[Bugfix] Fix FP16 overflow for DeepSeek V2 (vllm-project#13232)

Signed-off-by: Yida Wu <[email protected]>
Signed-off-by: Louis Ulmer <[email protected]>

Author: Concurrensee

vllm/model_executor/models/deepseek_v2.py

@@ -155,11 +155,21 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
             shared_output = self.shared_experts(hidden_states)
         # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
-        final_hidden_states = self.experts(
-            hidden_states=hidden_states,
-            router_logits=router_logits) * self.routed_scaling_factor
+        if hidden_states.dtype != torch.float16:
+            final_hidden_states = self.experts(
+                hidden_states=hidden_states,
+                router_logits=router_logits) * self.routed_scaling_factor
+        else:
+            # This is a special case to avoid FP16 overflow
+            final_hidden_states = self.experts(hidden_states=hidden_states,
+                                               router_logits=router_logits)
         if shared_output is not None:
-            final_hidden_states = final_hidden_states + shared_output
+            if hidden_states.dtype != torch.float16:
+                final_hidden_states = final_hidden_states + shared_output
+            else:
+                # This is a special case to avoid FP16 overflow
+                final_hidden_states = final_hidden_states + shared_output \
+                    * (1. / self.routed_scaling_factor)
         if self.tp_size > 1:
             final_hidden_states = tensor_model_parallel_all_reduce(
                 final_hidden_states)
@@ -531,6 +541,7 @@ def __init__(
                                        eps=config.rms_norm_eps)
         self.post_attention_layernorm = RMSNorm(config.hidden_size,
                                                 eps=config.rms_norm_eps)
+        self.routed_scaling_factor = config.routed_scaling_factor
 
     def forward(
         self,
@@ -551,9 +562,18 @@ def forward(
         )
 
         # Fully Connected
+        if isinstance(self.mlp, DeepseekV2MoE) and \
+            hidden_states.dtype == torch.float16:
+            # This is a special case to avoid FP16 overflow
+            hidden_states *= 1. / self.routed_scaling_factor
         hidden_states, residual = self.post_attention_layernorm(
             hidden_states, residual)
         hidden_states = self.mlp(hidden_states)
+        if isinstance(self.mlp, DeepseekV2MLP) and \
+            hidden_states.dtype == torch.float16:
+            # This is a special case to avoid FP16 overflow
+            hidden_states *= 1. / self.routed_scaling_factor
+            residual *= 1. / self.routed_scaling_factor
         return hidden_states, residual