[PyTorch] Fuse eager EP prepare and dispatch into a single op

[phu0ngng]

Description

This allows us to pin-host the recv_total_tokens, which saves the cost of D2H sync.

TODO: Do so for non-Eagermode as well.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Greptile Summary

The PR fuses eager expert-parallel prepare and dispatch so the receive count can be written to pinned host memory and used to allocate outputs without a separate device-to-host copy.

• Adds a fused C++/pybind eager prepare-and-dispatch operation.
• Routes eager BF16 and MXFP8 dispatch through a new autograd wrapper.
• Moves eager receive-count storage to pinned host memory.
• Adds MXFP8 eager dispatch backward coverage.

Confidence Score: 4/5

The PR is not yet safe to merge because eager EP dispatch still breaks or is rejected by torch.compile full-graph compilation.

Eager dispatch unconditionally invokes a plain pybind function without a torch.library or fake registration, so the previously reported compilation failure remains reachable.

Files Needing Attention: transformer_engine/pytorch/ep.py and transformer_engine/pytorch/csrc/extensions/ep.cpp

Important Files Changed

Filename	Overview
transformer_engine/pytorch/ep.py	Adds the fused eager autograd path, but its direct pybind forward call leaves the previously reported torch.compile incompatibility outstanding.
transformer_engine/pytorch/csrc/extensions/ep.cpp	Implements and binds the fused pinned-host receive-count, allocation, and dispatch sequence.
transformer_engine/pytorch/csrc/extensions.h	Declares the new fused eager dispatch binding consistently with its C++ implementation.
tests/pytorch/distributed/run_ep.py	Adds eager MXFP8 forward/backward gradient coverage but does not exercise torch.compile.

Sequence Diagram

sequenceDiagram
  participant Python as ep_dispatch
  participant Autograd as _EpPrepareAndDispatchEager
  participant Binding as pybind extension
  participant CUDA as EP backend
  Python->>Autograd: apply(tokens, routing, pinned count)
  Autograd->>Binding: ep_prepare_and_dispatch_eager(...)
  Binding->>CUDA: prepare routing and write receive count
  CUDA-->>Binding: pinned-host receive count
  Binding->>Binding: synchronize and allocate outputs
  Binding->>CUDA: dispatch tokens
  CUDA-->>Python: receive tensors and expert counts

Loading

_{Reviews (2): Last reviewed commit: "[PyTorch] Fuse eager EP prepare and disp..." | Re-trigger Greptile}