contrib: Mixtral MoE (SDK 2.29) + Mistral-Small-4-119B-2603#133
Open
jimburtoft wants to merge 1 commit intoaws-neuron:mainfrom
Open
contrib: Mixtral MoE (SDK 2.29) + Mistral-Small-4-119B-2603#133jimburtoft wants to merge 1 commit intoaws-neuron:mainfrom
jimburtoft wants to merge 1 commit intoaws-neuron:mainfrom
Conversation
- Mixtral 8x7B: Updated README with SDK 2.29 results (40.4 tok/s, +5% over 2.28), added patch_moe.py for torch_block_wise workaround, documented TKG non-applicability - Mixtral 8x22B: New contrib directory with SDK 2.29 results (25.8 tok/s, +4%), patch_moe.py, NVMe storage instructions for 262GB model - Mistral-Small-4-119B-2603: New contrib with custom NeuronDeepseekV3ForCausalLM model class (MLA+MoE), FP8->BF16 extraction script, MLA bug fix, tokenizer fix, 74.5 tok/s on TP=16 (6.9x improvement over broken Phase 1 baseline)
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
patch_moe.pyscript and documented TKG non-applicability for MoE models.NeuronDeepseekV3ForCausalLM(429-line model class supporting MLA + 128-expert MoE). Achieves 74.5 tok/s on trn2.48xlarge TP=16 after fixing a critical MLA attention bug in stock NxDI code. Includes FP8→BF16 extraction, tokenizer fix, and all required patches.Key Findings
out_absorb = wkv_b[:, self.v_head_dim:, :]should bewkv_b[:, self.qk_nope_head_dim:, :]inmodeling_deepseek.py. Invisible for stock DeepSeek V3 (both=128) but crashes Mistral-Small-4 (v_head_dim=128, qk_nope_head_dim=64).Instance Requirements