cutlass-pinned backfill was needed on H200. On B200 six models required nvidia-cutlass-dsl==4.5.3; here all six pass unpinned. See "Failure attribution" in REPORT.md.

Runs that exist but are excluded

Only one, and it is not in this archive:

run why excluded
everything under /code/users/alexsu/vllm-validate/ That tree is the B200 experiment, a separate study with its own archive. The H200 sweep was deliberately written to /code/users/alexsu/vllm-validate-h200/ so the two could not mix.

There was no contaminated or voided round on H200. The clean sweep ran once, start to finish, and the only follow-up was the single-model gemma-4 pin.

How to re-derive the coverage

python3 - <<'PY'
import json, glob, os
LEG = {"0.18.0": ["vllm-0.18.0-vllm-only-h200-clean"],
       "0.26.0": ["vllm-0.26.0-vllm-only-h200-clean",
                  "vllm-0.26.0-vllm-only-h200-tfpin"],
       "0.29.0": ["vllm-0.29.0-vllm-only-h200-clean"]}
for leg, dirs in LEG.items():
    models = {json.load(open(p))["model"]
              for d in dirs for p in glob.glob(f"results/{d}/*/results.json")}
    print(f"{leg}: {len(models)}/14 with results")
PY

Expected: 0.18.0: 13/14, 0.26.0: 14/14, 0.29.0: 14/14.