Skip to content

[#17723][fix] Re-path dense Qwen quantization entries - #17727

Draft
mihai-chiorean wants to merge 1 commit into
NVIDIA:mainfrom
mihai-chiorean:fix/17723-qwen38-mlp-quant-paths
Draft

[#17723][fix] Re-path dense Qwen quantization entries#17727
mihai-chiorean wants to merge 1 commit into
NVIDIA:mainfrom
mihai-chiorean:fix/17723-qwen38-mlp-quant-paths

Conversation

@mihai-chiorean

@mihai-chiorean mihai-chiorean commented Aug 14, 2026

Copy link
Copy Markdown
Contributor

Summary

Translate dense Qwen3.5/3.8 MLP entries from the Hugging Face .mlp.{gate,up,down}_proj namespace to the runtime _DenseMlpAdapter .mlp.mlp.* namespace for every quantization algorithm.

Related to #17723.

Motivation

The existing translation was gated on W4A16_NVFP4. Qwen3.8 compressed-tensors checkpoints correctly parse packed W4A4 layers as NVFP4 and later dense layers as rowwise FP8, so neither entry matched the runtime module tree. The modules then silently inherited the unquantized MIXED_PRECISION global config.

Testing

Dependencies

The translation itself is independent. Qwen3.8 obtains these per-module entries from #17725.

Signed-off-by: Mihai Chiorean <mihai.v.chiorean@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant