9c723eeb29
Read all 20+ key source files in one pass: - 27 CCCL tuning headers (17000+ lines) with 199 benchmark annotations - 29 muh tuning headers (3618 lines) with BI-V100 adaptations - gen_patch.py (409 lines) — C++ injection DEAD, Triton injection ALIVE - muh_kernel_map.py (400+ lines) — CCCL→vllm algorithm mapping - muh_dispatch.py (200+ lines) — runtime policy dispatch - bench_bi100.py (713 lines) — PyTorch-based CCCL benchmark runner - prefix_prefill.py (895 lines) — Triton prefill kernel - paged_attn.py (794 lines) — decode attention dispatch - qwen3_5.py (588 lines) — Qwen3.6 MoE model adapter - computility-run.yaml, baseline.muh, Dockerfile Key findings: - C++ injection path is dead (no .cu source in enginex) - Real optimization paths: Triton params, model adapter, vllm Python config - CCCL value is parameter space knowledge + benchmark data patterns - bench_bi100.py is ready to run on Phanthy Cloud for real data - paged_attn.py line 99 use_v1=True disables V2 for long sequences