fallback dtype for quantization layer filter - #1629
Open
dxqb wants to merge 1 commit into
Open
Conversation
…tion layer filter Previously, layers excluded by the quantization layer filter were silently cast to train_dtype instead of a user-chosen dtype. Add QuantizationConfig.fallback_dtype (default bfloat16), which can itself be a quantized type (fp8/int8/w8a8/nf4) - replace_linear_with_quantized_layers now quantizes filtered-out layers with the fallback dtype's linear class instead of just skipping them. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Previously, layers excluded by the quantization layer filter were silently cast to train_dtype instead of a user-chosen dtype. Add QuantizationConfig.fallback_dtype (default bfloat16), which can itself be a quantized type (fp8/int8/w8a8/nf4) -
The purpose is models like Qwen and Z-Image that don't quantize well over many layers to int8 - but their attn and mlp layers still do. Selecting that filter caused the remaining layers to be loaded in bf16 which wastes vram because they still perform well in float8.
Test plan
pre-commit run --all-filespassesAI assistance