From 9effd44ccb7838c129af202dca5cfc177e5714ef Mon Sep 17 00:00:00 2001 From: Chenhan Yu Date: Mon, 15 Jun 2026 18:56:10 +0000 Subject: [PATCH] =?UTF-8?q?[OMNIML-5025]=20cell=5Ft0=5Fd7=20=E2=80=94=20ge?= =?UTF-8?q?mma-4-E4B-it=20MTP=20container=20fix?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Switch container in tools/launcher/examples/google/gemma-4-E4B-it/ specdec_bench_mtp_vllm.yaml from vllm/vllm-openai:v0.22.1 → vllm/vllm-openai:gemma. The generic v0.22.1 image fails Gemma 4 MTP engine startup; the Gemma-specific image is what the sister cell (OMNIML-5024 cell_t0_d3) ran successfully on. Per PR #1564, this benchmark family does NOT use per-cell _cells/.yaml files — cell-specific params are passed as pipeline.task_N.args+=[...] overrides at submit time, not committed to the repo. The parent yaml + the args-override CLI is the complete contract. Real run with this container: experiment cicd_1781548540, AL_qualitative_overall=3.2945, AL_throughput_32k_overall=3.3803. Surfaced 2026-06-15 (intern-agent job 341631795). Supersedes #1736. Signed-off-by: Chenhan Yu Co-Authored-By: pensieve-intern agent Signed-off-by: Chenhan Yu --- .../gemma-4-E4B-it/specdec_bench_mtp_vllm.yaml | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-) diff --git a/tools/launcher/examples/google/gemma-4-E4B-it/specdec_bench_mtp_vllm.yaml b/tools/launcher/examples/google/gemma-4-E4B-it/specdec_bench_mtp_vllm.yaml index d5d7f365c35..e44c5ee526c 100644 --- a/tools/launcher/examples/google/gemma-4-E4B-it/specdec_bench_mtp_vllm.yaml +++ b/tools/launcher/examples/google/gemma-4-E4B-it/specdec_bench_mtp_vllm.yaml @@ -1,13 +1,13 @@ # SPEED-bench MTP speculative-decoding run for gemma-4-E4B-it via vLLM. # # Gemma 4 MTP support landed in vLLM PR vllm-project/vllm#41745 (2026-05-06) -# and is in ``vllm/vllm-openai:v0.22.1`` (and later). Gemma 4 MTP uses a -# separate assistant model passed via ``--draft_model_dir``; vLLM -# auto-detects Gemma 4 from the assistant and does NOT take a ``method`` -# key in ``speculative_config``. The wrapper at -# ``examples/specdec_bench/specdec_bench/models/vllm.py`` routes to the -# assistant-model config shape when ``--speculative_algorithm MTP`` is -# paired with ``--draft_model_dir``. +# and is in the Gemma-specific vLLM image. Gemma 4 MTP uses a +# separate assistant model passed via ``--draft_model_dir``. The wrapper at +# ``examples/specdec_bench/specdec_bench/models/vllm.py`` emits +# ``method=mtp`` plus the assistant model for this family. +# +# Use ``vllm/vllm-openai:gemma`` for this family; generic vLLM images can +# fail Gemma 4 MTP startup or long-context throughput runs. # # Assistant model: ``google/gemma-4-E4B-it-assistant`` (public, ungated). # @@ -52,7 +52,7 @@ pipeline: nodes: 1 ntasks_per_node: 1 gpus_per_node: 1 - container: vllm/vllm-openai:v0.22.1 + container: vllm/vllm-openai:gemma # task_1: SPEED throughput_32k split task_1: @@ -80,4 +80,4 @@ pipeline: nodes: 1 ntasks_per_node: 1 gpus_per_node: 1 - container: vllm/vllm-openai:v0.22.1 + container: vllm/vllm-openai:gemma