perf(gemma4): add Gemma4ForConditionalGeneration to mm_disabled_models#10

Draft

pyc96 wants to merge 15 commits into

mainfrom

pyc/sota-gemma4-31b-mm-disabled

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
perf(gemma4 MTP H100): tune Triton extend tile for Lq=256 / sm_90
pyc96
committed
JIT custom_all_reduce/tp_qknorm: use reinterpret_cast not std::bit_cast
pyc96
committed

Commits on May 24, 2026

perf(gemma4): add Gemma4ForConditionalGeneration to mm_disabled_models
pyc96
committed