Gemma 4 31B current-template optimization probe¶
Captured: 2026-07-17
Host/topology: Primary Node, Docker Desktop on WSL2; one lab serve on the RTX PRO 6000 Blackwell Max-Q (96 GB, 300 W power limit)
Engine: vllm/vllm-openai:v0.25.1, V1 runner, FP8 KV cache, Triton attention selected by vLLM
Model: google/gemma-4-31B-it-qat-w4a16-ct at a766e9afa44931dfa9ff5de90af9494ca193e74c; current base tokenizer/template from google/gemma-4-31B-it at b9ea41a2887d8607f594846523f94c6cc75ac8a4
Outcome¶
The current official 31B QAT W4A16 continuous-batching checkpoint is healthy with the new Google chat template at 128K and passed the smoke plus structured-JSON preflight with thinking disabled. Its warmed, single-request long generation was 62.3 tok/s: two equal 512-token responses took 8.270 s and 8.171 s end-to-end. A prior 1,024-token run took 17.710 s (57.8 tok/s), including its first request's Triton JIT warm-up.
The 128K capacity probe recorded 74.97 s TTFT, but generated only 41 completion tokens, so it is useful as a prefill observation rather than a decode-rate comparison. The local TTFT is behind the current external single RTX PRO 6000 NVFP4 reports (about 46--48 s at 128K), although those reports use a different NVFP4 checkpoint and likely a higher-power workstation card. This Max-Q card is power-limited to 300 W, so the external number is an advisory target rather than a like-for-like regression threshold.
No production recipe, router policy, or tier changed. The managed 12B Heavy serve remains the rollback/production target and is restored after this lab capture.
Native MTP result: blocked for this target¶
vLLM 0.25.1 successfully recognized method=mtp and loaded Google's
google/gemma-4-31B-it-qat-q4_0-unquantized-assistant (one speculative token). It then failed
during engine profiling with an incompatible projection dimension: target activations of 6400
versus assistant weights expecting 10752. This is a real incompatibility between the W4A16
continuous-batching target and that Q4 assistant, not a quoting or startup failure.
Consequently, do not add this assistant to the W4A16 recipe. Google's QAT documentation pairs the Q4 assistant with the matching Q4 unquantized target; testing that distinct, much larger target is a separate capacity and quality experiment, not a drop-in speed flag.
Reproduction and raw evidence¶
The long-generation suite asks for a trailing marker. Gemma continued emitting x through the
entire cap and omitted it, so the deterministic marker check correctly failed. Both warmed
attempts nevertheless returned exactly 512 completion tokens with finish_reason: length; their
recorded API timing is valid diagnostic throughput evidence, not a quality pass.
Research implications¶
- Google's current template's thinking-off primer is a behavior/control change, not an expected
source of this scale of decode slowdown. Retain the pinned current tokenizer revision and send
chat_template_kwargs={"enable_thinking": false}for matched disabled-thinking benchmarks. - At multi-turn boundaries, strip prior thought-channel content as required by Google's template guidance; retain it only across tool-call subturns. This avoids unnecessarily inflating context and preserves tool semantics.
- WSL2 is not the immediate explanation for the 128K gap: models are on an ext4 Docker volume,
the recipe already enables
VLLM_WSL2_ENABLE_PIN_MEMORY=1, and this model is fully GPU resident. Updating WSL and Docker Desktop is worthwhile for stability and GPU fixes, but raising WSL system-memory limits will not directly make GPU decode faster. Treat pinned-memory changes as a separate controlled A/B because WSL/WDDM page-locked memory can pressure the host.
Sources¶
- Google, Gemma thinking control, accessed 2026-07-17.
- Google, Gemma 4 prompt formatting, accessed 2026-07-17.
- vLLM, MTP speculative decoding, accessed 2026-07-17.
- Millstone AI, 31B NVFP4 RTX PRO 6000 reference, advisory external comparison, accessed 2026-07-17.
External results are not a local quality or promotion result. The official checkpoint/template and the exact local artifact links above are the decision evidence for this finding.