()
| 307 | |
| 308 | #[test] |
| 309 | fn test_dequantize_fp8_blockwise_scaling() { |
| 310 | let f32_weight = |
| 311 | Tensor::from_vec(vec![1.0f32; 128 * 128], (128, 128), &Device::Cpu).unwrap(); |
| 312 | let weight = f32_weight.to_dtype(DType::F8E4M3).unwrap(); |
| 313 | let scale_inv = Tensor::from_vec(vec![2.0f32], (1, 1), &Device::Cpu).unwrap(); |
| 314 | let result = dequantize_fp8_blockwise(&weight, &scale_inv).unwrap(); |
| 315 | let first: f32 = result.flatten_all().unwrap().to_vec1::<f32>().unwrap()[0]; |
| 316 | assert!( |
| 317 | (first - 2.0).abs() < 0.1, |
| 318 | "scale=2.0 should double value, got {first}" |
| 319 | ); |
| 320 | } |
| 321 | |
| 322 | #[test] |
| 323 | fn test_is_fp8_quantized() { |
nothing calls this directly
no test coverage detected