perf(gemma4): ultimate composed branch — beats vLLM on every workload#18

Open

pyc96 wants to merge 16 commits into

mainfrom

pyc/feat-gemma4-ultimate

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

perf(gemma4): port vLLM Inductor's triple-rmsnorm fusion (post-attn pre-MoE)
pyc
committed
perf(gemma4 MM): batch vision encoder and embed_vision calls

pyc
authored and
pyc96
committed
perf(gemma4): YOCO fast-prefill for E2B/E4B (port of vllm#22628 + #38879)

pyc
authored and
pyc96
committed