[WIP][Common, PyTorch] Improve mHC to match DeepSeek's implementation#2953
Draft
kainzhong wants to merge 2 commits intoNVIDIA:mainfrom
Draft
[WIP][Common, PyTorch] Improve mHC to match DeepSeek's implementation#2953kainzhong wants to merge 2 commits intoNVIDIA:mainfrom
kainzhong wants to merge 2 commits intoNVIDIA:mainfrom
Conversation
…plementation Signed-off-by: Kaining Zhong <kainingz@nvidia.com>
for more information, see https://pre-commit.ci
kainzhong
changed the title
kainzhong
changed the title
kainzhong
commented
May 1, 2026
|
|
||
| @staticmethod | ||
| def forward(ctx, f, bias, H_post, x, H_res, n, use_tf32=True): | ||
| def forward(ctx, f, bias, H_post, x, H_res, n, use_tf32=True, fuse_grad_x_acc=True): |
Collaborator
Author
There was a problem hiding this comment.
Make it default to False, and also add this to other two functions to make users aware
| return grad_f, grad_bias, grad_H_post, grad_x, grad_H_res, None, None | ||
| if ctx.fuse_grad_x_acc: | ||
| # When fused x gradient accumulation is enabled, use fp32 for the accumulation buffer | ||
| x.untyped_storage().grad_x_acc = grad_x.to(torch.float32) |
Collaborator
Author
There was a problem hiding this comment.
Check if this is overwriting anything that is already there
| M = s * b | ||
|
|
||
| grad_x = torch.empty_like(x) | ||
| fuse_grad_x_acc = hasattr(x.untyped_storage(), "grad_x_acc") |
Collaborator
Author
There was a problem hiding this comment.
Add fuse_grad_x_acc param in other functions as well
kainzhong
commented
May 1, 2026
|
|
||
| ctx.dtype = H.dtype | ||
| H = H.to(torch.float32) | ||
| ctx.alpha_main_grad = getattr(alpha, "main_grad", None) |
Collaborator
Author
There was a problem hiding this comment.
With FSDP this will break
Check linear.py for reference
Talk with Varun
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
Some enhancement for mHC to better align with DeepSeek's tilelang implementation: https://github.com/deepseek-ai/TileKernels/tree/main/tile_kernels/mhc
Fixes # (issue)
Type of change
Changes
mhc_fused_projectionto accept arguments with mixed dtype: x.dtype=bf16, phi.dtype=fp32, which matches DeepSeek's implementationmhc_fused_projectionnow outputs fp32 regardless of the input dtype, matching DeepSeek's implementationfuse_grad_x_accoptimization, which will reuse the same grad_x buffer to accumulate the initial mHC input x's gradient formhc_fused_expand_combine,mhc_fused_aggregateandmhc_fused_projectionnorm_weightformhc_fused_projection, which would be equivalent to apply RMSNorm in the unfused manner withelementwise_affine=True, which would be the learnable per-element affine parameters for RMSNormmain_gradoptimization for Megatron-LM integration, which will accumulate the gradient of phi, alpha, beta (they are all supposed to betorch.nn.Parameter) tomain_gradif such attribute exists.learned_output_contractin [dev] [DeepSeek-v4] Part 3: MTP support with mHC and new mHC contract Megatron-LM#4518Checklist: