From 0e36863fd1210c0a2412ca19d65a5282fc80a293 Mon Sep 17 00:00:00 2001 From: Chenhan Yu Date: Mon, 15 Jun 2026 11:35:12 -0700 Subject: [PATCH] =?UTF-8?q?launcher:=20gemma-4-E4B-it=20MTP=20=E2=80=94=20?= =?UTF-8?q?switch=20vLLM=20image=20to=20vllm-openai:gemma?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The generic vllm-openai:v0.22.1 image fails Gemma 4 MTP engine startup with an assertion error. The Gemma-specific image (vllm-openai:gemma) ships the patches Gemma 4 MTP needs; same image was used successfully on the sister cell (OMNIML-5024 cell_t0_d3, branch pensieve-intern/OMNIML-5023/t0_d3). Surfaced 2026-06-15 on OMNIML-5025 cell_t0_d7 (intern-agent job 341631795): agent submitted with v0.22.1, task 0 failed at vLLM engine init, agent diff'd against sister branch and patched locally. This commit applies the same patch to main so the next cell to dispatch doesn't re-derive the workaround. Signed-off-by: Chenhan Yu --- .../specdec_bench_mtp_vllm.yaml | 19 ++++++++++--------- 1 file changed, 10 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..0fc599d19c2 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 +1,14 @@ # 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 +# (``v0.22.1`` and similar) fail Gemma 4 MTP engine startup. Surfaced +# 2026-06-15 on OMNIML-5025 cell_t0_d7 (intern-agent job 341631795). # # Assistant model: ``google/gemma-4-E4B-it-assistant`` (public, ungated). # @@ -52,7 +53,7 @@ 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 +81,4 @@ pipeline: nodes: 1 ntasks_per_node: 1 gpus_per_node: 1 - container: vllm/vllm-openai:v0.22.1 + container: vllm/vllm-openai:gemma