Update README.md (#2)

- Update README.md (ce34483c2fb7b4f0154e581f6a3d3bdff00caf8b)


Co-authored-by: laixinn <yuanzu@users.noreply.huggingface.co>

README.md

@@ -2,6 +2,37 @@
 license: mit
 library_name: transformers
 ---
+
+# Block-wise INT8 DeepSeek-R1
+
+The INT8 data type is both friendly and efficient for most hardware platforms.
+
+**We provide a block-wise INT8 weight for DeepSeek-R1.**
+
+In benchmarking, we observe **no accuracy loss** and up to **30\%** performance enhancement.
+
+[SGLang](https://github.com/sgl-project/sglang/tree/main) will soon support the block-wise INT8 quantization operation once our [PULL REQUEST](https://github.com/sgl-project/sglang/pull/3730) is merged.
+
+## 1. Benchmarking Result (detailed in [PULL REQUEST](https://github.com/sgl-project/sglang/pull/3730)):
+| Model  | Config | Accuracy (GSM8K) | Accuracy (MMLU) | Output Throughput(qps=128) | Output Throughput(bs=1) |
+|--------|--------|-------------------|----------------|------------------------------|--------------------------|
+| BF16 R1 | (A100\*16)x2 | 95.8              | 87.1           | 4450.02 (+33%)                | 44.18 (+18%)             |
+| INT8 R1 | A100\*32  | 95.5              | 87.1           | 3342.29                       | 37.20                     |
+
+## 2. Quantization Process
+
+We apply INT8 quantization to the BF16 checkpoints.
+
+The weight scales are determined by dividing the block-wise maximum of element values by the INT8 type maximum.
+
+To generate this weight, run the provided script in the ``./inference`` directory:
+
+``
+python3 bf16_case_int8.py --input-bf16-hf-path /path/to/bf16-weights/ --output-int8-hf-path /path/to/save-int8-weight/
+``
+
+---
+
 # DeepSeek-R1
 <!-- markdownlint-disable first-line-h1 -->
 <!-- markdownlint-disable html -->
@@ -46,11 +77,6 @@ library_name: transformers
   <a href="https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf"><b>Paper Link</b>👁️</a>
 </p>
 
-## 0. INT8 Quantization
-
-We apply a INT8 quantization on the BF16 checkpoints, where weight scales are determined by dividing he block-wise maximum of element values by the INT8 type maximum.
-The quantization script is provided in inference/bf16_case_int8.py.
-
 ## 1. Introduction
 
 We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1.