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

[Yukun-Cui]

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

Summary

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

Testing

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

silu_backward

dtype	Size	Torch Latency (ms)	Gems Latency (ms)	Speedup	TFLOPS
float16	1073741824	4.484840	4.669640	0.960x	0.230
float16	64, 64	0.006560	0.003040	2.158x	0.001
float16	4096, 4096	0.109920	0.105760	1.039x	0.159
float16	64, 512, 512	0.110280	0.106560	1.035x	0.157
float16	1024, 1024, 1024	4.480120	4.618220	0.970x	0.233
float32	1073741824	9.036880	9.022460	1.002x	0.119
float32	64, 64	0.005560	0.008680	0.641x	0.000
float32	4096, 4096	0.187520	0.180520	1.039x	0.093
float32	64, 512, 512	0.187720	0.180560	1.040x	0.093
float32	1024, 1024, 1024	9.038420	9.009620	1.003x	0.119
bfloat16	1073741824	4.474480	4.788980	0.934x	0.224
bfloat16	64, 64	0.006800	0.003400	2.000x	0.001
bfloat16	4096, 4096	0.110040	0.108120	1.018x	0.155
bfloat16	64, 512, 512	0.110480	0.109000	1.014x	0.154
bfloat16	1024, 1024, 1024	4.467360	4.682340	0.954x	0.229

Operator	Arithmetic Mean Speedup
silu_backward	1.12x

Files Changed

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

[Dongxu-H]

/test | silu_backward:mthreads

[Dongxu-H]

/test | silu_backward:mthreads

[github-actions]

❌ On-demand test failed

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

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

[Dongxu-H]

LGTM

[Dongxu-H]

LGTM

[Dongxu-H]

/test | silu_backward:mthreads