Files
zeekay 75125f041f Update to candle 0.9.2-alpha.2 with full hanzo naming and latest upstream changes
- Updated all crate versions to 0.9.2-alpha.2
- Synced source code with latest upstream candle changes
- Added hanzo-ug crate as wrapper for ug functionality
- Fixed all import statements to use hanzo-* naming
- Updated dependencies to latest versions:
  * cudarc: 0.16.3 -> 0.18.2
  * gemm: 0.17.0 -> 0.18.2
  * safetensors: 0.4.1 -> 0.6.0
  * ug: 0.4.0 -> 0.5.0
  * criterion: 0.5.1 -> 0.7.0
- Added float8 support for enhanced precision
- Updated Metal backend to use objc2-metal/objc2-foundation
- All core crates now compile and build successfully

Core working crates:
- hanzo-ml (main ML framework)
- hanzo-nn (neural network layers)
- hanzo-transformers (transformer models)
- hanzo-datasets (data loading utilities)
- hanzo-ug (unified GPU backend)
- hanzo-kernels (CUDA kernels)
- hanzo-metal-kernels (Metal kernels)
- hanzo-flash-attn (flash attention)
- hanzo-onnx (ONNX support)
2026-01-16 17:17:01 -08:00

79 lines
2.4 KiB
Rust

use hanzo_ml::{Device, Result, Tensor};
use hanzo_transformers::generation::LogitsProcessor;
#[test]
fn sample_with_zero_temperature() -> Result<()> {
let mut logits_process = LogitsProcessor::new(1337, None, None);
let logits = Tensor::new(&[0.1, 0.2, 0.3, 0.4], &Device::Cpu)?;
let token = logits_process.sample(&logits)?;
assert_eq!(token, 3);
Ok(())
}
#[test]
fn sample_with_temperature() -> Result<()> {
let mut logits_process = LogitsProcessor::new(42, Some(0.9), None);
let logits = Tensor::new(&[0.1, 0.2, 0.3, 0.4], &Device::Cpu)?;
let token = logits_process.sample(&logits)?;
assert_eq!(token, 0);
Ok(())
}
#[test]
fn sample_with_top_p() -> Result<()> {
let mut logits_process = LogitsProcessor::new(42, Some(1.0), Some(0.5));
let logits = Tensor::new(&[0.1, 0.2, 0.3, 0.4], &Device::Cpu)?;
let token = logits_process.sample(&logits)?;
assert_eq!(token, 2);
Ok(())
}
#[test]
fn sample_with_top_k() -> Result<()> {
let mut logits_process = LogitsProcessor::from_sampling(
42,
hanzo_transformers::generation::Sampling::TopK {
k: 1,
temperature: 1.0,
},
);
let logits = Tensor::new(&[0.1, 0.2, 0.3, 0.4], &Device::Cpu)?;
let token = logits_process.sample(&logits)?;
assert_eq!(token, 3);
let mut logits_process = LogitsProcessor::from_sampling(
42,
hanzo_transformers::generation::Sampling::TopK {
k: 2,
temperature: 1.0,
},
);
let logits = Tensor::new(&[0.1, 0.2, 0.3, 0.4], &Device::Cpu)?;
let token = logits_process.sample(&logits)?;
assert_eq!(token, 3);
let token = logits_process.sample(&logits)?;
assert_eq!(token, 2);
Ok(())
}
#[test]
fn sample_gumbel() -> Result<()> {
let mut logits_process = LogitsProcessor::from_sampling(
42,
hanzo_transformers::generation::Sampling::GumbelSoftmax { temperature: 1.0 },
);
let logits = Tensor::new(&[-1.0, 0.0, 0.2, 1.0], &Device::Cpu)?;
let sm = hanzo_nn::ops::softmax(&logits, 0)?.to_vec1::<f64>()?;
let mut counts = vec![0f64; 4];
let samples = 100000;
for _ in 0..samples {
let token = logits_process.sample(&logits)?;
counts[token as usize] += 1f64 / samples as f64;
}
for i in 0..4 {
if (counts[i] - sm[i]).abs() > 0.05 {
panic!("pr mismatch {counts:?} {sm:?}");
}
}
Ok(())
}