Repository navigation

fix(w8a16): correct MMA A-fragment register mapping + MoE improvements - #125

Merged

m96-chan merged 21 commits into

mainfrom

feature/v0.2.18

Dec 30, 2025

Merged

fix(w8a16): correct MMA A-fragment register mapping + MoE improvements#125
m96-chan merged 21 commits into
mainfrom
feature/v0.2.18

Commits on Dec 28, 2025
- perf(fp8): add cached scale buffer variants (v5-v8) for FP8xFP8 GEMM
  
  m96-chan
  and
  claude
  committed
  772c10c
- feat(gemv): add accurate FP8/FP8 GEMV kernel (#123)
  
  m96-chan
  and
  claude
  committed
  7232145
- perf(gemv): optimize accurate FP8 kernel to match fast version speed
  
  m96-chan
  and
  claude
  committed
  09aed7b
- Merge main: fix GEMV error rates and add rel-error-check skill
  m96-chan
  committed
  e88f282
- fix(test): correct W8A16 scale format and update error rate
  
  m96-chan
  and
  claude
  committed
  92dabaf
- perf(gemv): add optimized BF16 GEMV kernel with B[N,K] layout
  
  m96-chan
  and
  claude
  committed
  247db16
- feat(llm): use optimized BF16 GEMV in LinearBF16 layer
  
  m96-chan
  and
  claude
  committed
  65d8f1d
Commits on Dec 29, 2025
- fix(matmul): W8A16 GEMM scalar kernel LUT initialization
  
  m96-chan
  and
  claude
  committed
  0e1ad49
- wip(moe): disable system prompt in chat_cli_moe.py
  
  m96-chan
  and
  claude
  committed
  5877cf0
- fix(w8a16): correct MMA A-fragment register mapping for m16n8k16
  
  m96-chan
  and
  claude
  committed
  39ee5c9
- test(moe): add MoE inference test for various prompt lengths
  
  m96-chan
  and
  claude
  committed
  ef4b3a0
- chore: sync working state for FP8 MoE inference
  
  m96-chan
  and
  claude
  committed
  66a3d6c
- style: format test_moe_inference.py with ruff
  
  m96-chan
  and
  claude
  committed
  f105ab2
- docs: update README with v0.2.18 GEMV benchmarks
  
  m96-chan
  and
  claude
  committed
  22748fa
- test: skip MoE inference test in CI
  
  m96-chan
  and
  claude
  committed
  a3043d6
- test: skip FP8 accurate GEMV tests in CI
  
  m96-chan
  and
  claude
  committed
  8746f98
- fix: use correct function name has_native_module
  
  m96-chan
  and
  claude
  committed
  801b900
- bench: update GEMV benchmarks with latest measurements
  
  m96-chan
  and
  claude
  committed
  47a2d06
- docs: update GEMV benchmarks with optimized BF16 kernel
  
  m96-chan
  and
  claude
  committed
  162ad95
- refactor(gemv): switch gemv_bf16 to optimized B[N,K] kernel
  
  m96-chan
  and
  claude
  committed
  c451f67
- refactor(gemv): remove old BF16 GEMV kernel (B[K,N] layout)
  
  m96-chan
  and
  claude
  committed
  971a68b

Commits

Commits on Dec 28, 2025

Commits on Dec 29, 2025