mirror of
https://github.com/hanzoai/ml.git
synced 2026-11-20 23:08:29 +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]>
61 lines
1.8 KiB
Rust
61 lines
1.8 KiB
Rust
use hanzo_ml::{DType, Device, Result, Tensor};
|
|
|
|
fn rel(a: &Tensor, b: &Tensor) -> Result<f32> {
|
|
let d = (a - &b.to_dtype(DType::F32)?)?
|
|
.abs()?
|
|
.max_all()?
|
|
.to_scalar::<f32>()?;
|
|
let s = a.abs()?.max_all()?.to_scalar::<f32>()?.max(1.0);
|
|
Ok(d / s)
|
|
}
|
|
|
|
#[test]
|
|
fn bf16_cpu_op_sweep() -> Result<()> {
|
|
let dev = Device::Cpu;
|
|
let x32 = Tensor::rand(-2f32, 2f32, (4, 8, 2560), &dev)?;
|
|
let xb = x32.to_dtype(DType::BF16)?;
|
|
|
|
let checks: Vec<(&str, f32)> = vec![
|
|
(
|
|
"silu",
|
|
rel(&hanzo_nn::ops::silu(&x32)?, &hanzo_nn::ops::silu(&xb)?)?,
|
|
),
|
|
(
|
|
"softmax",
|
|
rel(
|
|
&hanzo_nn::ops::softmax_last_dim(&x32)?,
|
|
&hanzo_nn::ops::softmax_last_dim(&xb)?,
|
|
)?,
|
|
),
|
|
("mul", rel(&(&x32 * &x32)?, &(&xb * &xb)?)?),
|
|
("add", rel(&(&x32 + &x32)?, &(&xb + &xb)?)?),
|
|
("exp", rel(&x32.exp()?, &xb.exp()?)?),
|
|
];
|
|
let w32 = Tensor::rand(0.5f32, 1.5f32, (2560,), &dev)?;
|
|
let rms32 = hanzo_nn::ops::rms_norm(&x32, &w32, 1e-6)?;
|
|
let rmsb = hanzo_nn::ops::rms_norm(&xb, &w32.to_dtype(DType::BF16)?, 1e-6)?;
|
|
let mut checks = checks;
|
|
checks.push(("rms_norm", rel(&rms32, &rmsb)?));
|
|
|
|
for (name, r) in &checks {
|
|
println!("{name}: rel {r:.4}");
|
|
assert!(*r < 3e-2, "{name} broken: rel {r}");
|
|
}
|
|
Ok(())
|
|
}
|
|
|
|
#[test]
|
|
fn bf16_cpu_matmul_works() -> Result<()> {
|
|
let dev = Device::Cpu;
|
|
let a = Tensor::rand(-1f32, 1f32, (2, 33, 64), &dev)?;
|
|
let b = Tensor::rand(-1f32, 1f32, (2, 64, 17), &dev)?;
|
|
let want = a.matmul(&b)?;
|
|
let got = a
|
|
.to_dtype(DType::BF16)?
|
|
.matmul(&b.to_dtype(DType::BF16)?)?
|
|
.to_dtype(DType::F32)?;
|
|
let d = (&want - &got)?.abs()?.max_all()?.to_scalar::<f32>()?;
|
|
assert!(d < 0.15, "bf16 matmul diff {d}");
|
|
Ok(())
|
|
}
|