fix: restore PP process group on batched pipeline P2P ops - #83
Open
Laz4rz wants to merge 33 commits into
Open
Conversation
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…rk#4) Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com> Co-authored-by: Yueming Yuan <yym022502@gmail.com>
…adixark#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 (radixark#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 (radixark#18) Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: fzyzcjy <5236035+fzyzcjy@users.noreply.github.com>
…adixark#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>
…xark#58) Co-authored-by: Zhiyao Jiang <jessicajiang324@gmail.com>
…xark#60) Co-authored-by: zyzshishui <82826991+zyzshishui@users.noreply.github.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…rk#63) Co-authored-by: Zhichenzzz <zczeng@uw.edu>
Co-authored-by: yueming-yuan <yym022502@gmail.com>
…e host memory instead of pinned memory to avoid SegFault on ROCm (radixark#74)
The group argument was dropped from all four torch.distributed.P2POp calls in _batched_p2p_ops in b4d401c under the rationale "Remove group param from P2POp calls (PyTorch new API)". That rationale is mistaken: P2POp(op, tensor, peer, group) remains valid in current PyTorch (verified on torch 2.11) — peer stays a global rank and group selects the communicator. With group omitted, every op defaults to the WORLD communicator, so pipeline-parallel payload traffic (activations/grads between PP stages) loses its process-group isolation and is routed over the world group. This is also inconsistent within the same file: _communicate_shapes and _p2p_ops never dropped the group argument. This fork is consumed by CUDA/NCCL images only (the NPU flow patches NVIDIA upstream, not this fork), so restoring group carries no HCCL compatibility risk.
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
9 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.
Summary
_batched_p2p_opsinmegatron/core/pipeline_parallel/p2p_communication.pybuilds its fourtorch.distributed.P2POpcalls without thegroupargument. It was dropped in b4d401c (#2, "misc compatibility fixes for PyTorch and TE") with the rationale "Remove group param from P2POp calls (PyTorch new API)".That rationale is mistaken:
P2POp(op, tensor, peer, group)is still valid in current PyTorch (verified on torch 2.11).peerstays a global rank;groupselects the communicator. The API did not change.groupomitted, every op defaults to the WORLD communicator — pipeline-parallel payload traffic (activations/grads between PP stages) loses its process-group isolation._communicate_shapespassesself.pp_groupand_p2p_opspassesgroupto this day, so the batched path is internally inconsistent.This PR restores exactly the pre-b4d401c71 form (matches NVIDIA upstream
main):(same for the other three ops; 4 lines changed, nothing else from b4d401c is touched)
NPU safety
This fork is consumed by CUDA/NCCL images only — the NPU flow patches NVIDIA upstream, not this fork — so restoring
groupcarries no HCCL compatibility risk.Downstream note
Downstream (arcee) currently carries a runtime compat shim that monkey-patches
_batched_p2p_opsto re-add the group; it self-disables once it detects this fix has landed, so no coordinated rollout is needed.