Skip to content

[BACKEND][TLE] add sunrise backend with tle - #726

Merged
zhzhcookie merged 6 commits into
flagos-ai:triton_v3.6.xfrom
xdali-adora:sunrise/triton_v3.6.x_with_tle_os
Jun 30, 2026
Merged

[BACKEND][TLE] add sunrise backend with tle#726
zhzhcookie merged 6 commits into
flagos-ai:triton_v3.6.xfrom
xdali-adora:sunrise/triton_v3.6.x_with_tle_os

Conversation

@xdali-adora

@xdali-adora xdali-adora commented Jun 29, 2026

Copy link
Copy Markdown
Contributor

This PR adds sunrise backend based on Triton3.6 in FlagTree and supports the major tle extentions.

Main changes:

  • add sunrise backend across third_party
  • support AABS for sunrise
  • support tle primitives
  • support shared_memory hint
  • add related tests and examples

TLE-Lite

Python DSL 状态
tle.load(is_async=True) ✅已支持
tle.cumsum ✅已支持
tle.extract_tile ✅已支持
tle.insert_tile ✅已支持
#@hint: shared_memory ✅已支持
tle.pipe ❌未支持
tle distributed ❌未支持

TLE-Struct (GPU)

Python DSL 状态
tle.gpu.alloc ✅已支持
tle.gpu.local_ptr ✅已支持
tle.gpu.copy (without TMA) ✅已支持
tle.gpu.memory_space ✅已支持
tle.gpu.copy (with TMA) ❌未支持
tle.gpu.pipeline ❌未支持
tle.gpu.warp_specialize ❌未支持

TLE-Raw

Python DSL 状态
mlir extension ❌未支持

@CLAassistant

CLAassistant commented Jun 29, 2026

Copy link
Copy Markdown

CLA assistant check
All committers have signed the CLA.

Comment thread .github/workflows/sunrise-build-and-test.yml Outdated
Comment thread lib/Dialect/Triton/Transforms/RewriteTensorPointer.cpp Outdated
Comment thread python/setup_tools/setup_helper.py Outdated
Comment thread python/setup_tools/setup_helper.py Outdated
Comment thread python/setup_tools/setup_helper.py Outdated
Comment thread python/setup_tools/setup_helper.py Outdated
Comment thread python/setup_tools/setup_helper.py Outdated
Comment thread python/setup_tools/setup_helper.py Outdated
@sunnycase

Copy link
Copy Markdown
Collaborator

Thanks for the contribution and for adding TLE support for the sunrise backend.

Could you please update the PR description or add supporting documentation to explain which TLE primitives are implemented/supported by this work, and include performance benefit data so reviewers can evaluate whether the implementation scope matches the expected value?

It would be helpful to include:

  • The list of implemented TLE primitives, their semantic coverage, and any partial support or known limitations.
  • The lowering/runtime path for each key primitive, especially where it differs from the native Triton path.
  • Performance data: benchmark cases, input sizes, hardware/driver environment, baseline, before/after results, improvement ratio, and any regression cases.
  • If this PR is currently only structural enablement and has no measurable performance gain yet, please state that explicitly and describe the follow-up validation plan.

Comment thread setup.py Outdated
Comment thread python/setup_tools/setup_helper.py
Comment thread CMakeLists.txt Outdated
Comment thread CMakeLists.txt Outdated
@xdali-adora
xdali-adora force-pushed the sunrise/triton_v3.6.x_with_tle_os branch from df25a4a to 67517a8 Compare June 29, 2026 11:54

@zhzhcookie zhzhcookie left a comment

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.

LGTM

@zhzhcookie
zhzhcookie merged commit 9f4c91e into flagos-ai:triton_v3.6.x Jun 30, 2026
12 of 13 checks passed
cjh9368 pushed a commit to cjh9368/FlagTree that referenced this pull request Jul 14, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

5 participants