[Bugfix] fix deepseek fp16 scale bug (vllm-project#14809)

Signed-off-by: Jinzhen Lin <[email protected]>
Co-authored-by: mgoin <[email protected]>

Author: 2 people

vllm/model_executor/models/deepseek_v2.py

@@ -161,14 +161,16 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
                 hidden_states=hidden_states,
                 router_logits=router_logits) * self.routed_scaling_factor
         else:
-            # This is a special case to avoid FP16 overflow
+            # Fix FP16 overflow
+            # See DeepseekV2DecoderLayer for more details.
             final_hidden_states = self.experts(hidden_states=hidden_states,
                                                router_logits=router_logits)
         if shared_output is not None:
             if hidden_states.dtype != torch.float16:
                 final_hidden_states = final_hidden_states + shared_output
             else:
-                # This is a special case to avoid FP16 overflow
+                # Fix FP16 overflow
+                # See DeepseekV2DecoderLayer for more details.
                 final_hidden_states = final_hidden_states + shared_output \
                     * (1. / self.routed_scaling_factor)
         if self.tp_size > 1:
@@ -500,6 +502,7 @@ def __init__(
         # DecoderLayers are created with `make_layers` which passes the prefix
         # with the layer's index.
         layer_idx = int(prefix.split(sep='.')[-1])
+        self.layer_idx = layer_idx
         if model_config.use_mla:
             attn_cls = DeepseekV2MLAAttention
         else:
@@ -562,19 +565,30 @@ def forward(
             hidden_states=hidden_states,
         )
 
-        # Fully Connected
-        if isinstance(self.mlp, DeepseekV2MoE) and \
-            hidden_states.dtype == torch.float16:
-            # This is a special case to avoid FP16 overflow
+        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.
             hidden_states *= 1. / self.routed_scaling_factor
+            if self.layer_idx == 0:
+                # The residual is shared by all layers, we only scale it on
+                # first layer.
+                residual *= 1. / self.routed_scaling_factor
+
+        # Fully Connected
         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
+
+        if isinstance(self.mlp,
+                      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.
+            # The scaling of DeepseekV2MOE output would be done in the forward
+            # of DeepseekV2MOE
             hidden_states *= 1. / self.routed_scaling_factor
-            residual *= 1. / self.routed_scaling_factor
+
         return hidden_states, residual