mirror of
https://github.com/hanzoai/ml.git
synced 2026-11-19 23:50:51 +00:00
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]>
112 lines
2.9 KiB
Rust
112 lines
2.9 KiB
Rust
#[cfg(feature = "mkl")]
|
|
extern crate intel_mkl_src;
|
|
|
|
#[cfg(feature = "accelerate")]
|
|
extern crate accelerate_src;
|
|
|
|
use clap::{Parser, ValueEnum};
|
|
|
|
use hanzo_ml::{DType, IndexOp, D};
|
|
use hanzo_nn::{Module, VarBuilder};
|
|
use hanzo_transformers::models::repvgg;
|
|
|
|
#[derive(Clone, Copy, Debug, ValueEnum)]
|
|
enum Which {
|
|
A0,
|
|
A1,
|
|
A2,
|
|
B0,
|
|
B1,
|
|
B2,
|
|
B3,
|
|
B1G4,
|
|
B2G4,
|
|
B3G4,
|
|
}
|
|
|
|
impl Which {
|
|
fn model_filename(&self) -> String {
|
|
let name = match self {
|
|
Self::A0 => "a0",
|
|
Self::A1 => "a1",
|
|
Self::A2 => "a2",
|
|
Self::B0 => "b0",
|
|
Self::B1 => "b1",
|
|
Self::B2 => "b2",
|
|
Self::B3 => "b3",
|
|
Self::B1G4 => "b1g4",
|
|
Self::B2G4 => "b2g4",
|
|
Self::B3G4 => "b3g4",
|
|
};
|
|
format!("timm/repvgg_{name}.rvgg_in1k")
|
|
}
|
|
|
|
fn config(&self) -> repvgg::Config {
|
|
match self {
|
|
Self::A0 => repvgg::Config::a0(),
|
|
Self::A1 => repvgg::Config::a1(),
|
|
Self::A2 => repvgg::Config::a2(),
|
|
Self::B0 => repvgg::Config::b0(),
|
|
Self::B1 => repvgg::Config::b1(),
|
|
Self::B2 => repvgg::Config::b2(),
|
|
Self::B3 => repvgg::Config::b3(),
|
|
Self::B1G4 => repvgg::Config::b1g4(),
|
|
Self::B2G4 => repvgg::Config::b2g4(),
|
|
Self::B3G4 => repvgg::Config::b3g4(),
|
|
}
|
|
}
|
|
}
|
|
|
|
#[derive(Parser)]
|
|
struct Args {
|
|
#[arg(long)]
|
|
model: Option<String>,
|
|
|
|
#[arg(long)]
|
|
image: String,
|
|
|
|
/// Run on CPU rather than on GPU.
|
|
#[arg(long)]
|
|
cpu: bool,
|
|
|
|
#[arg(value_enum, long, default_value_t=Which::A0)]
|
|
which: Which,
|
|
}
|
|
|
|
pub fn main() -> anyhow::Result<()> {
|
|
let args = Args::parse();
|
|
|
|
let device = hanzo_ml_examples::device(args.cpu)?;
|
|
|
|
let image = hanzo_ml_examples::imagenet::load_image224(args.image)?.to_device(&device)?;
|
|
println!("loaded image {image:?}");
|
|
|
|
let model_file = match args.model {
|
|
None => {
|
|
let model_name = args.which.model_filename();
|
|
let api = hanzo_ml_examples::hub::Api::new()?;
|
|
let api = api.model(model_name);
|
|
api.get("model.safetensors")?
|
|
}
|
|
Some(model) => model.into(),
|
|
};
|
|
|
|
let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
|
|
let model = repvgg::repvgg(&args.which.config(), 1000, 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(())
|
|
}
|