perf(gemma4): ULTIMATE v2 -- ties or beats vLLM no-MTP on 3 models (31B-it, 26B-A4B, E4B), MMLU tied#21

Draft

pyc96 wants to merge 23 commits into

mainfrom

pyc/feat-gemma4-ultimate-v2

Commits on May 23, 2026

debug: trtllm_mha page_table bounds-check (SGLANG_TRTLLM_MHA_DEBUG=1)
pyc96
committed
debug: SWA allocator OOB instrumentation (companion to bounds-check trap)
pyc96
committed
fix(trtllm_mha): clamp page_table to k_cache page range to prevent SWA crash
pyc96
committed
fix(trtllm_mha + FROZEN_KV_MTP): swap SWA-aware state with target pool
pyc96
committed
Revert "fix(trtllm_mha): clamp page_table to k_cache page range"
pyc96
committed
fix(gemma4): only apply swa_full_tokens_ratio=0.15 to MoE variants
pyc96
committed

Commits on May 24, 2026

perf(gemma4): close triton-attn TPOT gap (fused PLE tail + piecewise CG opt-in)
pyc
committed

Commits on May 25, 2026