Support zero-element graphs (e.g. SDPA with batch size 0)

[Anerudhan]

Fixes #101

Problem

Building a graph containing a tensor with a 0-sized dimension (e.g. SDPA with batch_size = 0, as hit by PyTorch's cuDNN SDPA backend in pytorch/pytorch#133780) fails at backend tensor creation with a generic CUDNN_STATUS_BAD_PARAM, because cuDNN backend tensor descriptors do not accept 0-sized dimensions.

Solution

Detect zero-element graphs during validate(), after shape inference has filled in all tensor dims. If the graph references at least one zero-element tensor and every non-virtual output tensor is zero-element, executing the graph could never write an output byte — so the whole graph is a well-defined no-op:

• build_operation_graph(), create_execution_plans() / create_execution_plan(), check_support(), build_plans() / build_plan_at_index(), and build() all succeed while skipping backend lowering entirely.
• get_workspace_size() reports 0.
• execute() (all overloads), autotune(), and warmup() return success without launching work; variant-pack pointers are ignored (zero-element tensors may be null).
• populate_cuda_graph() leaves the provided CUDA graph empty (valid and instantiable); update_cuda_graph() is a no-op.
• New Graph::is_zero_element_graph() (C++) and pygraph.is_zero_element_graph() (Python) report whether the graph was flagged.

Graphs that mix zero-element tensors with non-zero-element outputs (e.g. matmul contracting over a 0-sized dimension, whose output would need zero-filling) are rejected at validate() with a clear GRAPH_NOT_SUPPORTED message instead of the late backend BAD_PARAM.

Guarded limitations (clear errors instead of misbehavior):

• Plan serialization (serialize(std::vector<uint8_t>&)) errors out — there is no plan to serialize (previously this path would have indexed plans.execution_plans[-1]). JSON structure serialization is unaffected; a graph deserialized from JSON re-detects the no-op in its own validate().
• Runtime shape overrides (override_uids/shapes/strides) on a no-op graph error out rather than silently skipping real work.

Testing

• New Catch2 cases in test/cpp/validate.cpp ([zero_element]): pointwise no-op pipeline (fully deviceless, including execute with a null handle), SDPA with batch 0, and the mixed zero/non-zero rejection. Full tests binary passes: 290 assertions in 27 test cases (cuDNN 9.19.1).
• New pytest test/python/test_zero_element_graph.py: SDPA batch 0 end-to-end via pygraph (mirrors the issue repro), pointwise zero-element, and mixed-graph rejection. Verified on GPU via the pygraph bindings.
• docs/zero-element-graphs.md documents the behavior; linked from llms.txt.
• pre-commit (clang-format, black) clean on changed files.

🤖 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: 25d677bc-48ef-4fef-b885-27b07d77ae3a

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

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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