feat(gemma): add official Google QAT TP=4 profile

[Whamp]

Summary

Adds a supported four-card profile for Google's official google/gemma-4-31B-it-qat-w4a16-ct checkpoint:

• slug: vllm/gemma-31b-multi-google-qat-w4a16
• stock vLLM v0.25.1, TP=4, BF16 KV, native 262,144 context
• official Gemma assistant drafter at MTP n=2
• canonical Google chat template and native Gemma ParserEngine
• ⚠️ Production w/ caveats

This solves the missing full-native-context Gemma option for four 3090s using Google's own QAT checkpoint. Compared with the same-session no-MTP baseline, n=2 changes narrative/code decode from 82.45/82.68 to 80.61/92.34 TPS: −2.2% narrative, +11.7% code, approximately +4.7% across the canonical pair. The trade-off is use of all four GPUs and a thin 1.2 GiB/card margin at near-max context.

Closes #814.

Type of change

• New compose variant
• New patch / sidecar
• New script or tool
• New model
• Documentation / benchmark data

Verification

• Actual registered compose booted through switch.sh
• Profile header complete and status/registry drift guard passes
• BENCHMARKS row added
• models/gemma-4-31b/CHANGELOG.md added
• Full rig and validation evidence attached below

MTP n-sweep

Every n=1–4 arm passed 20/20 repeated streamed two-tool calls with both argument objects preserved, confirming vLLM #45588 resolves the legacy Gemma MTP parser failure on v0.25.1.

Arm	Narrative decode	Code decode
no MTP	82.45	82.68
n=1	78.75	86.34
n=2	80.61	92.34
n=3	73.18	85.36
n=4	69.02	84.43

The profile embeds a measured MTP_ACCEPT_MIN=1.8. The generic 2.0 acceptance floor was tuned for deeper Qwen drafting; n=2 is independently throughput-positive and correctness-gated here.

Operational gate

• verify-full.sh: PASS without host overrides
• verify-stress.sh: PASS through 257,544 prompt tokens (98%)
• KV pool: 650,262 tokens / 14.84 GiB; 2.48× native-window concurrency
• VRAM free at 257K: 1,213 MiB/card
• vision smoke: PASS — identified a blue circle centered on red
• soak: 100/100; 0 errors, silent outputs, or VRAM growth; 104.2% retention
• 10K/90K prefill: 981/692 tok/s

Full 8-pack quality

Mode	TC	IF	SO	DE	RM	BF	HA	CLI	Total
thinking off	12/15	14/15	15/15	12/15	13/15	14/15	14/20	22/40	116/150
thinking on	15/15	14/15	14/15	12/15	13/15	15/15	13/20	26/40	122/150

Repository guards

The full scripts/tests/*.sh suite passes except test-switch-explain.sh. That test fails identically on unchanged master on this four-GPU host: vllm/default detects multi4, but Qwen has no vLLM TP=4 DEFAULTS row. No profile-relevant guard fails.

Full report and selected raw logs: https://gist.github.com/Whamp/742569247ce18c234bd3c7170d18b724

Caveats

1. Uses all four GPUs.
2. MTP n=2 trades −2.2% narrative for +11.7% code decode.
3. Vision passes at short context, but vision plus near-max text is unvalidated at the 1.2 GiB/card margin.

Cross-links

• Closes Profile: official Gemma 4 31B Google QAT W4A16 on 4x3090 TP=4 #814
• vLLM #45588: Gemma ParserEngine migration
• vLLM #41745: Gemma MTP support