[QUARK-479] Fix per-layer quant dtype in DeepSeek attention init#268
Open
thpereir wants to merge 2 commits intothpereir/quark_quant_layerfrom
Open
[QUARK-479] Fix per-layer quant dtype in DeepSeek attention init#268thpereir wants to merge 2 commits intothpereir/quark_quant_layerfrom
thpereir wants to merge 2 commits intothpereir/quark_quant_layerfrom
Conversation
1 task
For mixed-precision models (e.g. MXFP4 MoE + FP8 attention), the attention block must resolve its own per-layer quant spec rather than using the global quant_config['quant_dtype']. - Add _attn_spec / _attn_quant_dtype via quant_config.resolve(prefix) - Use resolved dtype for FP4/FP8 decision in attention init - Pass prefix to MergedReplicatedLinear for fused_qkv_a_proj - Use resolved dtype for fuse_qknorm_quant decision Tested with DeepSeek-R1-0528-moe-mxfp4-other-ptpc on TP=4.
Instead of checking the global quant_dtype to decide whether to bypass FP4 quantization for MTP layers, use quant_config.resolve(prefix) to check the per-layer spec. This correctly preserves FP8 quantization for MTP layer 61 when the global config is MXFP4 but the layer has an FP8 per_Token override (as in the PTPC model format).
thpereir
force-pushed
the
thpereir/quark_quant_layer
branch
from
028d617 to
7392aec
Compare
thpereir
force-pushed
the
thpereir/deepseek_r1_mxfp4_ptpc
branch
from
f51003c to
6045cef
Compare
1 task
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
For mixed-precision models (e.g. MXFP4 MoE + FP8 attention), the attention block must resolve its own per-layer quant spec rather than using the global quant_config['quant_dtype'].
Tested with DeepSeek-R1-0528-moe-mxfp4-other-ptpc on TP=4.
Depends on #236
Motivation
Technical Details
Test Plan
Test Result
Submission Checklist