Exclude multimodal vision branch from quantization by default (NVBug 6293731, 6293762)

[Edwardf0t1]

What does this PR do?

Type of change: Bug fix

Fixes two sglang deployment failures on multimodal Gemma (gemma-4-31B-it) caused by general PTQ presets leaking quantization into the SigLIP vision branch via broad wildcards:

• NVBug 6293731 — general/ptq/fp8_default-kv_fp8: the w8a8_fp8_fp8 unit enables bare *weight_quantizer / *input_quantizer, which also match the vision tower (model.vision_tower.*, model.visual.*) and the vision embedding projection (model.embed_vision.*). The exported checkpoint deploys but emits garbled text in sglang.
• NVBug 6293762 — general/ptq/nvfp4_mlp_only-kv_fp8: the *mlp* enables also match the vision tower's block MLPs (model.vision_tower.encoder.layers.*.mlp), and an image request crashes the FP4 kernel at decode: ValueError: too many values to unpack (expected 2) in sglang's modelopt_quant.py apply.

Fix

Add *embed_vision* / *vision_tower* / *visual* disable rules to the shared configs/ptq/units/default_disabled_quantizers unit, alongside the existing *router* / *lm_head* entries.

Both the composed general/ptq/* recipes and the configs/ptq/presets/model/* presets import this unit, so:

• every general recipe (fp8_default, nvfp4_default, nvfp4_mlp_only, nvfp4_omlp_only, …) keeps the vision branch in BF16 by default — fixing the whole vision-overreach class, not just the two reported recipes;
• the test_general_ptq_yaml_matches_config_dicts YAML↔preset parity test stays satisfied (both sides pick up the new entries from the one shared unit).

The rules are no-ops on text-only models (nothing matches). A recipe that intentionally wants to quantize the vision branch can re-enable these after importing the unit.

Files changed:

• modelopt_recipes/configs/ptq/units/default_disabled_quantizers.yaml (+14)

Testing

Re-export of gemma-4-31B-it with the affected recipes and re-deploy in sglang (the env from the bug reports: lmsysorg/sglang:v0.5.12.post1, GB200) to confirm fp8_default no longer garbles text and nvfp4_mlp_only no longer crashes on image requests. (Results to be appended.) Unit-level: tests/unit/recipe/test_loader.py::test_general_ptq_yaml_matches_config_dicts (parity) passes for all four general presets.

Before your PR is "Ready for review"

• Is this change backward compatible?: ✅ (text-only checkpoints unaffected; new rules only match vision modules that should never have been quantized by a general recipe)
• If you copied code from any other sources or added a new PIP dependency: N/A
• Did you write any new necessary tests?: N/A (recipe data fix; covered by the existing parity test + verified by real PTQ export + sglang deploy)
• Did you update Changelog?: N/A
• Did you get Claude approval on this PR?: ❌ (pending)

Additional Information

NVBug 6293731 and 6293762. Reported on modelopt 0.45.0rc0, GB200, gemma-4-31B-it, sglang 0.5.12.post1. Tracked under OMNIML-5034. Companion to PR #1690 (same vision-overreach class on the gemma-specific w4a8_awq recipe, NVBug 6294017).

🤖 Generated with Claude Code

Summary by CodeRabbit

• Chores
  • Updated quantization configuration to preserve BF16 precision for vision encoder components in multimodal models.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 3c9f8ae1-09ab-40c2-9240-ae3f30f5b2ec

📥 Commits

Reviewing files that changed from the base of the PR and between 9e2acad and 513862e.

📒 Files selected for processing (1)

• modelopt_recipes/configs/ptq/units/default_disabled_quantizers.yaml

---

📝 Walkthrough

Walkthrough

This PR extends the PTQ default quantizer disable configuration to explicitly exclude vision and multimodal components from quantization by adding three new pattern-matching rules (*embed_vision*, *vision_tower*, *visual*) with documentation that these components remain in BF16 format unless downstream recipes re-enable them.

Changes

Quantization Recipe Configuration Update

Layer / File(s)	Summary
Vision component exclusion patterns in default quantizer config modelopt_recipes/configs/ptq/units/default_disabled_quantizers.yaml	Three new quantizer disable entries for *embed_vision*, *vision_tower*, and *visual* patterns are added to the default disabled quantizers configuration, with accompanying comments explaining that vision encoders and multimodal embedding projections remain in BF16 by default.

Possibly related PRs

• NVIDIA/Model-Optimizer#1687: Both PRs modify the same PTQ quantizer-disabling YAML configuration to add rules keeping vision and multimodal components unquantized for NVFP4.

Suggested reviewers

• shengliangxu

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~3 minutes

🚥 Pre-merge checks | ✅ 6

✅ Passed checks (6 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately summarizes the main change: excluding multimodal vision branches from quantization in PTQ recipes.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Security Anti-Patterns	✅ Passed	git diff vs origin/main shows no Python changes under modelopt/ or examples/ (only YAML plus non-scope test/tool files). No SECURITY.md anti-patterns can be introduced in-scope Python.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix-gemma4-fp8-nvfp4-vision-exclude

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 67.73%. Comparing base (dd49a46) to head (513862e).
⚠️ Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1691      +/-   ##
==========================================
+ Coverage   67.72%   67.73%   +0.01%     
==========================================
  Files         511      511              
  Lines       56168    56168              
==========================================
+ Hits        38037    38043       +6     
+ Misses      18131    18125       -6     

Flag	Coverage Δ
unit	54.34% <ø> (+0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[meenchen]

I recently added vision tower and visual for qwen3.6:

Model-Optimizer/modelopt_recipes/configs/ptq/units/default_disabled_quantizers.yaml

Line 41 in 60b1af5

- quantizer_name: '*visual*'

Could you rebase and resolve the overlap?

[Edwardf0t1]

Let me do it in #1690

[Edwardf0t1]

Resolved in #1690 (commit 0cf494b). I rebased #1690 onto main and removed the duplicate bare *visual* / *vision_tower* entries your qwen3.6 change added, keeping the single documented block that disables *vision_tower* / *visual* / *embed_vision* — so each glob appears exactly once. I also dropped the now-redundant explicit vision excludes from the new huggingface/gemma4/ptq/w4a8_awq-kv_fp8_cast.yaml recipe since they're inherited from the shared unit. Verified load_recipe still resolves all three globs as disabled with *weight_quantizer enabled (INT4).

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-06-12 17:21 UTC