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

83 lines
2.6 KiB
Rust

//! EVA-02: Explore the limits of Visual representation at scAle
//! https://github.com/baaivision/EVA
#[cfg(feature = "mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]
extern crate accelerate_src;
use clap::Parser;
use hanzo_ml::{DType, Device, IndexOp, Result, Tensor, D};
use hanzo_nn::{Module, VarBuilder};
use hanzo_transformers::models::eva2;
/// Loads an image from disk using the image crate, this returns a tensor with shape
/// (3, 448, 448). OpenAI normalization is applied.
pub fn load_image448_openai_norm<P: AsRef<std::path::Path>>(p: P) -> Result<Tensor> {
let img = image::ImageReader::open(p)?
.decode()
.map_err(hanzo_ml::Error::wrap)?
.resize_to_fill(448, 448, image::imageops::FilterType::Triangle);
let img = img.to_rgb8();
let data = img.into_raw();
let data = Tensor::from_vec(data, (448, 448, 3), &Device::Cpu)?.permute((2, 0, 1))?;
let mean =
Tensor::new(&[0.48145466f32, 0.4578275, 0.40821073], &Device::Cpu)?.reshape((3, 1, 1))?;
let std = Tensor::new(&[0.26862954f32, 0.261_302_6, 0.275_777_1], &Device::Cpu)?
.reshape((3, 1, 1))?;
(data.to_dtype(hanzo_ml::DType::F32)? / 255.)?
.broadcast_sub(&mean)?
.broadcast_div(&std)
}
#[derive(Parser)]
struct Args {
#[arg(long)]
model: Option<String>,
#[arg(long)]
image: String,
/// Run on CPU rather than on GPU.
#[arg(long)]
cpu: bool,
}
pub fn main() -> anyhow::Result<()> {
let args = Args::parse();
let device = hanzo_ml_examples::device(args.cpu)?;
let image = load_image448_openai_norm(args.image)?.to_device(&device)?;
println!("loaded image {image:?}");
let model_file = match args.model {
None => {
let api = hanzo_ml_examples::hub::Api::new()?;
let api = api.model("vincent-espitalier/hanzo-ml-eva2");
api.get("eva02_base_patch14_448.mim_in22k_ft_in22k_in1k_adapted.safetensors")?
}
Some(model) => model.into(),
};
let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
let model = eva2::vit_base(vb)?;
println!("model built");
let logits = model.forward(&image.unsqueeze(0)?)?;
let prs = hanzo_nn::ops::softmax(&logits, D::Minus1)?
.i(0)?
.to_vec1::<f32>()?;
let mut prs = prs.iter().enumerate().collect::<Vec<_>>();
prs.sort_by(|(_, p1), (_, p2)| p2.total_cmp(p1));
for &(category_idx, pr) in prs.iter().take(5) {
println!(
"{:24}: {:.2}%",
hanzo_ml_examples::imagenet::CLASSES[category_idx],
100. * pr
);
}
Ok(())
}