Skip to content

[ROCm][Perf] gfx942: use FlyDSL fp8 MQA logits kernel (ROCm/aiter#3913) #49544

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
tjtanaa merged 2 commits into from
Aug 16, 2026
+2 −9
Merged

[ROCm][Perf] gfx942: use FlyDSL fp8 MQA logits kernel (ROCm/aiter#3913) #49544

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0405394
[ROCm][Perf] gfx942: use FlyDSL fp8 MQA logits kernel (ROCm/aiter#3913)
akii96 Jul 22, 2026
df116ba
Merge branch 'main' into aakif/flydsl-fp8-mqa-logits-gfx942
tjtanaa Aug 16, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 2 additions & 9 deletions vllm/v1/attention/ops/rocm_aiter_mla_sparse.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -563,21 +563,14 @@ def rocm_fp8_mqa_logits(
Logits tensor of shape [M, N], dtype `torch.float32`.
"""

# TODO(ganyi): Temporarily workaround, will remove the module check and reference
# path after aiter merge this kernel into main
from vllm._aiter_ops import rocm_aiter_ops

k_fp8, scale = kv

# Temporarily route gfx942 to the vendored ROCm/aiter#3257 workaround.
# Remove this branch once vLLM bumps AITER to a version that includes
# ROCm/aiter#3257.
if _ON_GFX942 and rocm_aiter_ops.is_enabled():
from vllm.v1.attention.ops.triton_fp8_mqa_logits import (
fp8_mqa_logits_gfx942,
)
from aiter.ops.flydsl import flydsl_fp8_mqa_logits

return fp8_mqa_logits_gfx942(
return flydsl_fp8_mqa_logits(
q, k_fp8, scale, weights, cu_seqlen_ks, cu_seqlen_ke
)

Expand Down
Loading