Extend FP32 MoE numerics to low-precision recipes and registered activations - #77
Draft
zianglih wants to merge 31 commits into
Draft
Extend FP32 MoE numerics to low-precision recipes and registered activations#77zianglih wants to merge 31 commits into
zianglih wants to merge 31 commits into
Conversation
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com> Co-authored-by: Yueming Yuan <yym022502@gmail.com>
…VIDIA#5) Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
- Detach output layer params to prevent MTP gradient flowing to output layer - Add mtp_kwargs interface for flexible MTP label/loss_mask passing - Roll mtp_labels and loss_mask for RL training compatibility Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com> Co-authored-by: Yueming Yuan <yym022502@gmail.com>
…yers (NVIDIA#10) - Add is_mtp flag to MoE layers and multi_token_prediction module - Bypass routing replay for MTP layers (MTP uses fresh routing) - Replace rdxa/dev's built-in RouterReplay with miles.utils.routing_replay: - moe_utils.py: use get_routing_replay_compute_topk() wrapper - router.py: use register_routing_replay() for initialization Co-authored-by: Yueming Yuan <yym022502@gmail.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com> Co-authored-by: Yueming Yuan <yym022502@gmail.com>
After bumping Megatron (rdxa/dev), colocated IPC weight update fails with torch.AcceleratorError: CUDA error: invalid argument during torch.multiprocessing serialization of CUDA tensors. Root cause: Megatron's new TMS hook (PR NVIDIA#3048) alters allocator behavior in training flow, causing allocations via cuMemCreate/cuMemMap which are incompatible with CUDA IPC (_share_cuda_() fails). Fix: resolve mapping.py and dynamic_context.py conflicts to isolate hook side effects so TMS/allocator state remains IPC-compatible during the weight update phase. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- merge(): truncate dp_reshardable padding on optimizer/param_state path - load_parameter_state_from_dp_reshardable: tolerate missing 'padding' key - ShardedTensor: relax flattened_range to deprecation warning Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This PR rebases from radixark Megatron fork [miles-20260218](https://github.com/radixark/Megatron-LM/tree/miles-20260218) and resolve conflicts. Upgrade Megatron from Dec 17 (3714d81) to Feb 13 (1dcf0da) PR link: radixark#13 Co-authored-by: Yueming Yuan <yym022502@gmail.com> Made-with: Cursor
…se `--disable-weight-backuper` in miles (NVIDIA#18) Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: fzyzcjy <5236035+fzyzcjy@users.noreply.github.com>
…VIDIA#20) Co-authored-by: fzyzcjy <5236035+fzyzcjy@users.noreply.github.com>
Squash merge of the dense true-on-policy Megatron branch. Co-authored-by: zju-stu-lizheng <lizheng.cs@zju.edu.cn> Co-authored-by: zyxiyy02 <282300612+zyxiyy02@users.noreply.github.com> Co-authored-by: Yi Zhang <1109276519@qq.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: fzyzcjy <5236035+fzyzcjy@users.noreply.github.com>
…IA#58) Co-authored-by: Zhiyao Jiang <jessicajiang324@gmail.com>
…IA#60) Co-authored-by: zyzshishui <82826991+zyzshishui@users.noreply.github.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
yueming-yuan
force-pushed
the
miles-main
branch
from
August 24, 2026 19:59
2e64d78 to
235952d
Compare
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
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 ?
@HumansAnd
Generalizes the
moe_activation_in_fp32andmoe_combine_in_fp32options introduced by radixark/Megatron-LM#68 across supported Transformer Engine low-precision routed-MoE paths. It also replaces the SwiGLU-specific activation implementation with a reusable registration surface, with SwiGLU and squared ReLU as built-ins.Changes
delayed,tensorwise,blockwise, andmxfp8) in E4M3 and HYBRID formats, plus the standard NVFP4 recipe.TEGroupedMLProuted experts; a registered activation/gating pair for activation-in-FP32; and unfused AlltoAll with FP32 router probabilities for combine-in-FP32.tests.functional_tests.test_cases.common.moe_fp32_low_precisionfor the low-precision runtime matrix.Scope
This is a model- and architecture-neutral Megatron-core feature: there are no Inkling-, Nemotron-, or other model-specific checks or call paths. No Miles, SGLang, or Transformer Engine source changes are required. The activation option covers routed
TEGroupedMLPexperts; shared experts remain outside this path. Custom and per-module quantization recipes remain outside the supported contract.Validation
git diff --checkradixark/miles:dev-202608041247: all passedpy_compileon all six changed Python filesdelayed,tensorwise,blockwise,mxfp8, andnvfp4) in SwiGLU activation+combine mode and combine-only modeThe runtime harness verifies the active Transformer Engine recipe class, quantized FC1 and FC2 grouped GEMMs, one local expert per EP8 rank, cross-rank routing, quantization padding, FP32 router probabilities reaching the final unfused AlltoAll unpermute when combine-in-FP32 is enabled, the registered FP32 activation autograd node, and finite nonzero hidden/router/expert gradients over two iterations.
Contribution process
flowchart LR A[Pre-checks] --> B[PR Tests] subgraph Code Review/Approval C1[Expert Review] --> C2[Final Review] end B --> C1 C2 --> D[Merge]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.