Weight gradient kernels for dense and MoE models

[zheanxu]

This Pull Request introduces deepgemm.wgrad_gemm_fp8_fp8_fp32_nt and k_grouped_wgrad_gemm_fp8_fp8_fp32_nt, optimized weight gradient kernels for dense and MoE models. These kernels achieve a ~20% speedup compared to the internal CUTLASS implementation.

For detailed usage, refer to the function documentation.

Weight gradient GEMMs for dense models

M	N	K	Opti BMxBN	Computation (TFLOPS)	Memory Bandwidth (GB/s)
7168	2112	4096	128x152	920	507
1536	24576	4096	128x152	986	582
512	32768	4096	128x152	878	1086
16384	7168	4096	128x152	994	342
7168	4096	4096	128x152	942	411
2048	7168	4096	128x152	920	513
7168	2112	8192	128x152	1052	451
1536	24576	8192	128x152	1092	511
512	32768	8192	128x152	1014	1129
16384	7168	8192	128x152	1079	240
7168	4096	8192	128x152	1061	333
2048	7168	8192	128x152	1037	452

Grouped weight gradient GEMMs for MoE models

Groups	M	N	K	Opti BMxBN	Computation (TFLOPS)	Memory Bandwidth (GB/s)
4	7168	4096	4096	128x152	939	409
4	2048	7168	4096	128x152	900	502
4	7168	4096	8192	128x152	1044	328
4	2048	7168	8192	128x152	1033	450
8	7168	4096	4096	128x152	942	411
8	2048	7168	4096	128x152	902	503

[LyricZhao]

I plan to merge it after #94, thanks!

[LyricZhao]

These kernels achieve a ~20% speedup compared to the internal CUTLASS implementation.

To clarify, you can refer to the profile-data repo for internal CUTLASS impl performance comparison.

[hxdtest]

m = 4096 
n = 1024
k = 4096


x = torch.ones((m, k), device='cuda', dtype=torch.bfloat16)
y = torch.rand((n, k), device='cuda', dtype=torch.bfloat16)
 
ref_out =  (x.float() @ y.float().t())
out1 = ref_out.clone() 
x_fp8 = per_token_cast_to_fp8(x)
y_fp8 = per_token_cast_to_fp8(y)
 
 
deep_gemm.wgrad_gemm_fp8_fp8_fp32_nt(x_fp8, y_fp8, out1)
print(out1/ref_out)  

why the difference is nearly twice ?

out1/ref_out
tensor([[1.9995, 2.0002, 2.0004,  ..., 2.0006, 1.9999, 2.0009]

--update--
it should be out1 = torch.zeros_like(ref_out)

[zheanxu]

@hxdtest Thank you very much for your feedback.
During backpropagation, W needs to accumulate W_grad, so deep_gemm.wgrad_gemm_fp8_fp8_fp32_nt(x, y, out) was designed to perform out += [email protected]() instead of out = [email protected](). This detail was omitted in the documentation.

[hxdtest]

@hxdtest Thank you very much for your feedback. During backpropagation, W needs to accumulate W_grad, so deep_gemm.wgrad_gemm_fp8_fp8_fp32_nt(x, y, out) was designed to perform out += [email protected]() instead of out = [email protected](). This detail was omitted in the documentation.

Thank you for your reply. After fix the test code, the results are close.

[hxdtest]

@hxdtest Thank you very much for your feedback. During backpropagation, W needs to accumulate W_grad, so deep_gemm.wgrad_gemm_fp8_fp8_fp32_nt(x, y, out) was designed to perform out += [email protected]() instead of out = [email protected](). This detail was omitted in the documentation.

Fantastic Work！I used DeepGemm and built a fp8 Linear layer to replace torch.nn.Linear and run a rl job. It seems evaluation scores with mixed fp8 precision are close to scores with mixed bf16 experiment.

[ajWithNucleus]

@hxdtest can you please share your Linear layer wrapper as a quick start util. It will be helpful.

[hxdtest]

@hxdtest can you please share your Linear layer wrapper as a quick start util. It will be helpful.

https://github.com/hxdtest/fp8_verl/blob/add_fp8/verl/third_party/deep_gemm/fp8_linear.py

[LyricZhao]

Thanks for your work on the FP8 linear module. But the implements have lots of unfused kernels, e.g. per_token_cast_to_fp8 and kernels inside the call of the DeepGEMM function, which may lead to an overall lower performance.

Just a reminder if you care about the end-to-end performance :)