fix: stream save_model to prevent OOM on large MoE models#18
Open
0xClandestine wants to merge 1 commit intoBlaizzy:pc/add-deepseekv4flash-modelfrom
Open
fix: stream save_model to prevent OOM on large MoE models#180xClandestine wants to merge 1 commit intoBlaizzy:pc/add-deepseekv4flash-modelfrom
0xClandestine wants to merge 1 commit intoBlaizzy:pc/add-deepseekv4flash-modelfrom
Conversation
When converting DeepSeek V4 Flash (256 experts × 43 layers) with -q, the process gets OOM-killed during save. The lazy computation graph from dequant → stack → quantize creates enormous BF16 intermediates that all materialize at once when saving. Build and save shards incrementally: pop weights from the dict as each shard is constructed, explicitly mx.eval before writing, then free. This bounds peak memory to ~one shard + one evaluation intermediate instead of the entire model's lazy graph.
Author
|
PR: ml-explore#1192 |
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
256 experts × 43 layers) with-q, the process gets OOM-killed duringsave_model. The lazy computation graph fromdequant → stack → quantizecreates enormous BF16 intermediates that all materialize at once.save_modelto build and save shards incrementally: pop weights as each shard is constructed, explicitlymx.evalbefore writing, then free. Bounds peak memory to ~one shard (~5 GB) + one evaluation intermediate (~4 GB) instead of the entire model's lazy graph.Test plan
test_utils.pytests passmlx_lm convert --hf-path deepseek-ai/DeepSeek-V4-Flash -qon a machine with sufficient disk space