Files
zeekayandHanzo Dev 6a363f7ee3 merge: upstream main ddf1b879 and the aarch64 repack branch 35d7ae7c
Three-way per file through the rename: base 39355c6c91, theirs the upstream
merge of main with 35d7ae7c (via 7ef6a694d), ours origin/main.

Lands the barrier thread pool, runtime CPU feature dispatch, aarch64 and x86
repacked quantized kernels (indexed_gemv, gemv_fused_shared_lhs), quantized
embedding on CPU/Metal/CUDA, paged flash-attn kernels, cutile fused MoE, the
in-place op aliasing API, parking_lot storage locks, Metal buffer builders and
debug labels, safetensors 0.8, hf-hub 1.0, tokenizers 0.23.

Ours keeps ROCm, Vulkan, wgpu, the IQ/MXFP4/NVFP4/TQ dtypes, and the i-quant
weight/activation block split in matmul. Our backends ride InplaceOpN and the
quantized embedding; Metal get_rows gains the IQ kernels. ug leaves, as upstream.

Co-authored-by: Hanzo Dev <[email protected]>
2026-09-15 15:52:28 -07:00
..

PaddleOCR-VL

PaddleOCR-VL is a state-of-the-art vision-language model for document parsing, developed by PaddlePaddle. With only 0.9B parameters, it achieves competitive performance against much larger models (72B+) while maintaining fast inference speeds.

Features

  • Multilingual: Supports 109 languages including Chinese, English, Japanese, Korean, Arabic, and more
  • Multi-element Recognition: Handles text, tables, formulas, and charts
  • Dynamic Resolution: NaViT-style encoder processes images at variable resolutions without distortion
  • Multi-Image Processing: Process multiple images (e.g., multi-page documents) in a single prompt
  • Video Support: Extract and process video frames with temporal position encoding
  • Efficient: Compact 0.9B parameters with grouped query attention (GQA)
  • Position Embedding Caching: LFU cache for interpolated position embeddings improves performance

Command Line Options

Option Description Default
--image Path to document image (can be specified multiple times) (required*)
--video Path to video file (required*)
--fps Frames per second to extract from video 1.0
--max-frames Maximum frames to extract from video 16
--task Task type: ocr, table, formula, chart ocr
--model-id HuggingFace model ID PaddlePaddle/PaddleOCR-VL
--revision Model revision main
--max-length Maximum generation length 1024
--cpu Run on CPU false
--bf16 Use bfloat16 precision false
--seed Random seed 299792458

* Either --image or --video is required (mutually exclusive).

Examples

Basic Recognition

cargo run --example paddleocr-vl --release -- \
    --image hanzo-ml-examples/examples/paddleocr-vl/test_ocr.png \
    --task ocr

Table Recognition

cargo run --example paddleocr-vl --release -- \
    --image hanzo-ml-examples/examples/paddleocr-vl/test_table.png \
    --task table

Formula Recognition

cargo run --example paddleocr-vl --release -- \
    --image hanzo-ml-examples/examples/paddleocr-vl/test_formula.png \
    --task formula

Chart Recognition

cargo run --example paddleocr-vl --release -- \
    --image hanzo-ml-examples/examples/paddleocr-vl/test_chart.png \
    --task chart

Multi-Image (combined output)

Multi-Image OCR works with any task and uses --task ocr by default.

# Process multiple images with combined output
cargo run --example paddleocr-vl --release -- \
    --image hanzo-ml-examples/examples/paddleocr-vl/test_ocr.png \
    --image hanzo-ml-examples/examples/paddleocr-vl/test_ocr_page2.png

Mutli-Image (batch)

# Process chosen images sequentially with distinct output
cargo run --example paddleocr-vl --release -- \
    --batch hanzo-ml-examples/examples/paddleocr-vl/test_ocr.png hanzo-ml-examples/examples/paddleocr-vl/test_ocr_page2.png

# With shell glob expansion
cargo run --example paddleocr-vl --release -- \
    --batch hanzo-ml-examples/examples/paddleocr-vl/test_ocr*.png

Video OCR

cargo run --example paddleocr-vl --release -- \
    --video hanzo-ml-examples/examples/paddleocr-vl/test_video.mp4 \
    --task video \
    --fps 0.6 \
    --max-frames 64 \
    --max-length 2048