diff --git a/python/sglang/srt/models/dflash.py b/python/sglang/srt/models/dflash.py index 3b0f16b998..62f823be4f 100644 --- a/python/sglang/srt/models/dflash.py +++ b/python/sglang/srt/models/dflash.py @@ -228,15 +228,7 @@ class DFlashAttention(nn.Module): attention_bias = bool(getattr(config, "attention_bias", False)) rms_norm_eps = float(getattr(config, "rms_norm_eps", 1e-6)) - self.qkv_proj = QKVParallelLinear( - hidden_size=hidden_size, - head_size=head_dim, - total_num_heads=self.total_num_heads, - total_num_kv_heads=self.total_num_kv_heads, - bias=attention_bias, - quant_config=quant_config, - prefix=f"{prefix}.qkv_proj" if prefix else "qkv_proj", - ) + self.qkv_proj = self._build_qkv_proj(attention_bias, quant_config, prefix) self.o_proj = RowParallelLinear( self.total_num_heads * head_dim, hidden_size, @@ -319,6 +311,17 @@ class DFlashAttention(nn.Module): quant_config=quant_config, ) + def _build_qkv_proj(self, bias: bool, quant_config, prefix: str) -> nn.Module: + return QKVParallelLinear( + hidden_size=self.hidden_size, + head_size=self.head_dim, + total_num_heads=self.total_num_heads, + total_num_kv_heads=self.total_num_kv_heads, + bias=bias, + quant_config=quant_config, + prefix=f"{prefix}.qkv_proj" if prefix else "qkv_proj", + ) + def forward_prepare_npu(self, positions, hidden_states): qkv, _ = self.qkv_proj(hidden_states)