Remove PyTorch from the cutedsl dependency extra

[Anerudhan]

Summary

Removes torch and torch-c-dlpack-ext from [project.optional-dependencies].cutedsl, and makes every module under python/cudnn importable without PyTorch. Missing PyTorch is reported only when a PyTorch-dependent entry point is actually invoked.

Implements torch-dependency-removal.md (included in this branch, with review comment C23 applied: the torch import is attempted on first probe, not at helper initialization).

Framework dependency boundary

• New cudnn/_deps/torch_dep.py is the sole runtime import torch. The import is attempted lazily on the first is_available() / require() call and cached (module or failure, never retried), so importing the helper -- and therefore import cudnn -- stays PyTorch-free.
• New public cudnn.TorchNotAvailableError (subclasses ImportError) with the canonical message. The lazy optional-symbol loader re-raises it unwrapped rather than burying it in the generic optional-dependency hint.

Import safety

• 72 modules converted off module-level import torch onto torch_dep.require(), using postponed annotations plus if TYPE_CHECKING: import torch.
• 84 PyTorch-valued parameter defaults became None, resolved immediately after require(), with the fixed effective default stated in the docstring (e.g. `None` means `torch.float32`) and annotations widened to Optional[...]. Effective values are unchanged.
• 23 module-scope PyTorch constants became lazily-resolved accessors.
• wrapper.Graph remains deliberately PyTorch-only (it allocates Torch outputs/workspace and manages the default Torch stream), but now fails fast with TorchNotAvailableError instead of a bare RuntimeError.
• datatypes.is_cutlass_available() no longer requires PyTorch -- a real bug fix: CUTLASS-native cutlass.Numeric types now keep working when CUTLASS is installed and Torch is not.

Custom operators (facade + guarded implementation)

ops/causal_conv1d.py, experimental/ops/sdpa.py and experimental/ops/moe_grouped_matmul.py are now import-safe facades that resolve their public callables on first attribute access, preserving the original function objects and therefore their signatures. The former bodies moved to _causal_conv1d_torch.py / _sdpa_torch.py / _moe_grouped_matmul_torch.py, each wrapped in one top-level if torch_dep.is_available(): block. Module caching means the schemas, fake implementations, compiler hooks and autograd formulas still register exactly once when PyTorch is present, with unchanged operator names. These modules deliberately keep real (non-postponed) annotations, which torch.library.custom_op needs for schema inference.

The two package __init__ files bind from the implementation modules rather than the same-named facade submodules -- importing cudnn.ops.causal_conv1d sets it as an attribute of the parent package and would otherwise shadow the function of the same name.

Packaging

The extra no longer installs PyTorch. Documented install order: (1) the PyTorch build matching your CUDA environment, (2) nvidia-cudnn-frontend[cutedsl], (3) optionally torch-c-dlpack-ext to restore its AOT Torch->DLPack fast path. Omitting (3) affects conversion startup/performance, not numerical behavior.

Test plan

Passing in this branch:

• test/python/tools/audit_no_runtime_torch_import.py -- PASSES across all 226 modules. Rejects runtime torch imports (plain / aliased / from / submodule / literal-dynamic), PyTorch-valued defaults, and unguarded module-scope PyTorch access outside the dependency helper. Type-checking-only imports and the single guarded block in the three implementation modules are permitted.
• python -m compileall python/cudnn -- clean.
• test/python/tools/smoke_no_torch_import.py (new) -- blocks torch with a MetaPathFinder raising ModuleNotFoundError for exactly torch/torch.*, imports every submodule, and asserts representative entry points raise TorchNotAvailableError.

Executed with PyTorch genuinely absent (and the unbuilt pybind module stubbed, since this branch was authored without a build):

• import cudnn succeeds and torch is not in sys.modules afterwards
• 212 submodules import cleanly; none hard-require torch
• is_torch_available() is False; cudnn.TorchNotAvailableError is exported and subclasses ImportError
• cudnn.Graph(), cudnn.ops.{causal_conv1d,b2b_causal_conv1d}, and both experimental ops each raise TorchNotAvailableError with the canonical message
• 5 modules fail for an unrelated pre-existing reason (they need a newer CUTLASS than 4.5.2 -- cannot import name 'CollectorOp'), which reproduces without the torch blocker

Still to run in a proper environment: the clean-venv wheel acceptance check (assert neither distribution appears in pip list), an enable_tvm_ffi=True conversion path with PyTorch but without torch-c-dlpack-ext, and the existing PyTorch-installed test suites.

Remaining before merge

• Docs-inventory sweep for the install instructions listed in the plan (docs/fe-oss-apis/*, benchmark/dsa/README.md, etc.). python/cudnn/__init__.py's optional-dependency hint and pyproject.toml are already updated.
• File the tracked follow-up issue "Add permanent no-Torch import CI coverage" and link it here; the two tools above are intended to become that coverage.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 2279dc3f-00d4-4f93-981e-067c015b77d1

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}