Upgrade to vllm 0.17.0 corex v4.1 overlay

vllm/model_executor/models/qwen3.py

@@ -51,7 +51,7 @@ from vllm.v1.attention.backend import AttentionType
 from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP
 from .qwen2 import Qwen2MLP as Qwen3MLP
 from .qwen2 import Qwen2Model
-from .utils import AutoWeightsLoader, PPMissingLayer, extract_layer_index, maybe_prefix, reparse_quant_config
+from .utils import AutoWeightsLoader, PPMissingLayer, extract_layer_index, maybe_prefix
 
 logger = init_logger(__name__)
 
@@ -142,7 +142,6 @@ class Qwen3Attention(nn.Module):
         self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
         self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
         self.qk_norm = RMSNormQK(self.head_dim, self.head_dim,  eps=rms_norm_eps)
-
     def forward(
         self,
         positions: torch.Tensor,
@@ -150,17 +149,19 @@ class Qwen3Attention(nn.Module):
     ) -> torch.Tensor:
         qkv, _ = self.qkv_proj(hidden_states)
         q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
-        # # Add qk-norm
+        # Add qk-norm
         # q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim, self.head_dim)
-        # q_by_head = self.q_norm(q_by_head)
+        # q_by_head = self.q_norm.forward_native(q_by_head) # TODO(gyf) check why
         # q = q_by_head.view(q.shape)
         # k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim, self.head_dim)
-        # k_by_head = self.k_norm(k_by_head)
+        # k_by_head = self.k_norm.forward_native(k_by_head)
         # k = k_by_head.view(k.shape)
-
-        q_by_head = q.view(*q.shape[:-1], q.shape[-1]//self.head_dim, self.head_dim)
-        k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim, self.head_dim)
-
+        q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim,
+                           self.head_dim)
+        k_by_head = k.view(*k.shape[:-1],
+                           k.shape[-1] // self.head_dim,
+                           self.head_dim)
+        
         out_q, out_k = self.qk_norm(
             q_by_head,
             k_by_head,
@@ -170,7 +171,6 @@ class Qwen3Attention(nn.Module):
 
         q = out_q.view(q.shape)
         k = out_k.view(k.shape)
-
         q, k = self.rotary_emb(positions, q, k)
         attn_output = self.attn(q, k, v)
         output, _ = self.o_proj(attn_output)
@@ -201,8 +201,6 @@ class Qwen3DecoderLayer(nn.Module):
         else:
             attn_type = AttentionType.ENCODER_ONLY
 
-        quant_config = reparse_quant_config(prefix, quant_config)
-
         self.self_attn = Qwen3Attention(
             hidden_size=self.hidden_size,
             num_heads=config.num_attention_heads,
@@ -236,23 +234,25 @@ class Qwen3DecoderLayer(nn.Module):
         hidden_states: torch.Tensor,
         residual: torch.Tensor | None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
         # Self Attention
         if residual is None:
             residual = hidden_states
             hidden_states = self.input_layernorm(hidden_states)
         else:
-            hidden_states, residual = self.input_layernorm(hidden_states, residual)
+            hidden_states, residual = self.input_layernorm(
+                hidden_states, residual)
         hidden_states = self.self_attn(
             positions=positions,
             hidden_states=hidden_states,
         )
 
         # Fully Connected
-        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states, residual = self.post_attention_layernorm(
+            hidden_states, residual)
         hidden_states = self.mlp(hidden_states)
         return hidden_states, residual
 
-
 ALL_DECODER_LAYER_TYPES = {
     "attention": Qwen3DecoderLayer,
 }