From f37d00e856266da3dff40b5adc76fad0571b76df Mon Sep 17 00:00:00 2001 From: littledgg <61149469+littledgg@users.noreply.github.com> Date: Thu, 24 Jul 2025 20:15:00 +0800 Subject: [PATCH] [Model] Provide clearer error for missing KV cache quantization scales (#3007) --- fastdeploy/model_executor/models/deepseek_v3.py | 1 + fastdeploy/model_executor/models/qwen2.py | 1 + fastdeploy/model_executor/models/qwen3.py | 1 + 3 files changed, 3 insertions(+) diff --git a/fastdeploy/model_executor/models/deepseek_v3.py b/fastdeploy/model_executor/models/deepseek_v3.py index 2c2b0efe1..4d75b03b9 100644 --- a/fastdeploy/model_executor/models/deepseek_v3.py +++ b/fastdeploy/model_executor/models/deepseek_v3.py @@ -420,6 +420,7 @@ class DeepseekV3MLAAttention(nn.Layer): # NOTE(Ryan):Make sure kv_b_proj_bmm loaded before kv_b_proj, # The same weight key will be poped after kv_b_proj. self.o_proj.load_state_dict(state_dict) + self.mla_attn.load_state_dict(state_dict) class DeepSeekV3DecoderLayer(nn.Layer): diff --git a/fastdeploy/model_executor/models/qwen2.py b/fastdeploy/model_executor/models/qwen2.py index 418f3cdfb..af2af00b1 100644 --- a/fastdeploy/model_executor/models/qwen2.py +++ b/fastdeploy/model_executor/models/qwen2.py @@ -113,6 +113,7 @@ class Qwen2Attention(nn.Layer): """ """ self.qkv_proj.load_state_dict(state_dict) self.o_proj.load_state_dict(state_dict) + self.attn.load_state_dict(state_dict) def forward( self, diff --git a/fastdeploy/model_executor/models/qwen3.py b/fastdeploy/model_executor/models/qwen3.py index 9dee0f8e4..4b106aea2 100644 --- a/fastdeploy/model_executor/models/qwen3.py +++ b/fastdeploy/model_executor/models/qwen3.py @@ -95,6 +95,7 @@ class Qwen3Attention(nn.Layer): self.o_proj.load_state_dict(state_dict) self.q_norm.load_state_dict(state_dict) self.k_norm.load_state_dict(state_dict) + self.attn.load_state_dict(state_dict) def forward( self,