DSA: probe Paddle, not sys.modules["torch"], for framework metadata

[ForFishes]

Follow-up to #12. 1.28.0's framework-neutral metadata refactor (python/cudnn/tensor_adapter.py, plus the _torch() helper in python/cudnn/api_base.py) identifies the framework with sys.modules.get("torch"). That probe does not work on this branch, and the failure is not subtle: both DSA indexer backward wrappers are unusable.

Why the probe fails here

paddle.enable_compat(scope={"cudnn"}) rewrites only the module-level import torch of the modules it loads — it never registers the shim in sys.modules. Meanwhile a real PyTorch is routinely installed alongside Paddle in the same environment (in ours: PyTorch 2.12 + Paddle 3.4). So sys.modules["torch"] resolves to real PyTorch, which describes neither Paddle's dtypes nor its devices, strides or current stream.

Note that a function-level import torch does not help either, for the same reason — the fix has to name Paddle explicitly.

Symptoms fixed

• APIBase._check_dtype took its "not a dtype" branch for every Paddle dtype, because isinstance(paddle.bfloat16, torch.dtype) is False for real torch:

  ValueError: Expected dtype to be a torch.dtype or list, got <class 'paddle.dtype'>
    at python/cudnn/api_base.py, from IndexerBackward.check_support()
  

  This fires in check_support(), i.e. before compile()/execute(), so indexer_backward_wrapper and dense_indexer_backward_wrapper both fail unconditionally. In our downstream test suite that was ~60 failing DSA/indexer tests.

• tensor_adapter.detect_framework returned "unknown" for Paddle tensors. Two silent downgrades follow: default_stream handed back the CUDA legacy default stream instead of Paddle's current stream (a correctness hazard for anything overlapping with framework work), and get_strides fell through to a C-contiguous guess instead of reading the real strides.

• _raw_stream / _get_default_stream read .cuda_stream off the stream object. Paddle exposes the raw handle as stream_base.raw_stream — the same handle deepseek_sparse_attention/utils/runtime.resolve_stream already uses, so the two now agree.

• TensorDesc.is_contiguous dereferenced torch.contiguous_format unconditionally; Paddle does not define it. The formats are now looked up defensively so the default (memory_format=None) path cannot break.

Verification

Downstream (PaddleFleet DSA/indexer/CSA, single card, B30Z sm_103), 22 test files:

• before this patch: test_cudnn_dsa_indexer_bwd 16 failed / 19 passed, test_hybrid_mla_warmup_kl_cudnn 28 failed / 5 passed, test_mqa_latent_attention 8 failed, test_indexer_loss_overlap 5 failed, test_block_sparse_dsa_gradcheck 13 failed, test_hysparse_mqa_gather_dsa 6 failed, test_cudnn_dsa_indexer_docmask 3 failed;
• after: all of the above pass. The one remaining failure in that suite is a stale negative-control assertion that fails identically on paddle/v1.27.0, so it is unrelated to 1.28.0.

We also added a host-side regression test downstream that asserts the probe resolves to Paddle, that _check_dtype accepts Paddle dtypes in both scalar and list form, that Paddle tensors are recognised by the adapter, that strides are read rather than guessed, and that default_stream matches Paddle's current stream. Four of its seven cases fail without this patch.

Perf on the same setup is unaffected by this patch; for the record, paddle/v1.28.0 is a solid win over paddle/v1.27.0 on the sparse DSA indexer backward at our production shape (b=1, s_q=s_k=8192, h=64, d=128, topk=2048): 4.757 ms -> 3.095 ms end to end.

pre-commit (black, line length 160) is clean.