PP-DocLayoutV3 checkpoint/load-path mismatch in GLM-OCR

Summary

GLM-OCR loads PP-DocLayoutV3 through PPDocLayoutV3ForObjectDetection.from_pretrained(...).
With transformers==5.4.0, that load path reports the decoder detection heads as missing,
even though the published PaddlePaddle/PP-DocLayoutV3_safetensors checkpoint contains the
corresponding trained head weights under enc_* names.

This leads to newly initialized decoder heads instead of loading the tied trained weights.

Evidence

Runtime startup reported missing decoder head keys such as:

model.decoder.class_embed.weight
model.decoder.class_embed.bias
model.decoder.bbox_embed.layers.0.weight
model.decoder.bbox_embed.layers.0.bias
...

Checkpoint inspection showed these relevant keys:

model.denoising_class_embed.weight
model.enc_score_head.weight
model.enc_score_head.bias
model.enc_bbox_head.layers.0.weight
model.enc_bbox_head.layers.0.bias
model.enc_bbox_head.layers.1.weight
model.enc_bbox_head.layers.1.bias
model.enc_bbox_head.layers.2.weight
model.enc_bbox_head.layers.2.bias

But the model class expects:

model.denoising_class_embed.weight
model.decoder.class_embed.weight
model.decoder.class_embed.bias
model.decoder.bbox_embed.layers.0.weight
model.decoder.bbox_embed.layers.0.bias
model.decoder.bbox_embed.layers.1.weight
model.decoder.bbox_embed.layers.1.bias
model.decoder.bbox_embed.layers.2.weight
model.decoder.bbox_embed.layers.2.bias

The installed transformers implementation also declares these tied mappings:

_tied_weights_keys = {
    "decoder.class_embed": "enc_score_head",
    "decoder.bbox_embed": "enc_bbox_head",
}

Root cause

The checkpoint stores the trained detection heads under encoder-head names:

• model.enc_score_head.*
• model.enc_bbox_head.layers.*

But the object-detection wrapper expects decoder-head names:

• model.decoder.class_embed.*
• model.decoder.bbox_embed.layers.*

GLM-OCR's original load path did not alias the checkpoint keys before constructing
PPDocLayoutV3ForObjectDetection, so the decoder heads were treated as missing and
initialized from scratch.

Symptoms

• startup warnings about missing decoder head weights
• degraded or unstable layout detection in self-hosted OCR runs
• deprecation warning for PPDocLayoutV3ImageProcessorFast

Minimal fix direction

1. load the PP-DocLayoutV3 config separately
2. load model.safetensors directly
3. alias:
   • model.enc_score_head.* -> model.decoder.class_embed.*
   • model.enc_bbox_head.layers.* -> model.decoder.bbox_embed.layers.*
4. construct the model with the prepared state dict

Separately, switch from PPDocLayoutV3ImageProcessorFast to PPDocLayoutV3ImageProcessor
to remove the transformers 5.4.0 deprecation warning.