[KernelGen][MThreads] Add permute_copy Moore Threads specialized operator

[Yukun-Cui]

[KernelGen][MThreads] Add permute_copy Moore Threads specialized operator

Summary

Add a Moore Threads (MUSA) specialized Triton kernel for permute_copy, overriding the generic
implementation via runtime.replace_customized_ops(). Implements the permute copy operation.

Testing

• Reused the existing upstream accuracy tests tests/test_permute_copy.py (-m permute_copy)
• Validated against reference on the MUSA device; specialization confirmed active via the
  GEMS_MTHREADS PERMUTE_COPY debug log
• Falls back to the generic implementation for unsupported dtype/device/shape
  (fp64/int64 are not supported on Moore Threads hardware)

Performance

Compared against the generic FlagGems implementation on Moore Threads (MUSA).

permute_copy

dtype	Size	Torch Latency (ms)	Gems Latency (ms)	Speedup
float16	64, 128	0.005640	0.002760	2.043x
float16	128, 256	0.002760	0.002720	1.015x
float16	32, 64, 128	0.004240	0.005240	0.809x
float32	64, 128	0.006000	0.002760	2.174x
float32	128, 256	0.002640	0.002760	0.957x
float32	32, 64, 128	0.005240	0.005320	0.985x
bfloat16	64, 128	0.005640	0.002760	2.043x
bfloat16	128, 256	0.002760	0.002760	1.000x
bfloat16	32, 64, 128	0.004240	0.005280	0.803x

Operator	Arithmetic Mean Speedup
permute_copy	1.31x

Files Changed

• src/flag_gems/runtime/backend/_mthreads/ops/permute_copy.py: Moore Threads Triton kernel + fallback
• src/flag_gems/runtime/backend/_mthreads/ops/__init__.py: Register import and __all__

[Dongxu-H]

/test | permute_copy:mthreads

[github-actions]

❌ On-demand test failed

Operator: permute_copy
Runner: mthreads
Backend: mthreads-musa520

The test failed to complete. Check the workflow run for details.

[Dongxu-H]

/test | permute_copy_moore_threads_specialized:mthreads

[github-actions]

❌ On-demand test failed

Operator: permute_copy_moore_threads_specialized
Runner: mthreads
Backend: mthreads-musa520

The test failed to complete. Check the workflow run for details.

[Dongxu-H]

Logger 命名不符合 MThreads 后端规范。当前使用的是主文件夹写法 logging.getLogger(name)，应该改为 MThreads 后端专用写法：logger = logging.getLogger(fflag_gems.runtime.backend._mthreads.ops.{name[-1]})。

[Dongxu-H]

/test | permute_copy:mthreads

[Dongxu-H]

LGTM

[Yukun-Cui]

2026.07.27已被移出待特化列表