bufix

Signed-off-by: tanqingshan (A) <50050625@china.huawei.com>

Author: tanqingshan (A)

vllm_ascend/eplb/adaptor/vllm_adaptor.py

@@ -221,7 +221,13 @@ 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].copy_(updated_expert_map)
+        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,
@@ -234,7 +240,15 @@ def do_update_expert_weight(self, layer_id, local_expert_to_replace,
 
     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:
@@ -313,4 +327,4 @@ def determine_expert_map_all(self):
             expert_map_all[:, r, start:end] = local_ids.unsqueeze(0).expand(
                 self.num_moe_layers, -1)
 
-        return expert_map_all
+        return expert_map_all

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
@@ -135,4 +131,4 @@ def update_expert_map_and_weight(self, reqs):
         self.state = ExpertWeightUpdateState.WAITING
 
     def load_impl(self, old_expert_table, new_expert_table):
-        raise NotImplementedError
+        raise NotImplementedError

vllm_ascend/ops/fused_moe/moe_mlp.py

@@ -127,14 +127,17 @@ 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,
             )
@@ -295,4 +298,4 @@ def unified_apply_mlp(hidden_states: torch.Tensor,
                                  group_list=group_list,
                                  group_list_type=group_list_type,
                                  topk_scales=topk_scales,
-                                 need_trans=need_trans)
+                                 need_trans=need_trans)