fix: ensure vision tokens present in text for Qwen2-VL pretraining#3
Open
vra wants to merge 1 commit into
Open
fix: ensure vision tokens present in text for Qwen2-VL pretraining#3vra wants to merge 1 commit into
vra wants to merge 1 commit into
Conversation
Qwen2-VL requires <|vision_start|><|image_pad|><|vision_end|> tokens in the input text to match image features from the vision encoder. When apply_chat_template fails to insert these tokens (either because the template doesn't handle multimodal content on certain model versions, or because the fallback _fallback_format strips image references), the processor produces pixel_values without corresponding image tokens in input_ids, causing 'Image features and image tokens do not match: tokens: 0, features 288'. The fix checks for missing vision tokens after apply_chat_template and inserts them manually after the user turn start marker.
Closed
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Problem
Fixes #1 - Pretraining crashes with
ValueError: Image features and image tokens do not match: tokens: 0, features 288.Root Cause
Qwen2-VL requires
<|vision_start|><|image_pad|><|vision_end|>tokens in the input text to match the image features produced by the vision encoder. Whenapply_chat_templatefails to insert these tokens (either because the template doesn't handle multimodal content on certain model versions, or because the fallback_fallback_formatstrips image references), the processor producespixel_values+image_grid_thwfrom the images, butinput_idscontain 0<|image_pad|>tokens, causing the mismatch inget_placeholder_mask.Fix
After
apply_chat_templateproduces the text, check whether<|vision_start|>is present. If missing and the sample has a real image, manually insert the vision tokens after<|im_start|>user\n.