eplb quick-fix

Author: mercykid

vllm_ascend/eplb/adaptor/vllm_adaptor.py

@@ -194,20 +194,34 @@ def _export_tensor_to_file(self, expert_maps, expert_map_record_path: str):
                 json.dump(record, f, indent=4)
 
     def do_update_expert_map(self, layer_id, updated_expert_map):
-        self.expert_map_per_layer[layer_id] = updated_expert_map.clone()
-        self.expert_map_per_layer_cpu[layer_id] = updated_expert_map.clone()
+        pad_len = self.expert_map_per_layer[layer_id].shape[0] - updated_expert_map.shape[0]
+        updated_expert_map_padded = torch.nn.functional.pad(
+                                    updated_expert_map,
+                                    pad=(0,pad_len),
+                                    mode='constant',
+                                    value=-1
+                                    )
+        self.expert_map_per_layer[layer_id].copy_(updated_expert_map_padded)
+        self.expert_map_per_layer_cpu[layer_id].copy_(updated_expert_map)
 
     def do_update_expert_weight(self, layer_id, local_expert_to_replace,
                                 buffer_tensor_id):
         for expert_tensor, buffer_tensor in zip(
                 self.expert_param_per_layer[layer_id][local_expert_to_replace],
                 self.buffer_tensor_list[buffer_tensor_id]):
-            expert_tensor = buffer_tensor.clone()
+            expert_tensor.copy_(buffer_tensor)
             logger.debug(f"Expert tensor shape is :{expert_tensor.shape}")
 
     def do_update_log2phy_map(self, layer_id, updated_log2phy_map):
         if self.log2phy_map_per_layer[layer_id] is not None:
-            self.log2phy_map_per_layer[layer_id].copy_(updated_log2phy_map)
+            pad_len = self.log2phy_map_per_layer[layer_id].shape[0] - updated_log2phy_map.shape[0]
+            updated_log2phy_map_padded = torch.nn.functional.pad(
+                                        updated_log2phy_map,
+                                        pad=(0,pad_len),
+                                        mode='constant',
+                                        value=-1
+                                        )
+            self.log2phy_map_per_layer[layer_id].copy_(updated_log2phy_map_padded)
 
     def global2local(self, placement: torch.Tensor,
                      E_local: int) -> torch.Tensor:

vllm_ascend/eplb/core/eplb_device_transfer_loader.py

@@ -50,10 +50,6 @@ def generate_expert_d2d_transfer_task(self, expert_send_info,
             )
             return
 
-        # If neither send nor receive task is needed for this layer on this rank, return
-        if not (expert_send_info or expert_recv_info):
-            return
-
         self.updated_expert_map = updated_expert_map
 
         self.layer_id = layer_id

vllm_ascend/ops/fused_moe/moe_mlp.py

@@ -112,14 +112,16 @@ def quant_apply_mlp(hidden_states: torch.Tensor,
             if quantized_hidden_states is not None:
                 dispose_tensor(quantized_hidden_states)
             # act_fn: swiglu
+            group_diff = torch.diff(group_list, dim=0)
+            new_group = torch.cat([group_list[0].unsqueeze(0), group_diff],dim=0)
             hidden_states, swiglu_out_scale = torch_npu.npu_dequant_swiglu_quant(
                 x=hidden_states,
                 weight_scale=w1_scale,
                 activation_scale=pertoken_scale,
                 bias=None,
                 quant_scale=None,
                 quant_offset=None,
-                group_index=group_list,
+                group_index=new_group,
                 activate_left=True,
                 quant_mode=1,
             )

vllm_ascend/quantization/w8a8_dynamic.py

@@ -238,7 +238,7 @@ def apply(
             hidden_states=x,
             pertoken_scale=pertoken_scale,
             w1=layer.w13_weight,
-            w1_scale=layer.w13_weight_scale_fp32,
+            w1_scale=layer.w13_weight_scale.to(torch.float32),
             w2=layer.w2_weight,
             w2_scale=layer.w2_weight_scale,
             topk_weights=topk_weights,