Modelpublic

glebkudr/PaddleOCR-VL-1.6-WebGPU

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

1likes

Model Card

PaddleOCR-VL-1.6 WebGPU runtime artifacts

Browser deployment artifacts derived from `PaddlePaddle/PaddleOCR-VL-1.6` at revision 66317acc4c9fc17bd154591ce650735cd2855f3e.

This repository is intended for a custom WebGPU inference runtime. It is not a drop-in Transformers checkpoint.

The model was not fine-tuned and its architecture was not modified.

—The pinned BF16 checkpoint was deterministically converted to IEEE FP16 using round-to-nearest, ties-to-even.
—Vision-encoder and projector matrices were transposed once into the input-major storage layout consumed by the WebGPU kernels.
—Decoder weights were packed into the custom PVLCPK01 container.
—Patch/position embeddings and the token embedding table were widened from BF16 to F32 where required by the current browser runtime.
—The upstream tokenizer is included byte-for-byte.

Path	Purpose
`m7u-vision-stack-full-fp16-input-major-official/`	27-layer FP16 vision encoder shards and manifests
`m7v-projector-full-fp16-input-major-official/`	FP16 projector weights and manifest
`m7q1-decoder-stack-logits-fp16-official.pvlc`	FP16 decoder, final norm and LM-head pack
`m6e10-vision-embed.safetensors`	F32 patch and position embeddings
`m6e9-embed-tokens.f32`	F32 token embedding table
`m6e9-embed-tokens.json`	Token embedding provenance and digest
`m6e9-tokenizer.json`	Byte-exact upstream tokenizer
`SHA256SUMS`	Transport checksums for every published artifact

The inference engine is intentionally not included in this model repository. It will be distributed separately under its own license.

—Base model: PaddlePaddle/PaddleOCR-VL-1.6
—Base revision: 66317acc4c9fc17bd154591ce650735cd2855f3e
—Converted FP16 checkpoint BLAKE3: 7eaf17cbadb7ff816491a3bfe8c96abc52c85ceea5506e73f3eb676cff57655a
—Conversion metadata: pvlc.conversion=bf16_to_ieee_f16_rne_v1

The per-component manifests contain exact tensor shapes, storage layouts, byte counts and BLAKE3 digests.

The derived model artifacts retain the parent model's Apache License 2.0. See LICENSE.

PaddleOCR-VL and PaddleOCR are works of their respective upstream authors. This repository does not imply endorsement by PaddlePaddle.