Port DeepSeek Sparse Attention to MambaModel#3553
Draft
janEbert wants to merge 7 commits intoNVIDIA:mainfrom
Draft
Port DeepSeek Sparse Attention to MambaModel#3553janEbert wants to merge 7 commits intoNVIDIA:mainfrom
MambaModel#3553janEbert wants to merge 7 commits intoNVIDIA:mainfrom
Conversation
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
DSA = DeepSeek Sparse Attention
And add corresponding tests. DSA = DeepSeek Sparse Attention
- New pytest test `test_dsa_gpt_mamba_equivalence.py` builds both a
GPTModel (DSA, 4 layers) and a MambaModel (pattern S-S-S-S-, 8 layers)
in-memory, remaps weights GPT→Mamba, and asserts logprob equivalence
across TP=1/PP=1, TP=2/PP=1, and TP=1/PP=2 distributed configs.
- New checkpoint conversion utility
`tools/checkpoint/remap_gpt_dsa_to_mamba.py` applies the same
layer-key remapping (decoder.layers.{N} → {2N}/{2N+1},
decoder.final_layernorm → decoder.final_norm) to DCP checkpoints.
- New functional test cases for CI: hybrid_dsa_mamba_logitsmatch_tp1_pp1
and _tp2_pp1, each with model_config.yaml (MambaModel inference) and
placeholder golden values.
- New CI recipe
`tests/test_utils/recipes/h100/mamba-dsa-static-inference.yaml` wiring
the two functional test cases to the h100 pipeline.
Extend the DSA GPT/Mamba logprob equivalence suite to cover mixed
dense+MoE architectures, mirroring the real DeepSeek-V3 layout where the
first N layers are dense and the remaining layers use MoE.
Key changes:
- Add `pre_mlp_layernorm.*` routing in `_remap_gpt_to_mamba_state_dict`
and `_remap_key` (checkpoint tool): MoE layers expose a real TENorm
for `pre_mlp_layernorm` (not fused), which maps to MoETransformerLayer
2N+1. Dense layers use IdentityOp and produce no keys, so existing
tests are unaffected.
- Add `_make_dsa_moe_config` with `moe_layer_freq=[0,0,1,1]` (first 2
GPT layers dense, last 2 MoE) and proxy MoE params matching the
DeepSeek-V3 style (4 experts, grouped-gemm, allgather dispatcher,
shared experts).
- Add `_MOE_MAMBA_PATTERN = "S-S-SESE"` and
`TestDSAMoEGPTMambaEquivalence` with the same three parametrized tests
as the dense suite
(tp=1/2 pp=1/2): logprob match, strict weight loading, and
golden-value recording/comparison.
- Add functional test
configs (`hybrid_dsa_moe_mamba_logitsmatch_tp{1,2}_pp1`) with
placeholder golden-value JSONs and corresponding CI recipe entries in
`mamba-dsa-static-inference.yaml`.
duncanriach
reviewed
Feb 24, 2026
Contributor
duncanriach
left a comment
duncanriach
left a comment
There was a problem hiding this comment.
Quick review.
I want this to merge after 3377. This will need to be adjusted to accommodate the changes in that PR
|
|
||
| MAMBA = "M" | ||
| ATTENTION = "*" | ||
| DSA_ATTENTION = "S" |
Contributor
There was a problem hiding this comment.
Wondering if 'S' should be reserved for sliding-window attention. Wondering if this should be 'D'. Of course, these choices are arbitrary and hopefully ultimately temporary.
|
|
||
| if self.config.fp8: | ||
| assert (2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads) % 16 == 0, ( | ||
| fp8_align_size = get_fp8_align_size(self.config.fp8_recipe) |
Contributor
There was a problem hiding this comment.
What prompts this fix in this PR?
| assert torch.all(padding_logits == 0.0), "Logits for padding tokens are not all zero." | ||
|
|
||
|
|
||
| class TestMambaBlockwiseFP8: |
Contributor
There was a problem hiding this comment.
Ugh, yeah. Checkpoint compatibility is an issue
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
3 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.
What does this PR do ?
Make experimental DeepSeek Sparse Attention (DSA) available to
MambaModel.Pre-checks
Core 0.8)Code review
The following process is enforced via the CODEOWNERS file for changes into
megatron/core. For changes outside ofmegatron/core, it is up to the PR author whether or not to tag the Final Reviewer team.For MRs into `main` branch
Feel free to message or comment the @mcore-oncall to help accelerate your merge into main. The less complex your PR is, the faster it will be approved and merged!
(Step 1): Add PR label
Expert Review(Step 2): Collect the expert reviewers reviews
Expert Reviewlabel when your PR is ready for review.Final Review might get declined if these requirements are not fulfilled.
(Step 3): Final Review
Final Reviewlabel(Optional Step 4): Cherry-pick into release branch
If this PR also needs to be merged into
core_r*release branches, after this PR has been merged, selectCherry-pickto open a new PR into the release branch.For MRs into `dev` branch
The proposed review process for `dev` branch is under active discussion.MRs are mergable after one approval by either
eharper@nvidia.comorzijiey@nvidia.com.Merging your PR
Any member of core-adlr and
core-nemowill be able to merge your PR.