yiqichen01 commited on
Commit
d88c9df
·
verified ·
1 Parent(s): 6ed1e30
Files changed (2) hide show
  1. README.md +1 -1
  2. README_CN.md +1 -1
README.md CHANGED
@@ -318,7 +318,7 @@ DFly2 is Hy4 flash's speculative-decoding drafter, trained with [AngelSpec](http
318
  - **PD-disaggregation ready**: not involved in prefill; the prefill-to-decode transfer adds only one last-layer hidden state.
319
 
320
  <p align="center">
321
- <a href="assets/dfly2.png"><img src="assets/dfly2.png" alt="DFly2 architecture: each draft layer reuses the KV cache and the indexer-selected positions of a sparse attention layer of Hy4 flash" width="40%"/></a>
322
  </p>
323
 
324
  ### Quantization
 
318
  - **PD-disaggregation ready**: not involved in prefill; the prefill-to-decode transfer adds only one last-layer hidden state.
319
 
320
  <p align="center">
321
+ <a href="assets/dfly2.png"><img src="assets/dfly2.png" alt="DFly2 architecture: each draft layer reuses the KV cache and the indexer-selected positions of a sparse attention layer of Hy4 flash" width="75%"/></a>
322
  </p>
323
 
324
  ### Quantization
README_CN.md CHANGED
@@ -304,7 +304,7 @@ DFly2 是基于 [AngelSpec](https://github.com/Tencent/AngelSpec) 框架训练
304
  - **适配 PD 分离**:不参与 prefill,P→D 仅额外传输一个末层隐状态。
305
 
306
  <p align="center">
307
- <a href="assets/dfly2.png"><img src="assets/dfly2.png" alt="DFly2 结构图:每个 draft 层复用 Hy4 flash 对应稀疏注意力层的 KV cache 与压缩索引器选出的位置" width="40%"/></a>
308
  </p>
309
 
310
  ### 量化工具
 
304
  - **适配 PD 分离**:不参与 prefill,P→D 仅额外传输一个末层隐状态。
305
 
306
  <p align="center">
307
+ <a href="assets/dfly2.png"><img src="assets/dfly2.png" alt="DFly2 结构图:每个 draft 层复用 Hy4 flash 对应稀疏注意力层的 KV cache 与压缩索引器选出的位置" width="75%"/></a>
308
  </p>
309
 
310
  ### 量化工具