onnxruntime
[DO NOT MERGE] mnb verify m
#29467
Closed

[DO NOT MERGE] mnb verify m #29467

jambayk wants to merge 13 commits into main from jambayk/mnb-verify-m
jambayk
jambayk Add batched small-M GEMV for CUDA MatMulNBits 4-bit and 8-bit
c8564004
jambayk Add accuracy_level=4 int8 dp4a MatMulNBits CUDA path for small-M verify
f2599963
jambayk Optimize int8 dp4a verify path: K-unrolling + per-block activation quant
93a781b5
jambayk Support block_size 32 in fpA_intB W4A16 MatMulNBits path
ca8a0729
jambayk Fix fpA_intB GEMV small-M occupancy (CtaN=8->4 for m=1,2,3)
3f7063c1
jambayk Make fpA_intB MatMulNBits prefill fast across block sizes
d3535e1b
jambayk Bring fpA_intB MatMulNBits prefill to parity for small-N and group_si…
350de0b3
jambayk Revert int8-activation (accuracy_level=4) MatMulNBits path
940d7c76
jambayk Route fpA_intB small-N large-M through dequant+cuBLAS to fix the M>=1…
c0941e5a
jambayk Add standard-layout batched verify GEMV for MatMulNBits (no prepacking)
8964560c
jambayk Tune verify GEMV: M-aware CtaN and CtaM=6
2fa7f4c2
jambayk Revert "Tune verify GEMV: M-aware CtaN and CtaM=6"
a4cc2f30
jambayk Merge branch 'main' into jambayk/mnb-verify-m
98a09732
jambayk jambayk closed this 12 days ago
jambayk jambayk deleted the jambayk/mnb-verify-m branch 12 days ago

Login to write a write a comment.

Login via GitHub

Reviewers
No reviews
Assignees
No one assigned
Labels
Milestone