simple optimization for moonlight moe decoding forward (#12891)

python/llm/src/ipex_llm/transformers/convert.py

@@ -2031,9 +2031,11 @@ def _optimize_post(model):
         from ipex_llm.transformers.models.common import rms_norm_forward
         from ipex_llm.transformers.models.deepseek import deepseek_model_forward
         from ipex_llm.transformers.models.deepseek import deepseek_attention_forward
+        from ipex_llm.transformers.models.deepseek import deepseek_moe_forward
         convert_forward(model, module.DeepseekV3RMSNorm, rms_norm_forward)
         convert_forward(model, module.DeepseekV3Model, deepseek_model_forward)
         convert_forward(model, module.DeepseekV3Attention, deepseek_attention_forward)
+        convert_forward(model, module.DeepseekV3MoE, deepseek_moe_forward)
 
     return model
 

python/llm/src/ipex_llm/transformers/models/deepseek.py

@@ -269,3 +269,35 @@ def deepseek_attention_forward(
         attn_weights = None
 
     return attn_output, attn_weights, past_key_value
+
+
+def moe_infer_decode(self, x: torch.Tensor, topk_ids: torch.Tensor, topk_weight: torch.Tensor):
+    idxs = topk_ids.flatten().tolist()
+    outputs = []
+    for i in idxs:
+        expert = self.experts[i]
+        expert_out = expert(x)
+        outputs.append(expert_out)
+    outs = torch.cat(outputs, dim=0)
+    reshaped_topk_weight = topk_weight.squeeze(0).unsqueeze(-1).to(outs.dtype)
+    final_out = (outs * reshaped_topk_weight).sum(dim=0, keepdim=True)
+    return final_out
+
+
+def deepseek_moe_forward(self, hidden_states: torch.Tensor):
+    identity = hidden_states
+    orig_shape = hidden_states.shape
+    topk_idx, topk_weight = self.gate(hidden_states)
+    hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+    flat_topk_idx = topk_idx.view(-1)
+    if not self.training:
+        # IPEX-LLM OPT start : add special moe_infer implementation for decoding
+        if topk_idx.size(0) == 1:
+            y = moe_infer_decode(self, hidden_states, topk_idx, topk_weight)
+        else:
+            y = self.moe_infer(hidden_states, topk_idx, topk_weight)
+        y = y.view(*orig_shape)
+        # IPEX-LLM OPT end
+    if self.config.n_shared_experts is not None:
+        y = y + self.shared_experts(identity)
+    return y