Issue Station
Sign In
Tools hereToolbox

vllm - ✅(Solved) Fix [Bug][Model] Eagle2.5-VL applies ImageNet normalization instead of SigLIP2 [1 pull requests, 1 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
vllm-project/vllm#37977•Fetched 2026-04-08 01:22:16
View on GitHub
Comments
0
Participants
1
Timeline
1
Reactions
0
Author
edwingao28edwingao28
Participants
edwingao28edwingao28
Timeline (top)
cross-referenced ×1

Fix Action

Fixed

PR fix notes

PR #37978: [Bugfix][Model] Fix Eagle2.5-VL using ImageNet normalization instead of SigLIP2

Description (problem / solution / changelog)

Summary

Eagle2.5-VL uses SigLIP2 as its vision encoder, which expects mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5] normalization. However, it inherits InternVLImageProcessor which hardcodes ImageNet normalization (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]).

This silently degrades all Eagle2.5-VL vision encoder outputs.

Fix: Add Eagle2_5_VLImageProcessor subclass (following the existing Nemotron-VL pattern in nemotron_vl.py) that overrides _images_to_pixel_values_lst() with SigLIP2 normalization constants from the model's preprocessor_config.json.

Fixes #37977

Evidence

  • nvidia/Eagle2-1B preprocessor_config.json: image_mean=[0.5, 0.5, 0.5], image_std=[0.5, 0.5, 0.5]
  • Eagle2.5 imports SiglipVisionModel (not InternViT)
  • Nemotron-VL (nemotron_vl.py:250-292) solves identical problem with same pattern

This PR is not duplicating existing work

  • Searched issues/PRs for "Eagle2.5 normalization", "Eagle2 image_mean", "Eagle2_5_vl normalization" — no matches found.

Test plan

  • Verify pixel value range is [-1, 1] after normalization (SigLIP2) instead of non-uniform ImageNet range
  • Run Eagle2.5-VL inference and compare outputs with HuggingFace reference

Changed files

  • vllm/model_executor/models/eagle2_5_vl.py (modified, +73/-1)

Code Example

Eagle2_5_VLProcessingInfo.get_image_processor()     [eagle2_5_vl.py:87]
  └─► InternVLImageProcessor(**kwargs)
        └─► _images_to_pixel_values_lst()            [internvl.py:257]
              └─► image_to_pixel_values_internvl()    [internvl.py:169]
                    └─► build_transform(input_size)   [internvl.py:27]
                          └─► T.Normalize(IMAGENET_MEAN, IMAGENET_STD)  ← wrong for Eagle2.5
RAW_BUFFERClick to expand / collapse

Your current environment

vLLM main branch (commit 5bf3c42d4)

Model

nvidia/Eagle2.5-8B, nvidia/Eagle2-1B (all Eagle2.5-VL variants)

🐛 Describe the bug

Eagle2.5-VL in vLLM applies ImageNet normalization (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) instead of SigLIP2 normalization (mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]).

Eagle2.5-VL uses SigLIP2 as its vision encoder, but inherits InternVL's InternVLImageProcessor which hardcodes ImageNet normalization. The model's official HuggingFace preprocessor_config.json specifies image_mean=[0.5, 0.5, 0.5] and image_std=[0.5, 0.5, 0.5].

Call chain:

Eagle2_5_VLProcessingInfo.get_image_processor()     [eagle2_5_vl.py:87]
  └─► InternVLImageProcessor(**kwargs)
        └─► _images_to_pixel_values_lst()            [internvl.py:257]
              └─► image_to_pixel_values_internvl()    [internvl.py:169]
                    └─► build_transform(input_size)   [internvl.py:27]
                          └─► T.Normalize(IMAGENET_MEAN, IMAGENET_STD)  ← wrong for Eagle2.5

Evidence:

  • nvidia/Eagle2-1B preprocessor_config.json: image_mean=[0.5, 0.5, 0.5], image_std=[0.5, 0.5, 0.5]
  • Eagle2.5 uses SigLIP2 vision encoder (eagle2_5_vl.py imports SiglipVisionModel)
  • InternVL uses InternViT (correct with ImageNet norm), Eagle2.5 uses SigLIP2 (needs 0.5/0.5 norm)

Impact: Wrong normalization silently degrades all Eagle2.5-VL vision encoder outputs in vLLM.

Precedent: Nemotron-VL (nemotron_vl.py:250-292) solves the same problem (SigLIP + InternVL tiling) by subclassing InternVLImageProcessor with SigLIP normalization.

Before submitting a new issue...

  • Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

extent analysis

Fix Plan

To fix the issue, we need to update the Eagle2_5_VLProcessingInfo class to use the correct normalization for the SigLIP2 vision encoder. We can achieve this by subclassing InternVLImageProcessor and overriding the build_transform method to use the correct normalization.

Code Changes

from transformers import InternVLImageProcessor
from PIL import Image

class Eagle2_5_VLImageProcessor(InternVLImageProcessor):
    def build_transform(self, input_size):
        # Override the build_transform method to use SigLIP2 normalization
        return Compose([
            Resize(input_size),
            CenterCrop(input_size),
            ToTensor(),
            Normalize(mean=[

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…
#cache error#pipeline error#runtime error#dependency conflict#environment setup

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING

vllm - ✅(Solved) Fix [Bug][Model] Eagle2.5-VL applies ImageNet normalization instead of SigLIP2 [1 pull requests, 1 participants]