Skip to content

add triton fallback for mi455 gptoss & dsfp4#2657

Merged
valarLip merged 7 commits intomainfrom
455_test
Apr 17, 2026
Merged

add triton fallback for mi455 gptoss & dsfp4#2657
valarLip merged 7 commits intomainfrom
455_test

Conversation

@HaonanWang98
Copy link
Copy Markdown
Contributor

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

@HaonanWang98 HaonanWang98 requested a review from a team April 8, 2026 10:14
@github-actions
Copy link
Copy Markdown
Contributor

github-actions bot commented Apr 8, 2026

🏷️ CI Guide

Runs automatically on every PR:

  • ✅ Pre-checks (submodule verification, code formatting)
  • ✅ Aiter op tests (gfx942 + gfx950)
  • ✅ Triton tests (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label Tests
ci:triton-355 Run Triton tests on MI355 in addition to MI325
ci:sglang SGLang integration tests
ci:atom ATOM benchmark (DeepSeek-R1 + GPT-OSS)
ci:vllm vLLM benchmark
ci:all All of the above

Add labels via the sidebar or gh pr edit 2657 --add-label <label>

@azaidy
Copy link
Copy Markdown
Contributor

azaidy commented Apr 8, 2026

@HaonanWang98 Have all these configs been tuned on gfx1250?

Copy link
Copy Markdown
Contributor

@azaidy azaidy left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If configs have not been tuned on gfx1250 please only create a fallback config per triton kernel. Do not create dedicated NK configs unless they are guaranteed to be best performing

@HaonanWang98 HaonanWang98 force-pushed the 455_test branch 2 times, most recently from 385ed1d to e8bacb0 Compare April 15, 2026 08:53
Comment thread aiter/jit/core.py Outdated
if hip_version > Version("6.2.41133"):
flags_hip += ["-mllvm -amdgpu-coerce-illegal-types=1"]
if get_gfx() == "gfx950" and int(os.getenv("AITER_FP4x2", "1")) > 0:
if (get_gfx() == "gfx950" or get_gfx() == "gfx1250") and int(
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

change to gfx942

Comment thread aiter/ops/mha.py Outdated
q, k, v = [maybe_contiguous(x) for x in (q, k, v)]

if can_impl_fmha_v3_fwd():
can_impl_fmha_v3_fwd_triton = (
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

...

Comment thread aiter/ops/mha.py
)


def flash_attn_varlen_func(
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

...

@valarLip valarLip merged commit 67f3217 into main Apr 17, 2026
45 of 46 checks passed
@valarLip valarLip deleted the 455_test branch April 17, 2026 10:18
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants