[KernelGen][MThreads] Add log10_ Moore Threads specialized operator

[Yukun-Cui]

[KernelGen][MThreads] Add log10_ Moore Threads specialized operator

Summary

Add a Moore Threads (MUSA) specialized Triton kernel for log10_, overriding the generic implementation via runtime.replace_customized_ops(). Implements the log10 operation.

Testing

• Reused the existing upstream accuracy tests tests/test_log10.py (-m log10_)
• Validated against reference on the MUSA device; specialization confirmed active via the GEMS_MTHREADS LOG10_ 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).

log10_

dtype	Size	Torch Latency (ms)	Gems Latency (ms)	Speedup	TFLOPS
float16	4096, 4096	0.061200	0.056880	1.076x	0.295
float16	64, 512, 512	0.055780	0.056400	0.989x	0.297
float16	1024, 65536	0.206600	0.213280	0.969x	0.315
float32	4096, 4096	0.108480	0.106800	1.016x	0.157
float32	64, 512, 512	0.107520	0.105960	1.015x	0.158
float32	1024, 65536	0.414460	0.413960	1.001x	0.162
bfloat16	4096, 4096	0.055520	0.056900	0.976x	0.295
bfloat16	64, 512, 512	0.055840	0.056520	0.988x	0.297
bfloat16	1024, 65536	0.208000	0.213220	0.976x	0.315

Operator	Arithmetic Mean Speedup
log10_	1.00x

Files Changed

• src/flag_gems/runtime/backend/_mthreads/ops/log10_.py: Moore Threads Triton kernel + fallback
• src/flag_gems/runtime/backend/_mthreads/ops/__init__.py: Register import and __all__

[Dongxu-H]

/test | log10_:mthreads

[github-actions]

❌ On-demand test failed

Operator: | log10_
Runner: mthreads
Backend: mthreads-musa520

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

[Dongxu-H]

/test | log10_:mthreads

[Dongxu-H]

/test | log10_:mthreads

[Dongxu-H]

PR has merge conflicts that must be resolved before merging. The mthreads ops init.py imports log10_ from both .log10 and .log10_ modules, which could cause confusion. Consider removing the duplicate import from .log10 if log10_ is intended to use the specialized implementation.

[Dongxu-H]

常数命名 LOG10_2 容易产生歧义，建议改为 LOG10_OF_2 或类似更清晰的名称。Arithmetic Mean Speedup 为 1.00x，几乎没有性能提升，请确认是否需要保留这个特化实现。

[Dongxu-H]

In init.py, both from .log10 import log10, log10_, log10_out and from .log10_ import log10_ are imported. This creates duplicate imports of log10_ which could be confusing. The specialized backend log10_ should override the default one, but the import order matters. Consider removing log10_ from the log10 import line to make the override explicit.

[Dongxu-H]

/test | log10_:mthreads