onnxruntime
Enable fpA_intB GEMM in CUDA builds and add configurable options
#29622
Merged
Merged
Enable fpA_intB GEMM in CUDA builds and add configurable options #29622
kunal-vaishnavi
merged
15 commits into
microsoft:main
from
tianleiwu:tlwu/20260708/fpa_intb_inmem_autotune
copilot-pull-request-reviewer
commented
on 2026-07-16
[CUDA] In-memory tactic autotuning for fpA_intB MatMulNBits
e4e013d6
Enable onnxruntime_USE_FPA_INTB_GEMM by default when CUDA is enabled
bc32e55a
Add ep.cuda.fpa_intb_gemm / fpa_intb_profile_m session-config keys
8a2ee181
Test ep.cuda.fpa_intb_gemm / fpa_intb_profile_m session-config keys
faf42962
Simplify fpA_intB enable option to boolean; prepacked forces the path
c1cb3320
Fix fpA_intB MatMulNBits CI failures and address review feedback
0ae5e83f
Clarify fpA_intB config-toggle test comment (numeric-equivalence scope)
8ad2e0a3
refactoring
5a8a30df
fix(cuda): use ORT env helper instead of std::getenv in cuda_runtime_…
d72d7a6b
fix Windows CI
eab48120
fix(cuda): use _dupenv_s on Windows to avoid C4996 in cuda_runtime_utils
abc2588c
fix build
6f40ffa2
fix lint warnings
774906f2
address feedbacks
fcf29207
remove stale test
ab7df6ec
kunal-vaishnavi
approved these changes
on 2026-07-17
Reviewers
Assignees
No one assigned
Labels
release:1.28.0
Milestone
No milestone
Login to write a write a comment.
Login via GitHub