fix(gemma4): only apply swa_full_tokens_ratio=0.15 to MoE variants

[pyc96]

Summary

Refine Patch 2 (#2) so the swa_full_tokens_ratio=0.15
override only applies to MoE Gemma-4 variants
(Gemma-4-26B-A4B-IT), not to dense variants (gemma-4-31B-it,
Gemma-4-E4B-IT).

Stacked on #7 (clamp revert). Staged on pyc96/sglang
only.

Motivation

Patch 2's 0.15 ratio was tuned for the MoE 26B-A4B-IT: MoE sparse
weights leave plenty of GPU memory for KV, so growing the full-attn
pool 3.6x (and shrinking the over-provisioned SWA pool) materially
improves long-context summarization TTFT.

For dense 31B and E4B, the same ratio is harmful:

• 31B-it (62 GB dense weights) + ratio 0.15 → SWA pool = 71808 tokens
  (~70 windows-worth). Concurrent 80-prompt chat → SWA usage 1.00 →
  scheduler stalls admission → output throughput collapses to 1135 tok/s.

This commit gates the override on num_experts > 0 (read from
hf_text_config), the standard SGLang MoE-detection pattern (mirrors
gemma4_causal.py:1166).

Per-model behavior (verified on 1x B200)

model	num_experts	branch	resulting ratio	resulting pool layout
Gemma-4-26B-A4B-IT (MoE)	128	override fires	0.15	full=2138240 swa=320704 (unchanged from Patch 2)
gemma-4-31B-it (dense)	0	override skipped	0.8 (upstream default)	full=132992 swa=106368
Gemma-4-E4B-IT (dense)	0	override skipped	0.8	(same skip path)

Log lines:

# 26B-A4B-IT:
Setting swa_full_tokens_ratio to 0.15 for Gemma4ForConditionalGeneration
(MoE Gemma-4 with num_experts=128; the default ratio over-provisions the
SWA pool and under-provisions the full-attention pool, ...).

# 31B-it:
Keeping default swa_full_tokens_ratio=0.8 for Gemma4ForConditionalGeneration
(dense Gemma-4; MoE-specific 0.15 override skipped to avoid SWA pool
starvation).

Benchmark: 31B-it + trtllm_mha + MTP + 1x B200 vs vLLM nightly

random 40 prompts, max-concurrency=32, seed 1:

chat 1k/1k

metric	SGLang (this PR)	vLLM nightly	delta
outcome	OK 40/40	OK 40/40	same
median TTFT	673 ms	901 ms	SGLang +25%
median TPOT	8.69 ms	9.69 ms	SGLang +10%
P99 TPOT	11.10 ms	13.16 ms	SGLang +16%
total throughput	4715 tok/s	4077 tok/s	SGLang +16%
accept length	3.13	n/a	--

summarization 8k/1k

metric	SGLang (this PR)	vLLM nightly	delta
outcome	OK 40/40	OK 40/40	same
median TPOT	17.02 ms	27.33 ms	SGLang +38%
P99 TPOT	27.13 ms	45.17 ms	SGLang +40%
total throughput	7475 tok/s	6468 tok/s	SGLang +16%
accept length	3.14	n/a	--
median TTFT	13613 ms	10162 ms	vLLM +25% (tradeoff)

Quality (MMLU @ N=500, seed=0, temp=0)

	accuracy
SGLang (this PR, 31B-it)	0.680
vLLM nightly (31B-it)	0.660

Within sampling noise.

Regression test for 26B-A4B-IT

The 26B path is unchanged. Verified by relaunching 26B and confirming:

• log: Setting swa_full_tokens_ratio to 0.15 for ... (MoE Gemma-4 with num_experts=128; ...)
• pool: full_layer_tokens=2138240 swa_layer_tokens=320704 (same as Patch 2)

Tests

test/srt/test_gemma4_swa_full_tokens_ratio.py — 4 cases pass, 2
skipped (full smoke tests skip when env lacks model-config stubs,
same as before):

test_moe_gemma4_default_overridden                       PASSED
test_dense_gemma4_default_preserved                      PASSED
test_user_override_preserved[Gemma4ForCausalLM]          PASSED
test_user_override_preserved[Gemma4ForConditionalGeneration] PASSED
test_full_method_runs_for_moe_gemma4                     SKIPPED (env)
test_full_method_runs_for_dense_gemma4                   SKIPPED (env)

Known limitation (out-of-scope)

The predicate if self.swa_full_tokens_ratio == ServerArgs.swa_full_tokens_ratio
still cannot distinguish "user passed 0.8 explicitly" from "user
didn't pass the flag" (same caveat as the upstream
apply_deepseek_v4_defaults pattern). Fixing this would require a
sentinel default (None) and resolution in __post_init__ — wider
surface area than this PR's MoE gating fix. For MoE Gemma-4, a user
passing 0.8 will still get the override; if they want the upstream
default explicitly, they can pass any other value like 0.81.

---

CI States

Latest PR Test (Base): ❌ Missing run-ci label -- add it to run CI tests.
Latest PR Test (Extra): ❌ Blocked -- run-ci is required first.