[CUDA] Enable SM80 FP4 QMoE grouped GEMM #29442
tianleiwu
wants to merge
27 commits into
main
from
tlwu/20260625/qmoe_fp4
Split-K2 SwiGLU GEMV
3de24c68
experiment of fp16 accumulation for split-k
bc4712b4
Merge remote-tracking branch 'origin/main' into tlwu/20260619/qmoe_sp…
4836d23c
env vars for fp32_accum and splik routing
64db1ce7
Merge remote-tracking branch 'origin/main' into tlwu/20260619/qmoe_sp…
ec607998
merge branch tlwu/qmoe_fpt_opt
9bdb7a11
enable gemv for fp4 by default
664cfa6c
add triton experiment
782f27df
WFP4A16 for FP16 with opt-in ORT_ENABLE_FP4_CUTLASS_GEMM=1
eaaf111c
safe FP4 QMoE CUTLASS gate
319088ac
native FP4 scale-layout fix
f7fb8aad
Dual prepack
52cd3e4a
add doc for env vars
1747aae1
EXP: FP4 GEMV converter packs 2 nibbles into one 32-bit store (bit-id…
0fb02a5f
docs: record FP4 GEMV converter 2-nibble store experiment (TBD, GPU c…
9777e64a
Update doc with re-run results
f678d88e
EXPERIMENT(qmoe_fp4): opt-in FP4 fc1 split-K + fp32-accum SwiGLU GEMV…
e8df8f4d
Update split-k experiment result
acdbadab
[CUDA] FP4 MoE GEMV: opt-in interleaved layout with dtype-conditional…
38e3e241
[CUDA] FP4 MoE: native WFP4A16 prefill for non-256-aligned shapes via…
16e868c4
add analysis of gemm kernel
6f3692a3
SM80 WFP4A16 kernel
f2cc2e4e
enable SM80 fp4 kernel by default
17694aa4
update WFP4A16 for A100 (FP16)
32d2fa7a
SM80 W-FP4-A-BF16 kernel
f15d8375
Fix FP4 SM80 PR lint cleanup
4a8104d2
Document BF16 SM80 FP4 A100 sweep
b5ada5b3
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub