PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Update auto-round-lib README.md (#1437)
chen, suyue232b1da93006
diff --git a/auto_round_extension/ark/README.md b/auto_round_extension/ark/README.mdindex b9aa6e755..6897a5f0d 100644--- a/auto_round_extension/ark/README.md+++ b/auto_round_extension/ark/README.md@@ -12,25 +12,25 @@ The kernels are optimized for the following GPUs:## Key FeaturesAutoRound Kernel provides weight-only linear computational capabilities for LLM inference. Specifically, the weight-only-quantization configs we support are given in the table below:### CPU-| Weight dtype | Compute dtype | Scale dtype | Algorithm<sup>[1]</sup> |-| ---------------------- | :----------------: | :---------------: | :--------: |-| INT8 | INT8<sup>[2]</sup> / BF16 / FP32 | BF16 / FP32 | sym / asym |-| INT4 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |-| INT3 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |-| INT2 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |-| INT5 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |-| INT6 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |-| INT7 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |-| INT1 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |-| FP8 (E4M3, E5M2) | BF16 / FP32 | FP32 / FP8 (E8M0) | NA |-| FP4 (E2M1) | BF16 / FP32 | BF16 / FP32 | NA |+| Weight dtype | Compute dtype | Scale dtype | Algorithm<sup>[1]</sup> |+|------------------|:--------------------------------:|:-----------------:|:-----------------------:|+| INT8 | INT8<sup>[2]</sup> / BF16 / FP32 | BF16 / FP32 | sym / asym |+| INT4 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |+| INT3 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |+| INT2 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |+| INT5 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |+| INT6 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |+| INT7 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |+| INT1 | INT8 / BF16 / FP32 | BF16 / FP32 | sym / asym |+| FP8 (E4M3, E5M2) | BF16 / FP32 | FP32 / FP8 (E8M0) | NA |+| FP4 (E2M1) | BF16 / FP32 | BF16 / FP32 | NA |### XPU-| Weight dtype | Compute dtype | Scale dtype | Algorithm |-| ---------------------- | :----------------: | :---------------: | :--------: |-| INT8 | INT8 / FP16 | FP16 | sym |-| INT4 | INT8 / FP16 | FP16 | sym |-| FP8 (E4M3, E5M2) | FP16 | FP16 / FP8 (E8M0) | NA |+| Weight dtype | Compute dtype | Scale dtype | Algorithm |+|------------------|:--------------:|:-----------------:|:----------:|+| INT8 | INT8 / FP16 | FP16 | sym |+| INT4 | INT8 / FP16 | FP16 | sym |+| FP8 (E4M3, E5M2) | FP16 | FP16 / FP8 (E8M0) | NA |<sup>[1]</sup>: Quantization algorithms for integer types: symmetric or asymmetric.<sup>[2]</sup>: Includes dynamic activation quantization; results are dequantized to floating-point formats.@@ -69,17 +69,18 @@ auto-round-lib-installThe version number of auto-round-lib follows the format:`{auto-round major version}.{auto-round minor version}.{oneAPI version}.{lib version}`-**For example: v0.9.1.1**-- The first two digits (0.9) correspond to the major and minor version of the auto_round framework.-- The third digit (1) represents the major version of Intel oneAPI: `1` indicates support for oneAPI 2025.1 (typically Torch 2.8), `2` indicates support for oneAPI 2025.2 (typically Torch 2.9).+**For example: v0.10.1.1**+- The first two digits (0.10) correspond to the major and minor version of the auto_round framework.+- The third component (1) represents the major version of Intel oneAPI: `1` indicates support for oneAPI 2025.1 (typically Torch 2.8), `2` indicates support for oneAPI 2025.2 (typically Torch 2.9), and `3` indicates support for oneAPI 2025.3 (typically Torch 2.10).- The final digit (1) is the patch version of auto-round-lib, reflecting updates, bug fixes, or improvements to the kernel package itself.**Version mapping table**| auto-round-lib Version | auto-round Version | oneAPI Version | Typical PyTorch Version |-|:----------------------:|:------------------:|:--------------:|:-------------------------:|-| 0.9.1.x | 0.9.x | 2025.1 | 2.8.x |-| 0.9.2.x | 0.9.x | 2025.2 | 2.9.x |+|:----------------------:|:------------------:|:--------------:|:-----------------------:|+| 0.10.1.x | 0.10.x | 2025.1 | 2.8.x |+| 0.10.2.x | 0.10.x | 2025.2 | 2.9.x |+| 0.10.3.x | 0.10.x | 2025.3 | 2.10.x |**Notes:** oneAPI version is aligned with PyTorch version during auto-round-lib binary build, but oneAPI toolkit is not required in runtime.