Skip to content

fix(recipes): exclude Llama-4 vision branch from default PTQ quantization #1826

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Edwardf0t1 merged 1 commit into from
Jun 25, 2026
Merged

fix(recipes): exclude Llama-4 vision branch from default PTQ quantization #1826

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 11 additions & 4 deletions modelopt_recipes/configs/ptq/units/default_disabled_quantizers.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,17 +43,24 @@
# Multimodal vision branch: keep the vision encoder (SigLIP / ViT) and any
# multimodal embedding projection in BF16 by default. Recipes that enable bare
# `*weight_quantizer` / `*input_quantizer` or `*mlp*` wildcards otherwise also
# match the vision tower (`model.vision_tower.*`, `model.visual.*`) and the
# embedding projection (`model.embed_vision.*`); quantizing the vision branch
# crashes export / produces garbage image embeddings on VL models (gemma-4,
# Qwen3.5-VL — NVBugs 6293731, 6293762, 6294017). A recipe that intentionally
# match the vision tower (`model.vision_tower.*`, `model.visual.*`,
# Llama-4 `vision_model.*`) and the embedding projection (`model.embed_vision.*`,
# Llama-4 `multi_modal_projector.*`); quantizing the vision branch crashes
# export / produces garbage image embeddings on VL models (gemma-4,
# Qwen3.5-VL — NVBugs 6293731, 6293762, 6294017; Llama-4-Scout — NVBugs
# 6359097, where `vision_model.patch_embedding.linear` has in_features=588,
# not divisible by the NVFP4 block size). A recipe that intentionally
# quantizes vision must re-enable these after importing this unit.
- quantizer_name: '*embed_vision*'
enable: false
- quantizer_name: '*vision_tower*'
enable: false
- quantizer_name: '*visual*'
enable: false
- quantizer_name: '*vision_model*'
enable: false
- quantizer_name: '*multi_modal_projector*'
enable: false

@juhi10071998 juhi10071998 Jun 25, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@shengliangxu - I was curious when do we add model specific architecture exclusions in the model specific YAMLs? I don't see us following a strict pattern, wondering if you have any recommendation/ thoughts on this.
For instance, for diffusionGemma I extended the default_disabled_quantizers to have it's own model specific exclusions

@shengliangxu shengliangxu Jun 25, 2026 •
edited
Loading

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, we should really not adding more and more model_type specific exclusions here. It will pollute all models. But not gonna block this PR for the sake that we are already doing it.

I think we need a design for a more systematic way for integrating model_type specific recipe components.

If you are interested, feel free to think about it. Also need to track it in Jira.

- parent_class: 'nn.BatchNorm1d'
quantizer_name: '*'
enable: false
Expand Down
Loading