The GitHub Actions job "CI" on tvm.git/yanxu/cudnn-sdpa-fixes has failed.
Run started by GitHub user YangXu1990uiuc (triggered by YangXu1990uiuc).

Head commit for run:
d76e4c620672928dfce7c762c739e0ff37b49d7b / Yang Xu <[email protected]>
[Relax][cuDNN] Reject unsupported attention offloads; fix default scale

The cuDNN BYOC backend matched relax.nn.attention on the op node alone, so
attention carrying a causal mask was offloaded to a cuDNN SDPA graph built
with .set_causal_mask(false) and silently computed bidirectionally.

- python/tvm/relax/backend/patterns.py:339: annotate the attention call so the
  partition check can read the op attributes.
- python/tvm/relax/backend/cuda/cudnn.py:89-97: do not offload when causal_mask
  or window_size is set, or when the input is not float16 (the runtime only
  builds a half-precision graph, attention.cc:42).
- src/runtime/extra/contrib/json/json_node.h:261-274: add HasAttrValue, since a
  None attribute is serialized as an empty string and HasAttr cannot tell. It is
  paired with GetAttr<T>, which throws on a type mismatch, so an attribute we
  cannot read is loud rather than silently replaced by the default.
- src/runtime/extra/contrib/cudnn/cudnn_json_runtime.cc:218-231: use it so the
  1/sqrt(head_size) default scale is actually applied, and reject masked
  attention that reaches the runtime.
- python/tvm/relax/testing/attention.py:79: optional causal_mask argument.
- tests/python/relax/test_codegen_cudnn.py:302-329: partition tests for the
  causal and fp32 cases, plus a positive control.

Report URL: https://github.com/apache/tvm/actions/runs/30531783693

With regards,
GitHub Actions via GitBox


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to