[bugfix] pin MeanMetric all_reduce tensor to float64 to avoid cross-rank dtype mismatch

[HaozheZhang6]

PR type

• Bug Fix

PR information

Fixes #9550.

MeanMetric.compute() (swift/metrics/utils.py) builds the all-reduce tensor with an inferred dtype:

tensor = torch.tensor([self.state, self.count], device=self.device)

In sequence-level accuracy, compute_acc() appends np.all(...) (a numpy.bool_). Accumulating that into self.state promotes it from a python float to numpy.float64 on the ranks that take that path. torch.tensor([...]) then infers float64 on those ranks but float32 on ranks where self.state stayed a python float. The ranks enter dist.all_reduce() with mismatched dtypes and the collective hangs (matching the debug screenshots in #9550).

Pinning the tensor to float64 makes every rank reduce with the same dtype regardless of self.state's runtime type. float64 also keeps count exact for large totals (float32 loses integer precision above 2^24).

Experiment results

Reproducing the hang needs a multi-rank run; minimal demo of the dtype inference (one rank's state promoted to numpy.float64 via np.bool_, the other left a python float):

rank-A state type: float | rank-B state type: float64
CURRENT  -> rankA=torch.float32  rankB=torch.float64  all_reduce MATCH=False
FIX      -> rankA=torch.float64  rankB=torch.float64  all_reduce MATCH=True

[gemini-code-assist]

Code Review

This pull request updates the tensor creation in swift/metrics/utils.py to use torch.float64 during distributed all-reduce operations. Feedback suggests changing this to torch.float32 to prevent runtime crashes on hardware backends like NPUs (using HCCL) that do not support double precision for collective operations.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Using torch.float64 for dist.all_reduce can cause runtime errors on certain accelerators and backends. Specifically, Huawei's HCCL (used for NPU execution, which this repository supports) does not support float64 (double) data types for collective operations.\n\nTo ensure compatibility across all supported hardware backends (including NPU/HCCL, GPU/NCCL, and CPU/Gloo), it is safer to use torch.float32. While float32 has a precision limit of 2^24 (approx. 16.7M) for exact integers, this is typically more than sufficient for evaluation sample/token counts, and any potential precision loss beyond that is negligible compared to a hard runtime crash on NPU.

Suggested change

	tensor = torch.tensor([self.state, self.count], dtype=torch.float64, device=self.device)
	tensor = torch.tensor([self.state, self.count], dtype=torch.float32, device=self.device)