apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 16:48:03 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

Simplify the Config code (#2770)

Browse Source 
* simplify the code

* fix vl

* delete config

* fix

* perfect code

* fix ci

* fix xpu

* fix xpu

* fix server

* resolve conflict

* fix mtp

* resolve conflict

* fix xpu

* fix xpu

* fix vl

* fix log

* fix qwen moe

* fix qwen moe

* fix qwen moe
This commit is contained in:
YuanRisheng
2025-07-14 19:50:05 +08:00
committed by GitHub
parent 2e81792d64
commit 4c7b8bc458
34 changed files with 551 additions and 911 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
fastdeploy/model_executor/layers/attention/mla_attention_backend.py
View File

@@ -113,18 +113,18 @@ class MLAAttentionBackend(AttentionBackend):
self.kv_num_heads: int = kv_num_heads
self.num_heads: int = num_heads
self.head_dim: int = fd_config.model_config.head_dim
self.num_layers: int = fd_config.model_config.num_layers
self.num_layers: int = fd_config.model_config.num_hidden_layers
# For Multi Head Latent Attention
self.kv_lora_rank: int = fd_config.model_config.deepseekv3.kv_lora_rank
self.qk_rope_head_dim: int = fd_config.model_config.deepseekv3.qk_rope_head_dim
self.qk_head_dim: int = fd_config.model_config.deepseekv3.qk_nope_head_dim \
+ fd_config.model_config.deepseekv3.qk_rope_head_dim
self.kv_lora_rank: int = fd_config.model_config.kv_lora_rank
self.qk_rope_head_dim: int = fd_config.model_config.qk_rope_head_dim
self.qk_head_dim: int = fd_config.model_config.qk_nope_head_dim \
+ fd_config.model_config.qk_rope_head_dim
self.attn_softmax_scale: float = self.qk_head_dim**-0.5
if fd_config.model_config.deepseekv3.rope_scaling:
mscale_all_dim = fd_config.model_config.deepseekv3.rope_scaling.get(
if fd_config.model_config.rope_scaling:
mscale_all_dim = fd_config.model_config.rope_scaling.get(
"mscale_all_dim", False) # 1.0
scaling_factor = fd_config.model_config.deepseekv3.rope_scaling[
scaling_factor = fd_config.model_config.rope_scaling[
"factor"] # 40
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
self.attn_softmax_scale = self.attn_softmax_scale * mscale * mscale