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

[Yukun-Cui]

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

Summary

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

Testing

• Reused the existing upstream accuracy tests tests/test_clip.py (-m clip_)
• Validated against reference on the MUSA device; specialization confirmed active via the GEMS_MTHREADS CLIP_ 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).

clip_

dtype	Size	Torch Latency (ms)	Gems Latency (ms)	Speedup
float16	1073741824	3.468320	3.460880	1.002x
float16	64, 64	0.005080	0.004200	1.210x
float16	4096, 4096	0.054520	0.054840	0.994x
float16	64, 512, 512	0.054680	0.055120	0.992x
float16	1024, 1024, 1024	3.467840	3.454600	1.004x
float32	1073741824	6.686300	7.151880	0.935x
float32	64, 64	0.004840	0.004120	1.175x
float32	4096, 4096	0.105040	0.106140	0.990x
float32	64, 512, 512	0.105200	0.106360	0.989x
float32	1024, 1024, 1024	6.663880	7.127280	0.935x
bfloat16	1073741824	3.447560	3.450960	0.999x
bfloat16	64, 64	0.005360	0.003960	1.354x
bfloat16	4096, 4096	0.054160	0.054760	0.989x
bfloat16	64, 512, 512	0.054480	0.054960	0.991x
bfloat16	1024, 1024, 1024	3.460440	3.458040	1.001x

Operator	Arithmetic Mean Speedup
clip_	1.04x

Files Changed

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

[Dongxu-H]

/test | clip_:mthreads

[Dongxu-H]

/test | clip_:mthreads

[github-actions]

❌ On-demand test failed

Operator: | clip_
Runner: mthreads
Backend: mthreads-musa520

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

[Dongxu-H]

/test | clip_:mthreads

[Dongxu-H]

PR includes 140 files changed for a single backend operator, which is unusual. The clip_.py implementation uses grid-stride loop pattern with hardcoded grid dimensions. Logger naming is correct. Implementation handles scalar min/max bounds with appropriate fallback for tensor bounds.

[Dongxu-H]

/test | clip_:mthreads

[Dongxu-H]

LGTM