删除DPSK_FP16_QUICK

vllm/model_executor/models/deepseek_v2.py

@@ -89,8 +89,6 @@ if current_platform.is_cuda_alike():
 elif current_platform.is_xpu():
     from vllm._ipex_ops import ipex_ops as ops
 
-os.environ['DPSK_FP16_QUICK'] = os.environ.get('DPSK_FP16_QUICK', '0')
-
 logger = init_logger(__name__)
 
 
@@ -191,7 +189,6 @@ class DeepseekV2MoE(nn.Module):
         # Load balancing settings.
         eplb_config = parallel_config.eplb_config
         self.enable_eplb = parallel_config.enable_eplb
-        self.dpsk_fp16_quick = os.environ.get('DPSK_FP16_QUICK') == '1'
 
         self.n_redundant_experts = eplb_config.num_redundant_experts
         self.n_logical_experts = self.n_routed_experts
@@ -288,7 +285,7 @@ class DeepseekV2MoE(nn.Module):
         
         # Fix FP16 overflow
         # See DeepseekV2DecoderLayer for more details.
-        if hidden_states.dtype != torch.float16 or self.dpsk_fp16_quick:
+        if hidden_states.dtype != torch.float16:
             final_hidden_states *= self.routed_scaling_factor
         elif self.shared_experts is not None:
             assert shared_output is not None
@@ -1054,7 +1051,6 @@ class DeepseekV2DecoderLayer(nn.Module):
             prefix=f"{prefix}.self_attn",
             topk_indices_buffer=topk_indices_buffer,
         )
-        self.dpsk_fp16_quick = os.environ.get('DPSK_FP16_QUICK') == '1'
         if (config.n_routed_experts is not None
                 and layer_idx >= config.first_k_dense_replace
                 and layer_idx % config.moe_layer_freq == 0):
@@ -1107,7 +1103,7 @@ class DeepseekV2DecoderLayer(nn.Module):
                 )
                 residual = new_residual
                 
-            if hidden_states.dtype == torch.float16 and not self.dpsk_fp16_quick:
+            if hidden_states.dtype == torch.float16:
                 # rmsnorm, and rmsnorm result would not affect by scale.
                 hidden_states *= 1. / self.routed_scaling_factor
                 if self.layer_idx == 0 or residual_fix_overflow:
@@ -1118,7 +1114,7 @@ class DeepseekV2DecoderLayer(nn.Module):
             hidden_states, new_resi = self.mlp(hidden_states, self.post_attention_layernorm.weight.data, residual)
 
             if isinstance(self.mlp,
-                        DeepseekV2MLP) and hidden_states.dtype == torch.float16 and not self.dpsk_fp16_quick:
+                        DeepseekV2MLP) and hidden_states.dtype == torch.float16:
                 # Fix FP16 overflow
                 # Scaling the DeepseekV2MLP output, it is the input of
                 # input_layernorm of next decoder layer.
@@ -1142,7 +1138,7 @@ class DeepseekV2DecoderLayer(nn.Module):
                 hidden_states=hidden_states,
             )
 
-            if hidden_states.dtype == torch.float16 and not self.dpsk_fp16_quick:
+            if hidden_states.dtype == torch.float16:
                 # Fix FP16 overflow
                 # We scale both hidden_states and residual before
                 # rmsnorm, and rmsnorm result would not affect by scale.
@@ -1157,7 +1153,7 @@ class DeepseekV2DecoderLayer(nn.Module):
                 hidden_states, residual)
             hidden_states = self.mlp(hidden_states)
             if isinstance(self.mlp,
-                        DeepseekV2MLP) and hidden_states.dtype == torch.float16 and not self.dpsk_fp16_quick:
+                        DeepseekV2MLP) and hidden_states.dtype == torch.float16:
                 # Fix FP16 overflow
                 # Scaling the DeepseekV2MLP output, it is the input of
                 # input_layernorm of next decoder layer.