[PyTorch] FP8 MHA with RoPE and Miscellaneous Improvements

[yaox12]

Description

1. FP8 MHA with RoPE. Handle Float8Tensor inputs depending on the dtype instead of the fp8_mha flag in DotProductAttention. fp8_mha still ensures the output of DPA is in FP8. With this PR:
   • FP8 DPA workflow (unchanged): LayerNormLinear -> DPA (cast BF16 input to FP8, FP8 DPA, cast output to BF16) -> Linear
   • FP8 MHA workflow:
     • Without RoPE (unchange): LayerNormLinear (output in FP8) -> DPA (FP8 DPA) -> (FP8 input) Linear.
     • With RoPE (new): LayerNormLinear (output in BF16) -> Apply RoPE (output in BF16) -> DPA (cast BF16 input to FP8, FP8 DPA) -> (FP8 input) Linear
2. Rename is_first_module_in_mha to fp8_output and add this flag to LayerNormLinear, otherwise even the LayerNormLinear in MLP (after MHA) would produce FP8 outputs when fp8_mha=True.
3. Avoid index_select ops in cast_to_fp8.
4. Avoid index_select ops in FP8 DPA. I only modified the fwd functions because in backward the CPU overheads are not exposed.
5. Changed the way we check strides of k and v to avoid creating PyTorch tensors.
6. Move transpose to backward for Float8Tensor inputs in Linear.

Timeline

As we can see, this PR greatly reduces the CPU overheads in red boxes.

• Before
  
• After
  

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refractor

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[timmoon10]

Overall I like this approach. The current FP8 MHA impl is brittle since it expects the modules to pass specific combinations of Float8Tensor/torch.Tensor. Adding logic so the modules can do casts internally makes this more flexible.

This is similar to how I envision the operation-based API to work. See how we cast inputs in the linear operation:

TransformerEngine/transformer_engine/pytorch/ops/basic/basic_linear.py

Lines 465 to 488 in ec49a52

	ifwith_fp8_computeandnotis_float8_tensor(x_local):
	fp8_dtype=get_fp8_te_dtype(
	input_fp8_meta["recipe"],
	fprop_tensor=True,
	)
	x_fp8=Float8Tensor(
	data=torch.empty_like(x_local, dtype=torch.uint8),
	fp8_meta=input_fp8_meta,
	fp8_meta_forward=True,
	fp8_meta_index=0,
	fp8_dtype=fp8_dtype,
	fp8_scale_inv=torch.empty([1], dtype=torch.float32, device=device),
	dtype=dtype,
	)
	with_cast_transpose=weight.requires_grad
	iftensor_parallel_mode=="column"andsequence_parallel:
	with_cast_transpose=False
	ifwith_cast_transpose:
	x_fp8.cast_transpose_(x_local)
	else:
	x_fp8.copy_(x_local)
	x_local=x_fp8
	elifnotwith_fp8_computeandis_float8_tensor(x_local):
	x_local=x_local.from_float8()

We're not there yet, but the goal is to be able to implement FP8 MHA with something like:

model=te.Sequential(
te.ops.LayerNorm(...), # fp8 outputte.ops.Linear(...),
te.ops.RoPE(...), # fp8 outputte.ops.SelfAttention(...), # fp8 outputte.ops.Linear(...),
)
withte.fp8_autocast():
y=model(x)

[timmoon10]

Regarding further optimizations: removing the select operations would be helpful if it's not too difficult. I've observed that they add non-trivial CPU overhead in other cases, so I recommend looking at #865. You should also be aware that I've made significant changes in the cpp_extensions functions in #1083.

The logic for torch.ops.tex_ts is for ONNX exports, which is based on TorchScript. We only register the ops needed for inference, so that's why FP8 cast is registered with TorchScript while FP8 cast-transpose is a plain Pybind11 function

[yaox12]

Regarding further optimizations: removing the select operations would be helpful if it's not too difficult. I've observed that they add non-trivial CPU overhead in other cases, so I recommend looking at #865. You should also be aware that I've made significant changes in the cpp_extensions functions in #1083.

The logic for torch.ops.tex_ts is for ONNX exports, which is based on TorchScript. We only register the ops needed for inference, so that's why FP8 cast is registered with TorchScript while FP8 cast-transpose is a plain Pybind11 function

Thanks for your explanation. cast_to_fp8 is doing index selection in Torchscript and thus calling PyTorch ops. I'll try to move it to plain C++.

[timmoon10]

/te-ci pytorch

[timmoon10]

LGTM

[cyanguwa]

LGTM

[cyanguwa]

@yaox12 do we have a test that particularly tests the functionality of FP8 MHA + RoPE? The test should be able to answer your question above as well, regarding the FP8GlobalStateManager.get_fp8_recipe().fp8_mha. Thanks.

[yaox12]

@yaox12 do we have a test that particularly tests the functionality of FP8 MHA + RoPE? The test should be able to answer your question above as well, regarding the FP8GlobalStateManager.get_fp8_recipe().fp8_mha. Thanks.

Thanks. Added RoPE tests.

[cyanguwa]

/te-ci pytorch

[yaox12]

@cyanguwa I find Flash Attention 3 is not installed in our CI container, so I just skip the FP8 DPA/MHA tests when FA3 is not available, otherwise they will throw the error "no attention backends available".

Another CI failure is tests/pytorch/fused_attn/test_fused_attn.py::test_dpa_mask[mask_9_0-model_configs0-dtype0], and it failes in other PRs too. Seems to be a CI issue.

[cyanguwa]

@cyanguwa I find Flash Attention 3 is not installed in our CI container, so I just skip the FP8 DPA/MHA tests when FA3 is not available, otherwise they will throw the error "no attention backends available".

Another CI failure is tests/pytorch/fused_attn/test_fused_attn.py::test_dpa_mask[mask_9_0-model_configs0-dtype0], and it failes in other PRs too. Seems to be a CI issue.

Yes, mask_9_0 is a cuDNN 9.4.0.47 issue, and the FP8 tests are getting fixed in #1141.

[yaox12]

@timmoon10 Can you review above unresolved comments?

[timmoon10]

LGTM

[yaox12]

@timmoon10@cyanguwa Can you trigger the CI?

[yaox12]

/te-ci pytorch

[yaox12]

/te-ci pytorch

[yaox12]

As Tim and Charlene have approved, and all comments have been resolved, and the CI has passed, I'll merge this PR.