DP: Fix for torch.compile

[xinyu-intel]

No description provided.

[Copilot]

Pull request overview

This PR implements fixes to enable torch.compile compatibility for data parallel operations on HPU devices. The main issue addressed is that stack-style all-gather operations are incompatible with torch.compile, requiring a switch to concat-style operations.

Key Changes:

• Modified all-gather operations to use concat-style instead of stack-style for torch.compile compatibility
• Added padding logic to ensure token counts are divisible by tensor parallel size for sequence parallel MOE
• Added null checks for DP metadata access to prevent potential errors during compilation

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
vllm_gaudi/v1/worker/hpu_dp_utils.py	Adds padding to ensure num_tokens is divisible by tp_size for sequence parallel MOE
vllm_gaudi/ops/hpu_fused_moe.py	Adds null-safety checks when accessing DP metadata fields
vllm_gaudi/ops/hpu_fp8.py	Adds null-safety checks when accessing DP metadata fields
vllm_gaudi/distributed/device_communicators/hpu_communicator.py	Converts all-gather from stack-style to concat-style for torch.compile compatibility
tests/full_tests/ci_tests.sh	Enables and adds tests for DP2xTP2 configuration with both lazy mode and torch.compile mode

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The code retrieves dp_metadata once but then checks if it's None on every subsequent attribute access. If dp_metadata is None on line 60, all three conditional expressions (lines 61, 64, 67) will set their variables to None, which could cause dispatch_tensor to fail since it may not handle None for the output parameter correctly. Consider adding an early check after line 60 or ensuring that dispatch_tensor properly handles None values.

[Copilot]

Same issue as in hpu_fused_moe.py: if dp_metadata is None, all three variables will be set to None, potentially causing issues in dispatch_tensor. The None handling should be consistent with the function's expectations or an early guard should be added.

[Copilot]

The error message is ambiguous as it doesn't distinguish between the lazy mode test and the torch.compile mode test. The message should specify 'torch.compile mode' to match the test being run.

Suggested change

	echo "Error: Test failed for data parallel size 2 + tensor parallel size 2" >&2
	echo "Error: Test failed for data parallel size 2 + tensor parallel size 2 (torch.compile mode)" >&2

[michalkuligowski]

if dp_size >1 the get_hpu_dp_metadata() should not return None, instead of checking if None, please add assert here

[xinyu-intel]

it's only available under lazy+hpugraph mode since it can reuse the memory.

[github-actions]

✅ CI Passed

All checks passed successfully against the following vllm commit:
17fec3af0942da83bcebe2ca0cb4f6ae81c634d8

[github-actions]

🚧 CI Blocked

The main CI workflow was not started for the following reason:

Your branch is behind the base branch. Please merge or rebase to get the latest changes.

[github-actions]

✅ CI Passed

All checks passed successfully against the following vllm commit:
f21f5ea38c6fa0e824bc00d5762d17e049199cd3