Add DeepSeek V4 Flash support for ART Megatron#745
Merged
Conversation
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Adds production DeepSeek V4 Flash support for ART Megatron, excluding context parallelism. This requires roughly 8 H200 training GPUs with streaming weight offloading enabled and 4 for inference. This can fit ~16K tokens and the trainer should run at ~4K tok/s. This is not particularly impressive MFU due to the complex kernels, but with shared prefix packing, overall this is decent.
The initial non-LoRA model implementation is adapted from Miles' DSV4 training implementation, with ART-specific integration for Megatron Bridge, model support, tokenizer behavior, validation, and rollout serving.
art.megatron.dsv4model code, including DSV4 attention/compression/indexer components, RoPE, hyper-connection, bridge glue, and tokenizer/encoding support.Notes
cp=1in model-support topologies.Validation