Skip to content

[XPU][CI] Disable test_xpu_graph until tc_piecewise is removed (#41634) - #42524

Merged
mingfeima merged 1 commit into
sgl-project:mainfrom
arathi-hlab:fix-xpu-graph-dynamo
Oct 5, 2026
Merged

mingfeima merged 1 commit into
sgl-project:mainfrom
arathi-hlab:fix-xpu-graph-dynamo

Conversation

@arathi-hlab

@arathi-hlab arathi-hlab commented Oct 4, 2026 •

Copy link
Copy Markdown
Collaborator

Motivation

XPU CI (stage-a-test-1-gpu-xpu) fails on every PR in test/registered/xpu/test_xpu_graph.py, e.g. https://github.com/sgl-project/sglang/actions/runs/37169947775/job/111340729252 (PR #40190):

torch._dynamo.exc.Unsupported: Unsupported object.__new__ user-defined class construction
  class=<class 'sglang.srt.layers.layer_boundary.residual.stream.Contribution'>,
  error=object.__new__(Contribution) is not safe, use Contribution.__new__()

The test runs Qwen2.5-1.5B with the tc_piecewise prefill graph, which compiles the decoder with fullgraph=True. Since #42301 the Qwen2 forward builds layer-boundary values that are msgspec.Structs, and Dynamo (torch 2.13) cannot construct a msgspec.Struct inside a compiled region.

Modifications

Per @mingfeima's review (#42524 (comment)): tc_piecewise is being removed in #41634, so this PR no longer reworks the layer-boundary classes. It only marks test_xpu_graph.py as disabled= in its register_xpu_ci(...) call, with a reason pointing at #41634, so run_suite.py skips the file (same mechanism as test/registered/xpu/test_triton_attention_backend.py). The XPU graph test will be refactored once #41634 lands.

The earlier version of this PR (plain __slots__ classes for the per-forward layer-boundary values, rebased from #41845) is dropped.

Verification

  • pre-commit run --files test/registered/xpu/test_xpu_graph.py passes.
  • The AST registry parser (ut_parse_one_file) returns the file with disabled= set, so run_suite.py --hw xpu lists it under skipped tests instead of running it.
  • XPU CI stage-a-test-1-gpu-xpu should be green on this PR.

CI States

Latest PR Test (Base): ⏳ Run #37253885968
Latest PR Test (Extra): ⚠️ Not enabled -- add run-ci-extra label to opt in.
Latest PR Test (AMD ROCm 10): ⏳ Run #37253886036

@mingfeima

Copy link
Copy Markdown
Collaborator

sglang is dropping tc_piecewise graph #41634. I recommend just disable the failed xpu test case in ci right now, and refactor once the above PR landed.

no need to fix upon existing tc piecewise any more.

@mingfeima mingfeima closed this Oct 5, 2026
@mingfeima mingfeima reopened this Oct 5, 2026
@arathi-hlab
arathi-hlab force-pushed the fix-xpu-graph-dynamo branch from 163c2ab to b578001 Compare October 5, 2026 01:57
@arathi-hlab arathi-hlab changed the title [XPU][Fix] Unbreak tc_piecewise prefill graph: build per-forward layer boundary values as plain classes [XPU][CI] Skip test_xpu_graph until tc_piecewise is removed (#41634) Oct 5, 2026
XPU CI (stage-a-test-1-gpu-xpu) fails on every PR in
test/registered/xpu/test_xpu_graph.py: the tc_piecewise prefill graph
compiles the Qwen2 decoder with fullgraph=True, and since sgl-project#42301 the
forward builds msgspec.Struct layer-boundary values that Dynamo (torch
2.13) cannot construct inside a compiled region.

tc_piecewise is being removed in sgl-project#41634, so rather than reworking the
layer-boundary classes for a graph mode that is going away, mark the
test disabled in its CI registration (run_suite.py then skips the file)
and refactor the XPU graph test once sgl-project#41634 lands.
@arathi-hlab arathi-hlab changed the title [XPU][CI] Skip test_xpu_graph until tc_piecewise is removed (#41634) [XPU][CI] Disable test_xpu_graph until tc_piecewise is removed (#41634) Oct 5, 2026
@arathi-hlab
arathi-hlab force-pushed the fix-xpu-graph-dynamo branch from b578001 to 87cb76f Compare October 5, 2026 02:03
@mingfeima mingfeima added intel xpu intel gpu with device `torch.xpu` labels Oct 5, 2026
@mingfeima
mingfeima merged commit b1bbd74 into sgl-project:main Oct 5, 2026
171 of 217 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

documentation Improvements or additions to documentation intel run-ci CI: run the baseline test suite on this PR xpu intel gpu with device `torch.xpu`

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants