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
..
2026-01-16 12:30:55 -08:00
2026-01-16 12:30:55 -08:00
2026-01-16 12:30:55 -08:00

hanzo-ml-llava

LLaVA (Large Language-and-Vision Assistant) is an end-to-end trained large multimodal model. This example is from hanzo-ml-llava

The code is based on https://github.com/haotian-liu/LLaVA, Hence the llava-hf version of config may perform differently.

model zoo

Right now this has been tested on liuhaotian/llava-v1.6-vicuna-7b and llava-hf/llava-v1.6-vicuna-7b-hf. Memory usage might have room for optimization.

Tokenizer Setup

The llava-hf models contain a tokenizer.json file so can be used directly with the -hf command line flag.

For the original llava models, you can use the following code to generate the tokenizer.json file.

conda create -n llava python=3.10  
pip install transformers protobuf
conda activate llava
python -c "from transformers import AutoTokenizer;tokenizer=AutoTokenizer.from_pretrained('liuhaotian/llava-v1.6-vicuna-7b');tokenizer.save_pretrained('tokenizer')"

Then the tokenizer.json file should be in tokenizer/tokenizer.json (which is the default path).

eval

cargo run --example llava --features cuda -- --image-file "llava_logo.png" --prompt "is this a cat?" --hf # default args, use  llava-hf/llava-v1.6-vicuna-7b-hf. image-file is required^_^
cargo run --example llava --features cuda -- --model-path liuhaotian/llava-v1.6-vicuna-7b  --image-file "llava_logo.png" --prompt "is this a cat?" # use liuhaotian/llava-v1.6-vicuna-7b, tokenizer setup should be done

Major Limitations

  1. Currently only support llama-2/vicuna llm. Haven't supoort Mistral yet.
  2. There are some ops like split, nonzero and where are not supported by hanzo.
  3. Lack of quantization and LoRA support.