Add support for fused Q Up-Proj GEMM/RoPE/Quant - #6213
Open
chaseblock wants to merge 15 commits into
Open
Conversation
svcnvidia-nemo-ci
marked this pull request as draft
August 3, 2026 15:47
Contributor
|
This PR has been automatically converted to draft because all PRs must start as drafts. When you are ready for review, click Ready for Review to begin the review process. This will:
See the contribution guide for more details. |
chaseblock
marked this pull request as ready for review
August 3, 2026 16:47
yaoyu-33
approved these changes
Aug 4, 2026
This commit adds support for fusing the q up proj gemm with the following rope and mxfp8 quantization operations. Relies on the TE functionality from NVIDIA/TransformerEngine#3303 Signed-off-by: Chase Block <cblock@nvidia.com>
Signed-off-by: Chase Block <cblock@nvidia.com>
chaseblock
force-pushed
the
qrope_fusion
branch
from
August 5, 2026 15:36
9c543f8 to
d76dee1
Compare
sraman-rgb
enabled auto-merge
August 5, 2026 17:31
Contributor
|
/ok to test d76dee1 |
santhnm2
approved these changes
Aug 5, 2026
Signed-off-by: Chase Block <cblock@nvidia.com>
auto-merge was automatically disabled
August 5, 2026 19:01
Head branch was pushed to by a user without write access
Contributor
|
/ok to test c262d0a |
Signed-off-by: Chase Block <cblock@nvidia.com>
Signed-off-by: Chase Block <cblock@nvidia.com>
jiemingz
reviewed
Aug 6, 2026
| # --- RoPE backward (unchanged: bf16, same rotary_bwd_q_kernel as the unfused path) --- | ||
| dq3 = dq.reshape(tokens, nh, q_head_dim).contiguous() | ||
| grid = lambda META: (tokens, triton.cdiv(nh, META["BLOCK_H"])) | ||
| rotary_bwd_q_kernel[grid]( |
Contributor
There was a problem hiding this comment.
are there plans to add a backward fusion? If not is there a way to avoid duplicating the backward pass that done here?
Author
There was a problem hiding this comment.
We haven't been considering a backwards fusion as far as I know, but I don't know that there's a better way to deduplicate this code at the moment.
I think this would require heavier changes to TE fully deduplicate this, and I'm not sure that this is worth it at this point.
Enabling this in the future will require either further modifications to the attention interface or reworking how the P2P communication works for CP. Signed-off-by: Chase Block <cblock@nvidia.com>
Signed-off-by: Chase Block <cblock@nvidia.com>
sraman-rgb
enabled auto-merge
August 7, 2026 14:56
Contributor
|
/ok to test 7822da3 |
Requires TE PR: NVIDIA/TransformerEngine#3330 Signed-off-by: Chase Block <cblock@nvidia.com>
auto-merge was automatically disabled
August 7, 2026 21:15
Head branch was pushed to by a user without write access
Signed-off-by: Chase Block <cblock@nvidia.com>
Contributor
|
/ok to test 9527108 |
Signed-off-by: Chase Block <cblock@nvidia.com>
Remove the duplicated Megatron autograd and RoPE-backward implementation because Transformer Engine now owns that operation. Keep Megatron responsible for routing and integration, enforce the MXFP8 DPA and zero-dropout contract found during B200 validation, and test the full Megatron-to-TE backward path. Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>
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
7 participants
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.
This commit adds support for fusing the q up proj gemm with the following rope and mxfp8 quantization operations. Relies on the TE functionality from here.
What does this PR do?
Allows us to fuse the DeepseekV3 Q up proj gemm with the succeeding rope and mxfp8 quantization operations, improving training performance.
Contribution process
Pre-checks
Code review
Feel free to message or comment @NVIDIA/mcore-oncall to help accelerate your merge into main. The less complex your PR is, the faster it will be approved and merged!
All PRs start as draft. If you open a non-draft PR, it will be automatically converted to draft.
Step 1: Mark PR as "Ready for Review"
.github/CODEOWNERS.Final Review might get declined if these requirements are not fulfilled.
Step 2: Final Review
For PRs that change
megatron/core, once all expert reviewers have approved, theFinal Reviewlabel is applied automatically and final reviewers are assigned.For PRs outside
megatron/core, this step is skipped.Step 3: Approved
Once all required reviewers have approved, the
Approvedlabel is applied automatically.Merge
Any member of mcore-engineers will be able to merge your PR.