Claude a8ca42b59c perf: hgemm_warptiling Config B — beats cublas on MoE-sized GEMM (0.7x) ...

probe_k10_configs.sh results on BI-V100:
  256x4096 @ 4096x11008:
    cublas:   10.554 ms
    Config B:  7.649 ms (0.7x cublas — FASTER)
    Config A:  2308 ms  (old broken config)

Config B: BM128 BN128 BK16 WM64 WN64 WNITER2 TM8 TN4 NT128
Root cause of Config A slowness: WN=128 WNITER=4 caused
excessive register pressure and smem bank conflicts.

2026-08-15 05:11:32 +00:00

..

bindings

feat: hgemm_warptiling.cu — siboehm kernel 10 ported to WARPSIZE=64 FP16

2026-08-14 17:05:59 +00:00

headers

fix: MoE kernel include paths — device_utils.cuh + arch_condition.h

2026-08-14 11:31:54 +00:00

moe

fix: MoE kernel include paths — device_utils.cuh + arch_condition.h

2026-08-14 11:31:54 +00:00

activation.cu

fix: activation.cu torch/extension.h + LOG(FATAL)→TORCH_CHECK, reshape_paged_cache.cu torch header

2026-08-14 10:52:54 +00:00

block_copy.cu

fix: add ceil_div + DEVICE_INLINE to device_utils.cuh

2026-08-14 11:11:15 +00:00

hgemm_blocktiling.cu

refactor: hgemm_blocktiling.cu — strict 1:1 from siboehm kernel 6

2026-08-14 16:24:22 +00:00

hgemm_warptiling.cu

perf: hgemm_warptiling Config B — beats cublas on MoE-sized GEMM (0.7x)

2026-08-15 05:11:32 +00:00

norm.cu

fix: remove glog/cuda_ops_api.h dependency from all xllm CUDA kernels

2026-08-14 10:30:10 +00:00

norm.cu.orig

fix: adapt xllm norm.cu for corex CUB (CUDA 10.2)

2026-08-14 10:19:51 +00:00

reshape_paged_cache.cu

fix: activation.cu torch/extension.h + LOG(FATAL)→TORCH_CHECK, reshape_paged_cache.cu torch header

2026-08-14 10:52:54 +00:00

rope.cu

fix: remove glog/cuda_ops_api.h dependency from all xllm CUDA kernels

2026-08-14 10:30:10 +00:00