[bug-fix] GLM OCR Patch Merger context_dim - #37962

Code Review

The pull request correctly updates the Glm4vVisionTransformer and GlmOcrVisionTransformer constructors to accept text_config, and modifies the GlmOcrPatchMerger to use text_config.intermediate_size for context_dim. These changes align with the stated purpose of fixing the GLM OCR Patch Merger context_dim and appear to be correct and necessary for the model's functionality.

Pull request overview

Updates the GLM-OCR vision tower wiring so the patch merger’s context_dim is derived from the text config (instead of a vision-derived product), aligning the merger MLP dimensions with the language-side configuration.

Changes:

• Pass text_config into GLM4V/GLM-OCR vision transformer construction paths.
• Change GLM-OCR patch merger context_dim to text_config.intermediate_size.
• Update transformers config imports/type references to include text config types.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

class Glm4vVisionTransformer(nn.Module):
    def __init__(
        self,
        text_config: Glm4vTextConfig,
        vision_config: Glm4vVisionConfig,
        norm_eps: float = 1e-6,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ) -> None:

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.