Route MoE Promote (Post/Pre) update 0110

environment.yml

@@ -1,4 +1,4 @@
-name: minigptv
+name: promptmoe
 channels:
   - pytorch
   - defaults

minigpt4/configs/datasets/coco/caption.yaml

@@ -17,14 +17,14 @@ datasets:
           # md5: aa31ac474cf6250ebb81d18348a07ed8
           storage: 
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/COCO_Cap/coco_karpathy_train.json
-        val:
-          url: https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_val.json
-          storage: 
-              - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/COCO_Cap/coco_karpathy_val.json
-        test:
-          url: https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_test.json
-          storage: 
-              - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/COCO_Cap/coco_karpathy_test.json
+        # val:
+        #   url: https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_val.json
+        #   storage: 
+        #       - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/COCO_Cap/coco_karpathy_val.json
+        # test:
+        #   url: https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_test.json
+        #   storage: 
+        #       - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/COCO_Cap/coco_karpathy_test.json
 
       images:
         storage: /mnt/pfs-guan-ssai/nlu/dingyifeng/data/COCO

minigpt4/configs/datasets/coco/defaults_vqa_eval.yaml

@@ -20,6 +20,7 @@ datasets:
               - https://storage.googleapis.com/sfr-vision-language-research/LAVIS/datasets/vqav2/v2_mscoco_val2014_annotations.json
           storage:
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/VQAv2/vqa_val_eval.json
+              # - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/VQAv2/vqa_train.json
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/VQAv2/answer_list.json
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/VQAv2/v2_OpenEnded_mscoco_val2014_questions.json
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/VQAv2/v2_mscoco_val2014_annotations.json
@@ -29,6 +30,7 @@ datasets:
               - https://storage.googleapis.com/sfr-vision-language-research/LAVIS/datasets/vqav2/answer_list.json
           storage:
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/VQAv2/vqa_test.json
+              # - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/VQAv2/vqa_train.json
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/VQAv2/answer_list.json
         
       images:

minigpt4/configs/datasets/okvqa/eval.yaml

@@ -20,6 +20,7 @@ datasets:
               - https://storage.googleapis.com/sfr-vision-language-research/LAVIS/datasets/okvqa/mscoco_val2014_annotations.json
           storage:
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/OKVQA/okvqa_val_eval.json
+              # - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/OKVQA/okvqa_train.json
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/OKVQA/okvqa_answer_list_train.json
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/OKVQA/OpenEnded_mscoco_val2014_questions.json
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/OKVQA/mscoco_val2014_annotations.json
@@ -32,6 +33,7 @@ datasets:
               - https://storage.googleapis.com/sfr-vision-language-research/LAVIS/datasets/okvqa/mscoco_val2014_annotations.json
           storage:
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/OKVQA/okvqa_val_eval.json
+              # - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/OKVQA/okvqa_train.json
               # - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/OKVQA/okvqa_val_eval_part100.json
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/OKVQA/okvqa_answer_list_train.json
               - /mnt/pfs-guan-ssai/nlu/wanghanzi/data/OKVQA/OpenEnded_mscoco_val2014_questions.json

minigpt4/datasets/datasets/caption_datasets.py

@@ -105,6 +105,8 @@ class COCOCaptionDataset(BaseDataset, __DisplMixin):
             'Using language, provide a short account of the image.',
             'Use a few words to illustrate what is happening in the picture.',
         ]
+        self.source = 'coco_cap'
+        
     def __getitem__(self, index):
 
         # TODO this assumes image input, not general enough
@@ -118,13 +120,20 @@ class COCOCaptionDataset(BaseDataset, __DisplMixin):
         image = self.vis_processor(image)
         caption = self.text_processor(ann["caption"])
 
-        instruction = random.choice(self.instruction_pool)
-        instruction = "<Img><ImageHere></Img> [caption] {} ".format(instruction)
+        # instruction = random.choice(self.instruction_pool)
+        # instruction = "<Img><ImageHere></Img> [caption] {} ".format(instruction)
+        q_input = ""
+        llm_input = random.choice(self.instruction_pool)
 
         return {
             "image": image,
+            "image_id": ann["image"],
             "answer": caption,
-            "instruction_input": instruction,
+            "q_input": q_input,
+            "llm_input": llm_input,
+            "text_input": llm_input,
+            "text_output": caption,
+            "source": 'coco_cap',
         }
 
 class CaptionEvalDataset(BaseDataset, __DisplMixin):

minigpt4/datasets/datasets/coco_caption.py

@@ -31,6 +31,7 @@ class COCOCapEvalDataset(CaptionEvalDataset):
         split (string): val or test
         """
         super().__init__(vis_processor, text_processor, vis_root, ann_paths)
+        self.source = 'coco_cap'
 
     def __getitem__(self, index):
         ann = self.annotation[index]

minigpt4/datasets/datasets/dataloader_utils.py

@@ -31,7 +31,6 @@ class MultiIterLoader:
         if ratios is None:
             ratios = [1.0] * len(loaders)
         else:
-            # import pdb; pdb.set_trace()
             assert len(ratios) == len(loaders)
             ratios = [float(ratio) / sum(ratios) for ratio in ratios]
 

minigpt4/eval_scripts/eval_vqa.py

@@ -12,7 +12,6 @@ from tqdm import tqdm
 import torch
 from torch.utils.data import DataLoader
 import torch.backends.cudnn as cudnn
-from datasets import load_dataset
 import sys
 sys.path.append("/mnt/pfs-guan-ssai/nlu/wanghanzi/multimodal/PromptMoE")
 
@@ -248,6 +247,7 @@ if 'vsr' in args.dataset:
     img_path = cfg.evaluation_datasets_cfg["vsr"]["img_path"]
     batch_size = cfg.evaluation_datasets_cfg["vsr"]["batch_size"]
     max_new_tokens = cfg.evaluation_datasets_cfg["vsr"]["max_new_tokens"]
+    from datasets import load_dataset
 
     annotation = load_dataset("cambridgeltl/vsr_zeroshot", split='test')
     data = VSREvalData(annotation, vis_processor, img_path)

minigpt4/models/QformerMoE.py

@@ -386,17 +386,23 @@ class BertOutput(nn.Module): # Add & Norm
 
 
 class FeedForward(nn.Module):
+    # remove LayerNorm
     def __init__(self, config):
-        nn.Module.__init__(self)
-        # first layer
-        self.intermediate_query = BertIntermediate(config)
-        # second layer
-        self.output_query = BertOutput(config)
+        super().__init__()
+        self.dense1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+        self.dense2 = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob) # adjust dropout ratio 0.1->0.2
+        # self.dropout = nn.Dropout(0.2) # adjust dropout ratio 0.1->0.2
 
     def forward(self, hidden_states: Tensor):
-        input_tensor = hidden_states
-        intermediate_output = self.intermediate_query(hidden_states)
-        hidden_states = self.output_query(intermediate_output, input_tensor)
+        hidden_states = self.dense1(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        hidden_states = self.dense2(hidden_states)
+        hidden_states = self.dropout(hidden_states)
         return hidden_states
 
 
@@ -440,6 +446,7 @@ class BertLayer(nn.Module):
             )
         else:
             self.experts = ffn
+        self.expert_ln = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
 
     def forward(
         self,
@@ -494,7 +501,8 @@ class BertLayer(nn.Module):
             moe_ffn_attention_input = query_attention_output[:, :query_length, :]
             moe_ffn_attention_mask = attention_mask.squeeze(dim=1).squeeze(dim=1)[:, :query_length]
             layer_output = self.feed_forward_query_moe(moe_ffn_attention_input, moe_ffn_attention_mask) # layer_output, gate_loss, gate_load
-
+            # import pdb; pdb.set_trace() # test0107
+            
             if attention_output.shape[1] > query_length: # have text input in Qformer
                 layer_output_text = apply_chunking_to_forward(
                     self.feed_forward_chunk,
@@ -503,6 +511,7 @@ class BertLayer(nn.Module):
                     attention_output[:, query_length:, :],
                 )
                 layer_output = (torch.cat([layer_output[0], layer_output_text], dim=1), layer_output[1], layer_output[2])
+            
         else:
             layer_output = apply_chunking_to_forward(
                 self.feed_forward_chunk,
@@ -524,15 +533,14 @@ class BertLayer(nn.Module):
 
     def feed_forward_query_moe(self, attention_output, expert_attention_mask):
         if not self.use_experts:
-            layer_output = self.experts(attention_output)
+            hidden_states = self.experts(attention_output)
+            layer_output = self.expert_ln(hidden_states + attention_output)
             return layer_output, 0.0, []
 
-        # if not self.importance_processor.is_moe:
-        #     raise RuntimeError("Need to turn the model to a MoE first.")
-
-        layer_output, gate_loss, gate_load = self.experts(
+        hidden_states, gate_loss, gate_load = self.experts(
             attention_output, expert_attention_mask
         )
+        layer_output = self.expert_ln(hidden_states + attention_output)
         return layer_output, gate_loss, gate_load
 
 class BertEncoder(nn.Module):

minigpt4/models/QformerRouteMoE.py

@@ -46,10 +46,9 @@ from transformers.utils import logging
 from transformers.models.bert.configuration_bert import BertConfig
 
 from minigpt4.models.moe.utils import (
-    FeedForward,
     MoEModelOutput,
     MoEModelOutputWithPooling,
-    use_experts,
+    use_experts_route,
     moe_layer_judge,
 )
 from minigpt4.models.moe.route_moe_layer import RouteMoELayer
@@ -378,13 +377,14 @@ class BertOutput(nn.Module): # Add & Norm
     def __init__(self, config):
         super().__init__()
         self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
-        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)  # 1
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
 
     def forward(self, hidden_states, input_tensor):
         hidden_states = self.dense(hidden_states)
         hidden_states = self.dropout(hidden_states)
-        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        # Move LayerNorm & ResNet out of FFN After MoEFFN
+        hidden_states = self.LayerNorm(hidden_states + input_tensor) # 1
         return hidden_states
 
 
@@ -429,7 +429,7 @@ class BertLayer(nn.Module):
         self.output_query = BertOutput(config)
 
         # Add MoE FFN
-        self.use_experts = use_experts(layer_num)
+        self.use_experts = use_experts_route(layer_num)
         self.layer_judge = moe_layer_judge(layer_num)
         self.num_beams = config.moebert_num_beams
         ffn = FeedForward(config)
@@ -442,10 +442,13 @@ class BertLayer(nn.Module):
                 num_beams=config.moebert_num_beams,
                 layer_judge = self.layer_judge,
                 route_method=config.route_method,
+                weight_type=config.moe_weight_type,
             )
         else:
             self.experts = ffn
 
+        # self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
     def forward(
         self,
         hidden_states,
@@ -463,8 +466,8 @@ class BertLayer(nn.Module):
         self_attn_past_key_value = (
             past_key_value[:2] if past_key_value is not None else None
         )
-        # import pdb;pdb.set_trace()
-
+        # import pdb; pdb.set_trace() # 0107test
+        
         # adjust the dimension of hidden_states, attention_mask, encoder_attention_mask and encoder_hidden_states to be the same 
         if self.num_beams > 1:
             if hidden_states.shape[0]== attention_mask.shape[0]*self.num_beams:
@@ -494,10 +497,6 @@ class BertLayer(nn.Module):
 
         present_key_value = self_attention_outputs[-1]
 
-        # import pdb;pdb.set_trace()
-        # print(self.layer_num, hidden_states.shape, attention_mask.shape)
-
-
         if query_length > 0:
             query_attention_output = attention_output[:, :query_length, :]
 
@@ -526,7 +525,8 @@ class BertLayer(nn.Module):
             moe_ffn_attention_input = query_attention_output[:, :query_length, :]
             moe_ffn_attention_mask = attention_mask.squeeze(dim=1).squeeze(dim=1)[:, :query_length]
             layer_output = self.feed_forward_query_moe(moe_ffn_attention_input, moe_ffn_attention_mask, beam_scores, expert_route) 
-            # layer_output = (layer_output, beam_scores, expert_route, beam_idx)
+            # layer_output = (layer_output, beam_scores, expert_route, beam_idx, importance_loss)
+            # import pdb; pdb.set_trace() # 0107test
 
             if attention_output.shape[1] > query_length: # have text input in Qformer
                 layer_output_text = apply_chunking_to_forward(
@@ -535,7 +535,8 @@ class BertLayer(nn.Module):
                     self.seq_len_dim,
                     attention_output[:, query_length:, :],
                 )
-                if layer_output[0].shape[0] == layer_output_text.shape[0]*self.num_beams and self.num_beams>1:
+                if self.layer_judge == 'first' and self.num_beams>1:
+                    # if layer_output[0].shape[0] == layer_output_text.shape[0]*self.num_beams and self.num_beams>1:
                     # adjust the dimension of layer_output_text to bz*num_beams
                     layer_output_text = self.adjust_layer_output_text(layer_output_text)
 
@@ -550,7 +551,8 @@ class BertLayer(nn.Module):
                     # layer_output & layer_output_text dimen_0 from bz*num_beams to bz
                     layer_output, layer_output_text = self.route_moe_last_layer_top1(layer_output, layer_output_text)
 
-                layer_output = (torch.cat([layer_output[0], layer_output_text], dim=1), layer_output[1], layer_output[2])
+                layer_output = (torch.cat([layer_output[0], layer_output_text], dim=1), layer_output[1], layer_output[2], layer_output[3],layer_output[4])
+                # import pdb; pdb.set_trace() # 0107test
         
         else:
             layer_output = apply_chunking_to_forward(
@@ -559,7 +561,7 @@ class BertLayer(nn.Module):
                 self.seq_len_dim,
                 attention_output,
             )
-            layer_output = (layer_output, None, None)
+            layer_output = (layer_output, None, None, None, 0.0)
         
         outputs = (layer_output,) + outputs
 
@@ -594,24 +596,27 @@ class BertLayer(nn.Module):
         beam_scores_new = beam_scores[selects]
         expert_route_new = expert_route[selects]
 
-        return (hidden_states_new, beam_scores_new, expert_route_new), layer_output_text
+        return (hidden_states_new, beam_scores_new, expert_route_new, layer_output[3], layer_output[4]), layer_output_text
 
 
     def feed_forward_chunk(self, attention_output):
         intermediate_output = self.intermediate(attention_output)
         layer_output = self.output(intermediate_output, attention_output)
+        # layer_output = self.LayerNorm(layer_output + attention_output)
         return layer_output
 
     def feed_forward_query_moe(self, attention_output, expert_attention_mask, beam_scores, expert_route):
-        
         if not self.use_experts:
             layer_output = self.experts(attention_output)
-            return layer_output, None, None, None
+            # layer_output = self.LayerNorm(layer_output + attention_output)
+            return layer_output, None, None, None, 0.0
 
-        layer_output, beam_scores, expert_route, beam_idx = self.experts(
+        layer_output, beam_scores, expert_route, beam_idx, importance_loss = self.experts(
             attention_output, expert_attention_mask, beam_scores, expert_route
         )
-        return layer_output, beam_scores, expert_route, beam_idx
+
+        # layer_output = self.LayerNorm(layer_output + attention_output)
+        return layer_output, beam_scores, expert_route, beam_idx, importance_loss
 
 class BertEncoder(nn.Module):
     def __init__(self, config):
@@ -645,6 +650,7 @@ class BertEncoder(nn.Module):
         next_decoder_cache = () if use_cache else None
         beam_scores=None
         expert_route=None
+        importance_loss = 0
         for i in range(self.config.num_hidden_layers):
 
             layer_module = self.layer[i]
@@ -693,6 +699,7 @@ class BertEncoder(nn.Module):
             hidden_states = layer_outputs[0][0]
             beam_scores = beam_scores if layer_outputs[0][1] == None else layer_outputs[0][1]
             expert_route = expert_route if layer_outputs[0][2] == None else layer_outputs[0][2]
+            importance_loss += layer_outputs[0][4]
 
             if use_cache:
                 next_decoder_cache += (layer_outputs[-1],)
@@ -724,6 +731,7 @@ class BertEncoder(nn.Module):
             cross_attentions=all_cross_attentions,
             beam_scores=beam_scores,
             expert_route=expert_route,
+            gate_loss=importance_loss,
         )
 
 
@@ -1103,6 +1111,7 @@ class BertModel(BertPreTrainedModel):
             cross_attentions=encoder_outputs.cross_attentions,
             beam_scores=encoder_outputs.beam_scores,
             expert_route=encoder_outputs.expert_route,
+            gate_loss=encoder_outputs.gate_loss
         )
 
 

minigpt4/models/blip2.py

@@ -62,7 +62,7 @@ class Blip2Base(BaseModel):
         return Qformer, query_tokens
     
     @classmethod
-    def init_RouteMoEQformer(cls, num_query_token, vision_width, moebert_expert_num, moebert_num_beams, route_method, cross_attention_freq=2):
+    def init_RouteMoEQformer(cls, num_query_token, vision_width, moebert_expert_num, moebert_num_beams, route_method, moe_weight_type, cross_attention_freq=2):
         moe_encoder_config = BertConfig.from_pretrained("/mnt/pfs-guan-ssai/nlu/wanghanzi/models/bert-base-uncased")
 
         moe_encoder_config.encoder_width = vision_width
@@ -74,6 +74,7 @@ class Blip2Base(BaseModel):
         moe_encoder_config.moebert_expert_num = moebert_expert_num
         moe_encoder_config.moebert_num_beams = moebert_num_beams
         moe_encoder_config.route_method = route_method
+        moe_encoder_config.moe_weight_type = moe_weight_type
 
         RouteMoEQformer = BertMoERouteLMHeadModel.from_pretrained(
             "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/bert-base-uncased", config=moe_encoder_config

minigpt4/models/blip2_vicuna_instruct.py

@@ -99,6 +99,7 @@ class Blip2VicunaInstruct(Blip2Base):
                     moebert_expert_num=moebert_expert_num,
                     moebert_num_beams=moebert_num_beams,
                     route_method=moebert_route_method,
+                    moe_weight_type=moe_weight_type,
                     cross_attention_freq=2
                 )
             else:
@@ -118,7 +119,6 @@ class Blip2VicunaInstruct(Blip2Base):
                 num_query_token, self.visual_encoder.num_features
             )
 
-        # import pdb;pdb.set_trace()
         if not qformer_text_input:
             self.Qformer.bert.embeddings.word_embeddings = None
             self.Qformer.bert.embeddings.position_embeddings = None
@@ -178,6 +178,19 @@ class Blip2VicunaInstruct(Blip2Base):
                 if "_query" in name and "experts" not in name: # raw ffn_query not update
                     param.requires_grad = False
 
+                ln_pattern = r"bert\.encoder\.layer\.\d+\.expert_ln\.(weight|bias)"
+                if re.match(ln_pattern, name):
+                    key_orig = re.sub('expert_ln', 'output_query.LayerNorm', name)
+                    param.data.copy_(state_dict[key_orig])
+                d1_pattern = r"bert\.encoder\.layer\.(\d+)\.experts(\.|\.experts\.\d+\.)dense1\.(weight|bias)"
+                if re.match(d1_pattern, name):
+                    key_orig = re.sub(r'experts(\.|\.experts\.\d+\.)dense1', 'intermediate_query.dense', name)
+                    param.data.copy_(state_dict[key_orig])
+                d2_pattern = r"bert\.encoder\.layer\.(\d+)\.experts(\.|\.experts\.\d+\.)dense2\.(weight|bias)"
+                if re.match(d2_pattern, name):
+                    key_orig = re.sub(r'experts(\.|\.experts\.\d+\.)dense2', 'output_query.dense', name)
+                    param.data.copy_(state_dict[key_orig])
+
         # freeze qformer        
         if freeze_qformer:
             for name, param in self.Qformer.named_parameters():
@@ -205,6 +218,7 @@ class Blip2VicunaInstruct(Blip2Base):
         self.use_moeqformer = use_moeqformer
         self.use_route_moe = use_route_moe
         self.moebert_load_balance = moebert_load_balance
+        self.moebert_num_beams = moebert_num_beams
 
         self.gate_save_path = gate_save_path
         # if self.gate_save_path != None:
@@ -242,7 +256,7 @@ class Blip2VicunaInstruct(Blip2Base):
         # print(samples["text_input"])
         # print(samples["text_output"])
         # print('-----------------')
-        # import pdb;pdb.set_trace()
+        # import pdb;pdb.set_trace() # 0107test
         image = samples["image"]
         with self.maybe_autocast():
             image_embeds = self.ln_vision(self.visual_encoder(image))
@@ -278,10 +292,10 @@ class Blip2VicunaInstruct(Blip2Base):
                 return_dict=True,
                 output_hidden_states=True,
             )
-
+        # import pdb; pdb.set_trace()# 0107test
         query_output_to_linear = query_output.last_hidden_state[:,:query_tokens.size(1),:]
         
-        if self.use_moeqformer and not self.use_route_moe:
+        if self.use_moeqformer:
             gate_loss = query_output.gate_loss # only available in QformerMoE
         
         if self.gate_save_path != None:
@@ -312,7 +326,7 @@ class Blip2VicunaInstruct(Blip2Base):
                         # 'gate_route_1': prob_gate_normalized[0][i].tolist(),
                     })
                     # for layer in [6,8,10]:
-                    #     layer_data  = all_hidden_states[layer]
+                    #     layer_data  = all_hidden_states[layer]s
                     #     file_path = os.path.join(self.gate_save_path, f'{image_id}_{str(layer)}.npy')
                     #     x = layer_data.data.cpu().numpy()
                     #     np.save(file_path,x) 
@@ -323,7 +337,6 @@ class Blip2VicunaInstruct(Blip2Base):
                 print("Gate Save Error....")
                 print(e)
 
-
         inputs_llm = self.llm_proj(query_output_to_linear)
         atts_llm = torch.ones(inputs_llm.size()[:-1], dtype=torch.long).to(image.device)
 
@@ -380,7 +393,7 @@ class Blip2VicunaInstruct(Blip2Base):
                 labels=targets,
             )
 
-        if self.use_moeqformer and not self.use_route_moe:
+        if self.use_moeqformer:
             loss = outputs.loss + self.moebert_load_balance * gate_loss
         else:
             loss = outputs.loss
@@ -441,6 +454,8 @@ class Blip2VicunaInstruct(Blip2Base):
                 output_hidden_states=True,
             )
 
+        # import pdb; pdb.set_trace()
+
         if self.gate_save_path != None:
             all_hidden_states = query_output.hidden_states
             # prob_gate_normalized = query_output.gate_loads
@@ -471,11 +486,11 @@ class Blip2VicunaInstruct(Blip2Base):
                         # 'gate_route_3': prob_gate_normalized[2][i].tolist(),
                         # 'gate_route_1': prob_gate_normalized[0][i].tolist(),
                     })
-                    for layer in [6,8,10]:
-                        if layer == 6:
-                            layer_data  = all_hidden_states[layer][i, :32, :]
+                    for layer in [6,7,8,9,10,11]:
+                        if layer in [6,11]:
+                            layer_data  = all_hidden_states[layer][i, :, :]
                         else:
-                            layer_data  = all_hidden_states[layer][i*3, :32, :]
+                            layer_data  = all_hidden_states[layer][i*self.moebert_num_beams, :, :]
                         file_path = os.path.join(self.gate_save_path, f'{image_id}_{str(layer)}.npy')
                         x = layer_data.data.cpu().numpy()
                         np.save(file_path,x) # 大功告成
@@ -683,5 +698,6 @@ class Blip2VicunaInstruct(Blip2Base):
         for name, param in model.named_parameters():
             if param.requires_grad == True:
                 print(name)
-
+        # [name for name, param in model.named_parameters() if (param.requires_grad == False and 'Qformer' in name and 'intermediate_query' in name)]
+        # import pdb; pdb.set_trace()# 0107test
         return model

minigpt4/models/moe/beam_search.py

@@ -21,7 +21,6 @@ class MoELayer(nn.Module):
         else:
             raise KeyError("Routing method not supported.")
 
-
     def _forward_gate_sentence(self, x, attention_mask):
         """
             x: query_attention_output , torch.Size([bz, 32, 768])
@@ -77,7 +76,65 @@ class MoELayer(nn.Module):
         print('Layer Qformer MoE: \n',prob_gate)
         return moe_result, select_prob_gate, gate
 
+    def _forward_gate_sentence_post(self, x, attention_mask):
+        """
+            x: query_attention_output; torch.Size([bz, 32, 768])
+            attention_mask: torch.ones([bz, 32])
+            bz = 4
+            x = torch.randn(bz,32,768)
+            attention_mask = torch.ones([bz, 32])
 
+        """
+        attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
+        x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
+        
+        def forward_expert(input_x, expert_idx):
+            # input_x += torch.randn(4,32,768)
+            # return input_x
+            output_x = self.experts[expert_idx].forward(input_x)
+            return output_x
+
+        outputs = list()
+        logits_gate_lst = list()
+        for expert_idx in range(self.num_experts):
+            output_x = forward_expert(x_masked, expert_idx)
+            outputs.append(output_x.unsqueeze(0))
+            
+            output_x_aver = output_x.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz, 768])
+            # gate_acore = self.gates[expert_idx](output_x_aver)
+            gate_score = self.gate(output_x_aver)
+            logits_gate_lst.append(gate_score)
+
+        candidate_output = torch.cat(outputs) # torch.Size([num_expert, bz, 32, 768])
+        logits_gate = torch.cat(logits_gate_lst,dim=1)# torch.Size([bz, num_expert])
+        prob_gate = F.softmax(logits_gate, dim=-1) # torch.Size([bz, num_experts])
+        topk_values, gate = torch.topk(prob_gate, self.topk, dim=1) # gate, 每个样本被分配的expert: torch.Size([bz, topk])
+        num_sentences = F.one_hot(gate, self.num_experts).sum(1).gt(0).sum(0) # 每个expert被分配的样本数 torch.Size([num_expert])
+        gate_load = num_sentences.clone()
+
+        # load balancing loss
+        if self.use_balance_loss:
+            balance_loss = self._balancing_loss(prob_gate, num_sentences)
+        else:
+            balance_loss = 0.0
+
+        # importance loss
+        importance_loss = self._importance_auxiliary_loss(prob_gate)
+
+        # output_average = candidate_output.sum(2) / candidate_attn_mask.unsqueeze(-1).sum(2) # torch.Size([num_expert, bz, 768])
+        # output_average = torch.permute(output_average, (1, 0, 2)) # torch.Size([bz, num_expert, 768])
+        # logits_gate = self.gate(output_average) # torch.Size([bz, num_experts, 1])
+
+        prob_gate_topk = torch.zeros_like(prob_gate)
+        prob_gate_topk.scatter_(1, gate, topk_values)
+        prob_gate_normalized = prob_gate_topk / prob_gate_topk.sum(dim=1, keepdim=True) # torch.Size([bz, num_expert])
+        candidate_output_ad = torch.permute(candidate_output, (1, 0, 2, 3)) # torch.Size([bz, num_expert, 32, 768])
+        results = prob_gate_normalized.unsqueeze(-1).unsqueeze(-1) * candidate_output_ad # torch.Size([bz, num_expert, 32, 768])
+        moe_result = torch.sum(results, dim=1) # torch.Size([bz, 32, 768])
+        import pdb;pdb.set_trace()
+
+        return moe_result, (balance_loss+importance_loss), prob_gate_normalized
+    
     def forward(self, x, attention_mask):
         if self.route_method == "gate-token":
             x, balance_loss, gate_load = self._forward_gate_token(x)
@@ -95,7 +152,7 @@ class MoELayer(nn.Module):
     
 
 class RouteMoELayer(nn.Module):
-    def __init__(self, hidden_size, expert, gate, num_experts, num_beams=2, layer_judge=None, route_method="pre-route"):
+    def __init__(self, hidden_size, expert, num_experts, num_beams=2, layer_judge=None, route_method="pre-route", weight_type="ffn_prob"):
         # remove hash list
         nn.Module.__init__(self)
         self.num_experts = num_experts
@@ -103,13 +160,26 @@ class RouteMoELayer(nn.Module):
         self.num_beams = num_beams
         self.hidden_size = hidden_size
         self.layer_judge = layer_judge
+        self.weight_type = weight_type
 
         self.route_method = route_method
         if self.route_method == "pre-route":
             self.gate = nn.Linear(hidden_size, num_experts, bias=False).float()
         elif self.route_method == "post-route":
-            # gate = nn.Linear(hidden_size, 1, bias=False).float()
-            self.gates = nn.ModuleList([copy.deepcopy(gate) for i in range(num_experts)])
+            gate = nn.Linear(hidden_size, 1, bias=False).float()
+            self.gate = gate
+            # self.gates = nn.ModuleList([copy.deepcopy(gate) for i in range(num_experts)])
+
+    def _importance_auxiliary_loss(self, prob_gate):
+        # From VMOE
+        # _importance_auxiliary_loss
+        axis = tuple(range(prob_gate.ndim - 1))  # All except last.
+        importance_per_expert = torch.sum(prob_gate, dim=axis)
+        std_importance_per_expert = torch.std(importance_per_expert)
+        mean_importance_per_expert = torch.mean(importance_per_expert)
+        # Compute coefficient of variation (i.e. std/mean) squared.
+        return (std_importance_per_expert / mean_importance_per_expert)**2
+
 
     def forward_gate(self, x):
         """
@@ -123,19 +193,21 @@ class RouteMoELayer(nn.Module):
         prob_gate = F.softmax(logits_gate, dim=-1) #  torch.Size([bz*num_beams, num_experts])
         return prob_gate
 
-    def beam_search(self, current_scores_log, beam_scores, expert_route, batch_size):
-        import pdb;pdb.set_trace()
+
+    def beam_search_backup(self, current_scores_log, beam_scores, expert_route, batch_size):
         if self.layer_judge=='first' and self.route_method=='pre-route':
+            # current_scores_log torch.Size([bz, num_experts])
             assert beam_scores==None and expert_route==None
             current_scores = torch.exp(current_scores_log)
             topk_values, gate = torch.topk(current_scores, self.num_beams, dim=1) # gate, 每个样本被分配的expert: torch.Size([bz, topk])
             beam_scores = topk_values.view(self.num_beams * batch_size) # torch.Size([bz * num_beams])
             expert_route = gate.view(self.num_beams * batch_size).unsqueeze(1) # torch.Size([bz * num_beams,1])
-            beam_idx = None
+            beam_idx = torch.tensor(range(self.num_beams * batch_size))
+
         else:
             if self.layer_judge=='first' and self.route_method == 'post-route':
                 batch_size = batch_size
-                next_scores_raw1 = torch.exp(current_scores_log) # torch.Size([bz, num_experts])
+                next_scores_raw1 = torch.exp(current_scores_log) # torch.Size([bz, num_beams*num_experts])
             else:
                 batch_size = int(batch_size // self.num_beams)
                 next_scores_raw = current_scores_log + torch.log(beam_scores).unsqueeze(1)  # torch.Size([4*3, 5]) # 取log 之后，可以直接相加概率
@@ -147,9 +219,6 @@ class RouteMoELayer(nn.Module):
             next_scores, next_experts = torch.topk(next_scores_raw1, self.num_beams, dim=1, largest=True, sorted=True)
             # next_scores torch.Size([bz, num_beams])
             # next_tokens torch.Size([bz, num_beams])
-            print(next_scores_raw1)
-            print(next_scores)
-            print(next_experts)
 
             next_batch_beam = list()
             for batch_idx in range(batch_size):
@@ -166,7 +235,7 @@ class RouteMoELayer(nn.Module):
                 next_batch_beam.extend(next_sent_beam)
 
             import pdb;pdb.set_trace()
-
+            
             if self.layer_judge=='first' and self.route_method == 'post-route':
                 beam_scores = next_scores.view(self.num_beams * batch_size) # torch.Size([bz * num_beams])
                 expert_route = next_experts.view(self.num_beams * batch_size)
@@ -181,33 +250,91 @@ class RouteMoELayer(nn.Module):
                 pre_route = expert_route[beam_idx,:]
                 expert_route = torch.cat([pre_route, beam_experts.unsqueeze(1)], dim=-1)
 
-        import pdb;pdb.set_trace()
+        return beam_scores, expert_route, beam_idx
+    
+    def beam_search(self, current_scores_log, beam_scores, expert_route, batch_size):
+        if self.layer_judge=='first' and self.route_method in ['pre-route', 'post-route']:
+            # current_scores_log torch.Size([bz, num_experts])
+            assert beam_scores==None and expert_route==None
+            current_scores = torch.exp(current_scores_log)
+            topk_values, gate = torch.topk(current_scores, self.num_beams, dim=1) # gate, 每个样本被分配的expert: torch.Size([bz, topk])
+            beam_scores = topk_values.view(self.num_beams * batch_size) # torch.Size([bz * num_beams])
+            expert_route = gate.view(self.num_beams * batch_size).unsqueeze(1) # torch.Size([bz * num_beams,1])
+            beam_idx = torch.tensor(range(self.num_beams * batch_size))
+            import pdb;pdb.set_trace()
+
+        else:
+            batch_size = int(batch_size // self.num_beams)
+            next_scores_raw = current_scores_log + torch.log(beam_scores).unsqueeze(1)  # torch.Size([4*3, 5]) # 取log 之后，可以直接相加概率
+            next_scores_exp = torch.exp(next_scores_raw)
+            next_scores_raw1 = next_scores_exp.view(
+                batch_size, self.num_beams * self.num_experts
+            )  # torch.Size([bz, num_beams*num_experts])
+
+            next_scores, next_experts = torch.topk(next_scores_raw1, self.num_beams, dim=1, largest=True, sorted=True)
+            # next_scores torch.Size([bz, num_beams])
+            # next_tokens torch.Size([bz, num_beams])
+
+            next_batch_beam = list()
+            for batch_idx in range(batch_size):
+                next_sent_beam = list()
+                for rank, (expert_id, expert_score) in enumerate(
+                    zip(next_experts[batch_idx], next_scores[batch_idx])
+                ):
+                    expert_id = expert_id.item()
+                    beam_id = expert_id // self.num_experts
+                    ex_id = expert_id % self.num_experts
+                    effective_beam_id = batch_idx*self.num_beams + beam_id
+
+                    next_sent_beam.append((expert_score, ex_id, effective_beam_id))
+                next_batch_beam.extend(next_sent_beam)
+
+            # import pdb;pdb.set_trace()
+
+            beam_scores = beam_scores.new([x[0] for x in next_batch_beam])
+            beam_experts = expert_route[:,-1].new([x[1] for x in next_batch_beam])
+            beam_idx = expert_route[:,-1].new([x[2] for x in next_batch_beam])
+            pre_route = expert_route[beam_idx,:]
+            expert_route = torch.cat([pre_route, beam_experts.unsqueeze(1)], dim=-1)
+
+            print("next_scores_raw1:\n",next_scores_raw1)
 
         return beam_scores, expert_route, beam_idx
     
-    
-    def forward_expert_ffn(self, x, expert_select, beam_scores):
+
+
+    def forward_expert_ffn(self, x, expert_select, current_scores):
         """
             x_repeat : [bz*num_beams, 32,768]
             expert_select : [bz*num_beams]
+            current_scores : [bz*num_beams, num_experts] / [bz, num_experts]
         """
-        # add_1212 l2_normalization
-        # normalized_tensor = torch.nn.functional.normalize(beam_scores, p=2, dim=0) # L2 Normalization  torch.Size([bz, topk])
+        # add_1228 l2_normalization
+        # normalized_tensor = torch.nn.functional.normalize(current_scores, p=2, dim=0) # L2 Normalization  torch.Size([bz, topk])
         # tmp_prob = normalized_tensor.unsqueeze(-1).unsqueeze(-1)
-
+        import pdb;pdb.set_trace()
         outputs = list()
-        for i in range(x.shape[0]):
-            output_x = self.experts[expert_select[i]].forward(x[i])
-            outputs.append(output_x.unsqueeze(0))
-        candidate_output = torch.cat(outputs) 
+        for i  in range(self.num_experts):
+            output_x = self.experts[i].forward(x)
+            outputs.append(output_x.unsqueeze(1))
+        candidate_output = torch.cat(outputs, dim=1) 
+        expert_select_matrix = F.one_hot(expert_select, self.num_experts)
 
-        # candidate_output = candidate_output * tmp_prob
-        return candidate_output # torch.Size([bz*num_beams, 32, 768])
+        if self.weight_type == 'ffn_prob':
+            tmp_prob = current_scores * expert_select_matrix
+            candidate_output = candidate_output * tmp_prob.unsqueeze(-1).unsqueeze(-1)
+        else:
+            candidate_output = candidate_output * expert_select_matrix.unsqueeze(-1).unsqueeze(-1)
+        import pdb;pdb.set_trace()
+        output = torch.sum(candidate_output, dim=1)
 
+        return output # torch.Size([bz*num_beams, 32, 768])
 
     def forward_pre_route(self, x, beam_scores, expert_route, use_log=True):
-        
-        current_scores = self.forward_gate(x) # [bz*num_beams, 5]
+        import pdb;pdb.set_trace()
+        current_scores = self.forward_gate(x) # [bz, num_beams] / [bz*num_beams, num_beams]
+
+        importance_loss = self._importance_auxiliary_loss(current_scores)
 
         if use_log:
             current_scores_log = torch.log(current_scores) # 取log之后可以直接相加
@@ -215,42 +342,45 @@ class RouteMoELayer(nn.Module):
             current_scores_log = current_scores
 
         batch_size, num_tokens = x.shape[0], x.shape[1]
-        beam_scores, expert_route, _ = self.beam_search(current_scores_log, beam_scores, expert_route, batch_size)
-
+        beam_scores, expert_route, beam_idx = self.beam_search(current_scores_log, beam_scores, expert_route, batch_size)
         current_expert_select = expert_route[:,-1]
 
+        import pdb;pdb.set_trace()
+
         if self.layer_judge=='first': # expand first dim to batch_size * num_beams
             replicated_tensor = x.unsqueeze(1).expand(batch_size, self.num_beams, num_tokens, self.hidden_size)
             x = replicated_tensor.contiguous().view(-1, num_tokens, self.hidden_size) # [bz*num_beams, 32,768]
+            current_scores_t = current_scores.unsqueeze(1).expand(batch_size, self.num_beams, self.num_experts)
+            current_scores = current_scores_t.contiguous().view(-1, self.num_experts) # [bz*num_beams, num_experts]
 
-        candidate_output = self.forward_expert_ffn(x, current_expert_select, beam_scores) # [bz*num_beams, 32,768]
-        
-        return candidate_output, beam_scores, expert_route
+        input_x = x[beam_idx]
+        candidate_output = self.forward_expert_ffn(input_x, current_expert_select, current_scores) # [bz*num_beams, 32,768]
+        import pdb;pdb.set_trace()
 
+        return candidate_output, beam_scores, expert_route, beam_idx, importance_loss
 
     def forward_post_route(self, x, beam_scores, expert_route, use_log=True):
         
-        # if self.layer_judge=='first': # expand first dim to batch_size * num_beams
-        #     batch_size, num_tokens = x.shape[0], x.shape[1]
-        #     replicated_tensor = x.unsqueeze(1).expand(batch_size, self.num_beams, num_tokens, self.hidden_size)
-        #     x = replicated_tensor.contiguous().view(-1, num_tokens, self.hidden_size) # [bz*num_beams, 32,768]
-    
         attention_mask = torch.ones(x.shape[0], x.shape[1]).to(x.device)
         x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
-        
+
         def forward_expert(input_x, expert_idx):
             output_x = self.experts[expert_idx].forward(input_x)
             return output_x
 
+        import pdb; pdb.set_trace()
         outputs = list()
         logits_gate_lst = list()
         for expert_idx in range(self.num_experts):
             output_x = forward_expert(x_masked, expert_idx)
+            # output_x_aver = output_x.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz*num_beam, 768])
+            output_x_aver = torch.mean(output_x, dim=1)
+            # gate_score = self.gates[expert_idx](output_x_aver)
+            gate_score = self.gate(output_x_aver)
+            logits_gate_lst.append(gate_score)
             outputs.append(output_x.unsqueeze(0))
-            output_x_aver = output_x.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz*num_beam, 768])
-            gate_acore = self.gates[expert_idx](output_x_aver)
-            logits_gate_lst.append(gate_acore)
-        candidate_output = torch.cat(outputs) # torch.Size([num_expert, bz*num_beam, 32, 768])
+
+        candidate_output_raw = torch.cat(outputs) # torch.Size([num_expert, bz*num_beam, 32, 768])
         logits_gate = torch.cat(logits_gate_lst,dim=1)# torch.Size([bz*num_beam, num_expert])
         current_scores = F.softmax(logits_gate, dim=-1) # torch.Size([bz*num_beam, num_experts])
 
@@ -259,25 +389,39 @@ class RouteMoELayer(nn.Module):
         else:
             current_scores_log = current_scores
         
-        import pdb;pdb.set_trace()
+        # importance loss
+        importance_loss = self._importance_auxiliary_loss(current_scores)
+        
+        # import pdb; pdb.set_trace()
 
-        batch_size = x.shape[0] # bz*num_beam
+        batch_size, num_tokens = x.shape[0], x.shape[1] # bz*num_beam
         beam_scores, expert_route, beam_idx = self.beam_search(current_scores_log, beam_scores, expert_route, batch_size)
         # beam_scores torch.Size([bz*num_beam])
         # expert_route torch.Size([bz*num_beam, layer_n])
         current_select_expert = expert_route[:,-1]
+        # current_select_expert torch.Size([bz*num_beam, 1])
         
-        output = list()
-        for i in range(beam_idx.shape[0]):
-            b_idx = beam_idx[i]
-            ex_idx = current_select_expert[i]
-            ex_out = candidate_output[ex_idx, b_idx, :,:]
-            output.append(ex_out.unsqueeze(0))
-
-        final_output = torch.concat(output, dim=0)
-
-        return final_output, beam_scores, expert_route, beam_idx
+        # import pdb; pdb.set_trace()
+        
+        if self.layer_judge == 'first':
+            replicated_tensor = candidate_output_raw.unsqueeze(2).expand(self.num_experts, batch_size, self.num_beams, num_tokens, self.hidden_size)
+            candidate_output_raw = replicated_tensor.contiguous().view(self.num_experts, -1, num_tokens, self.hidden_size) # [bz*num_beams, 32,768]
+            current_scores_t = current_scores.unsqueeze(1).expand(batch_size, self.num_beams, self.num_experts)
+            current_scores = current_scores_t.contiguous().view(-1, self.num_experts) # [bz*num_beams, num_experts]
+        
+        candidate_output = candidate_output_raw.permute(1, 0, 2, 3)[beam_idx] # torch.Size([8, 2, 32, 768])
+        expert_select_matrix = F.one_hot(current_select_expert, self.num_experts)
+        if self.weight_type == 'ffn_prob':
+            tmp_prob = current_scores[beam_idx] * expert_select_matrix
+            output = candidate_output * tmp_prob.unsqueeze(-1).unsqueeze(-1)
+        else:
+            output = candidate_output * expert_select_matrix.unsqueeze(-1).unsqueeze(-1)
+        final_output = torch.sum(output, dim=1)
+        
+        import pdb; pdb.set_trace()
+        print("current_scores:\n",current_scores)
 
+        return final_output, beam_scores, expert_route, beam_idx, importance_loss
 
     def forward(self, x, attention_mask, beam_scores, expert_route, use_log=True):
         """
@@ -286,13 +430,12 @@ class RouteMoELayer(nn.Module):
         
         """
         if self.route_method == 'pre-route':
-            candidate_output, beam_scores, expert_route, _ = self.forward_pre_route(x, beam_scores, expert_route, use_log=True)
+            candidate_output, beam_scores, expert_route, beam_idx, importance_loss = self.forward_pre_route(x, beam_scores, expert_route, use_log=True)
         elif self.route_method == "post-route":
-            candidate_output, beam_scores, expert_route, beam_idx = self.forward_post_route(x, beam_scores, expert_route, use_log=True)
+            candidate_output, beam_scores, expert_route, beam_idx, importance_loss = self.forward_post_route(x, beam_scores, expert_route, use_log=True)
 
-        return candidate_output, beam_scores, expert_route, beam_idx
+        return candidate_output, beam_scores, expert_route, beam_idx, importance_loss
     
-
 if __name__ == '__main__':
 
     import sys
@@ -314,8 +457,8 @@ if __name__ == '__main__':
     config.add_cross_attention = True
     config.cross_attention_freq = cross_attention_freq
     config.query_length = num_query_token
-    config.moebert_expert_num = 3
-    config.moebert_num_beams = 3
+    config.moebert_expert_num = 2
+    config.moebert_num_beams = 2
     config.moebert_route_method = 'gate-sentence'
     config.moe_topk = 2
     config.use_balance_loss = False
@@ -332,40 +475,46 @@ if __name__ == '__main__':
     for layer_num in [6, 8, 10]:
         layer_judge = moe_layer_judge(layer_num)
         ffn = FeedForward(config)
-        gate = nn.Linear(768, config.moebert_expert_num, bias=False).float()
 
         # experts = RouteMoELayer(
         #             hidden_size=768,
         #             expert=ffn,
-        #             gate = gate,
         #             num_experts=config.moebert_expert_num,
         #             num_beams=config.moebert_num_beams,
         #             layer_judge = layer_judge,
-        #             route_method = "pre-route"
+        #             route_method = "pre-route",
+        #             weight_type="no_ffn_prob"
         #         )
         # layer_output = experts(x, None, beam_scores, expert_route)
-        # hidden_states1, beam_scores, expert_route,_ = layer_output
+        # hidden_states1, beam_scores, expert_route, beam_idx, importance_loss = layer_output
 
         # print(beam_scores)
         # print(expert_route)
+        # print(beam_idx)
+        # print(importance_loss)
+        # x = hidden_states1
 
         gate1 = nn.Linear(768, 1, bias=False).float()
         experts_post = RouteMoELayer(
                     hidden_size=768,
                     expert=ffn,
-                    gate = gate1,
                     num_experts=config.moebert_expert_num,
                     num_beams=config.moebert_num_beams,
                     layer_judge = layer_judge,
-                    route_method = "post-route"
+                    route_method = "post-route",
+                    weight_type="ffn_prob"
                 )
         layer_output = experts_post(x1, None, beam_scores1, expert_route1, False)
-        hidden_states2, beam_scores1, expert_route1, beam_idx = layer_output
+        hidden_states2, beam_scores1, expert_route1, beam_idx, importance_loss = layer_output
 
         print(beam_scores1)
         print(expert_route1)
         print(beam_idx)
+        print(importance_loss)
+        x1 = hidden_states2
 
+
+        # gate = nn.Linear(768, config.moebert_expert_num, bias=False).float()
         # experts_moe = MoELayer(
         #         hidden_size=config.hidden_size,
         #         expert=ffn,
@@ -382,11 +531,62 @@ if __name__ == '__main__':
         
         # print(select_prob_gate)
         # print(gate_load)
-
-
-        # x = hidden_states1
-        x1 = hidden_states2
         # x2 = hidden_states3
 
         print("------------------------------------")
+        import pdb; pdb.set_trace()
+
+
+
+    def forward_post_route_backup(self, x, beam_scores, expert_route, use_log=True):
+        
+        attention_mask = torch.ones(x.shape[0], x.shape[1]).to(x.device)
+        x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
+        
+        def forward_expert(input_x, expert_idx):
+            output_x = self.experts[expert_idx].forward(input_x)
+            return output_x
+
+        outputs = list()
+        logits_gate_lst = list()
+        for expert_idx in range(self.num_experts):
+            output_x = forward_expert(x_masked, expert_idx)
+            outputs.append(output_x.unsqueeze(0))
+            # output_x_aver = output_x.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz*num_beam, 768])
+            # gate_score = self.gates[expert_idx](output_x_aver)
+            output_x_aver = torch.mean(output_x, dim=1)
+            gate_score = self.gate(output_x_aver)
+            logits_gate_lst.append(gate_score)
+        candidate_output = torch.cat(outputs) # torch.Size([num_expert, bz*num_beam, 32, 768])
+        logits_gate = torch.cat(logits_gate_lst,dim=1)# torch.Size([bz*num_beam, num_expert])
+        current_scores = F.softmax(logits_gate, dim=-1) # torch.Size([bz*num_beam, num_experts])
+
+        if use_log:
+            current_scores_log = torch.log(current_scores) # 取log之后可以直接相加
+        else:
+            current_scores_log = current_scores
+        
+        # importance loss
+        importance_loss = self._importance_auxiliary_loss(current_scores)
+        
+        batch_size = x.shape[0] # bz*num_beam
+        beam_scores, expert_route, beam_idx = self.beam_search(current_scores_log, beam_scores, expert_route, batch_size)
+        # beam_scores torch.Size([bz*num_beam])
+        # expert_route torch.Size([bz*num_beam, layer_n])
+        current_select_expert = expert_route[:,-1]
+        # current_select_expert torch.Size([bz*num_beam, 1])
+        
+        output = list()
+        for i in range(beam_idx.shape[0]):
+            b_idx = beam_idx[i]
+            ex_idx = current_select_expert[i]
+            ex_out = candidate_output[ex_idx, b_idx, :,:]
+            if self.weight_type == 'ffn_prob':
+                prob = current_scores[b_idx, ex_idx]
+                ex_out = ex_out*prob
+            output.append(ex_out.unsqueeze(0))
+
+        final_output = torch.concat(output, dim=0)
+        # import pdb;pdb.set_trace()
+        return final_output, beam_scores, expert_route, beam_idx, importance_loss
 

minigpt4/models/moe/beam_search_test.py

@@ -1,155 +0,0 @@
-import torch
-import copy
-import pickle
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-
-device = torch.device("cuda:2" if torch.cuda.is_available() else "cpu")
-
-
-def forward_expert(input_x, expert_idx):
-    input_x += torch.randn(32,768)
-    return input_x
-    # output_x = self.experts[expert_idx].forward(input_x)
-    # return output_x
-
-
-def forward_ffn(x_repeat, expert_select):
-    """
-        x_repeat : [bz*num_beams, 32,768]
-        expert_select : [bz*num_beams]
-    """
-    outputs = list()
-    num_beams_bz = x_repeat.shape[0]
-    for i in range(num_beams_bz):
-        output_x = forward_expert(x_repeat[i], expert_select[i]) # (32,768)
-        outputs.append(output_x.unsqueeze(0))
-    candidate_output = torch.cat(outputs) 
-    return candidate_output # torch.Size([bz*num_beams, 32, 768])
-
-def forward_gate(x, num_expert):
-    """
-        x : torch.Size([bz*num_beams, 32, 768]) or torch.Size([bz, 32, 768]) 
-        prob_gate : torch.Size([bz*num_beams, num_experts])  or torch.Size([bz, num_experts]) 
-    """
-    # attention_mask = torch.ones(x.shape[0], x.shape[1]).to(x.device)
-    # x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz*num_beams, 32, 768])
-    # x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz*num_beams, 768])
-    # logits_gate = gate(x_average) # torch.Size([bz, num_experts])
-    logits_gate = torch.randn(x.shape[0], num_expert)
-    prob_gate = F.softmax(logits_gate, dim=-1) #  torch.Size([bz*num_beams, num_experts])
-    return prob_gate
-
-def beam_search(layer, current_scores, beam_scores, expert_route, num_beams):
-    if layer == 0 and beam_scores==None and expert_route==None:
-        topk_values, gate = torch.topk(current_scores, num_beams, dim=1) # gate, 每个样本被分配的expert: torch.Size([bz, topk])
-        beam_scores = topk_values.view(num_beams*batch_size) # torch.Size([bz * num_beams])
-        expert_route = gate.view(num_beams*batch_size).unsqueeze(1) # torch.Size([bz * num_beams])
-
-    else:
-        next_scores_raw = current_scores + beam_scores.unsqueeze(1)  # torch.Size([4*3, 5]) # 取log 之后，可以直接相加概率
-        next_scores_raw1 = next_scores_raw.view(
-            batch_size, num_beams * num_expert
-        )  # torch.Size([4, 3*5])
-        next_scores, next_experts = torch.topk(next_scores_raw1, num_beams, dim=1, largest=True, sorted=True)
-        # next_scores torch.Size([4, 3*num_beams])
-        # next_tokens torch.Size([4, 3*num_beams])
-
-        next_batch_beam = list()
-        for batch_idx in range(batch_size):
-            next_sent_beam = list()
-            print(batch_idx)
-            for rank, (expert_id, expert_score) in enumerate(
-                zip(next_experts[batch_idx], next_scores[batch_idx])
-            ):
-                expert_id = expert_id.item()
-                beam_id = expert_id // num_expert
-                ex_id = expert_id % num_expert
-                effective_beam_id = batch_idx*num_beams + beam_id
-
-                # print(expert_id, beam_id, ex_id, effective_beam_id, expert_score)
-
-                next_sent_beam.append((expert_score, ex_id, effective_beam_id))
-            next_batch_beam.extend(next_sent_beam)
-            
-            # print()
-
-        import pdb;pdb.set_trace()
-
-        beam_scores = beam_scores.new([x[0] for x in next_batch_beam])
-        beam_experts = expert_route[:,-1].new([x[1] for x in next_batch_beam])
-        beam_idx = expert_route[:,-1].new([x[2] for x in next_batch_beam])
-
-        pre_route = expert_route[beam_idx,:]
-        expert_route = torch.cat([pre_route, beam_experts.unsqueeze(1)], dim=-1)
-
-    return beam_scores, expert_route
-
-
-if __name__ == '__main__':
-
-    batch_size = 3
-    num_beams = 2
-    num_expert = 5
-    x = torch.randn(batch_size, 32, 768)
-    beam_scores, expert_route = None, None
-
-    for layer in range(0,3):
-        # import pdb;pdb.set_trace()
-
-        current_scores = forward_gate(x, num_expert)
-        import pdb;pdb.set_trace()
-
-        beam_scores, expert_route = beam_search(layer, current_scores, beam_scores, expert_route, num_beams)
-        current_expert_select = expert_route[:,-1]
-
-        if layer == 0:
-            replicated_tensor = x.unsqueeze(1).expand(batch_size, num_beams, 32, 768)
-            x = replicated_tensor.contiguous().view(-1, 32, 768) # [12,32,768] [bz*num_beams, 32,768]
-        else:
-            x = candidate_output
-
-        candidate_output = forward_ffn(x, current_expert_select) # torch.Size([4*3, 5])
-
-        x = candidate_output
-
-
-    scores = beam_scores.view(batch_size, num_beams)
-    topk_values, gate = torch.topk(scores, 1, dim=1)
-    # gate [batch_size, 1]
-    # topk_values [batch_size, 1]
-    selects = [ (bz_idx * num_beams + gate[bz_idx].item()) for bz_idx in range(batch_size)]
-    final_scores = beam_scores[selects]
-    final_expert_route = expert_route[selects]
-    final_output = candidate_output[selects]
-
-
-
-
-
-
-
-# def forward_ffn_post(x_repeat, expert_select):
-#     """
-#         x_repeat : [bz*num_beams, 32,768]
-#         expert_select : [bz*num_beams]
-#         prob_gate : torch.Size([bz*num_beams, num_experts])
-#     """
-#     outputs = list()
-#     logits_gate_lst = list()
-#     # attention_mask = torch.ones([batch_size, 32])
-#     for i in range(num_beams*batch_size):
-#         output_x = forward_expert(x_repeat[i], expert_select[i]) # (32,768)
-#         outputs.append(output_x.unsqueeze(0))
-#         # output_x_aver = output_x.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz, 768])
-#         # gate_acore = self.gates[expert_idx](output_x_aver)
-#         # gate_score = self.gate(output_x_aver)
-#         num_expert = 5
-#         gate_score = torch.randn(1,num_expert)
-#         logits_gate_lst.append(gate_score)
-
-#     candidate_output = torch.cat(outputs) # torch.Size([bz*num_beams, 32, 768])
-#     logits_gate = torch.cat(logits_gate_lst,dim=0)# torch.Size([bz*num_beams, num_expert])
-#     prob_gate = F.softmax(logits_gate, dim=-1) # torch.Size([bz*num_beams, num_experts])
-#     return prob_gate, candidate_output

minigpt4/models/moe/moe_layer.py

@@ -5,7 +5,7 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 class MoELayer(nn.Module):
-    def __init__(self, hidden_size, expert, num_experts, route_method, topk=1, use_balance_loss=True, weight_type='l2_norm'):
+    def __init__(self, hidden_size, expert, num_experts, route_method, topk=1, use_balance_loss=True, weight_type='raw_prob'):
         # remove hash list
         nn.Module.__init__(self)
         self.num_experts = num_experts
@@ -81,54 +81,6 @@ class MoELayer(nn.Module):
 
         return x, balance_loss, gate_load
 
-    def _forward_gate_sentence_top1_raw(self, x, attention_mask):
-        """
-            x: query_attention_output , torch.Size([bz, 32, 768])
-            attention_mask: torch.ones([bz, 32])
-            
-            ### Notice:
-            the raw version of expert_attention_mask is the extended_attention_mask, 
-            which will be add to attention_score directly
-            the values of extended_attention_mask are -0.0 or -10000
-            it should be adjust to 1/0 version to be processed by experts
-        """
-        attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
-        x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
-        x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz, 768])
-        logits_gate = self.gate(x_average) # torch.Size([bz, num_experts])
-        prob_gate = F.softmax(logits_gate, dim=-1) # torch.Size([bz, num_experts])
-        gate = torch.argmax(prob_gate, dim=-1) # torch.Size([bz])
-
-        order = gate.argsort(0)
-        num_sentences = F.one_hot(gate, self.num_experts).gt(0).sum(0)
-        gate_load = num_sentences.clone()
-        x = x[order]  # reorder according to expert number
-        x = x.split(num_sentences.tolist(), dim=0)  # a list of length self.num_experts
-
-        # compute the load balancing loss
-        P = prob_gate.mean(0)
-        temp = num_sentences.float()
-        f = temp / temp.sum(0, keepdim=True)
-        balance_loss = self.num_experts * torch.sum(P * f)
-
-        prob_gate = prob_gate.gather(dim=1, index=gate.unsqueeze(1))
-        prob_gate = prob_gate[order]
-        prob_gate = prob_gate.split(num_sentences.tolist(), dim=0)
-
-        def forward_expert(input_x, prob_x, expert_idx):
-            input_x = self.experts[expert_idx].forward(input_x)
-            input_x = input_x * prob_x.unsqueeze(-1)
-            return input_x
-
-        result = []
-        for i in range(self.num_experts):
-            if x[i].size(0) > 0:
-                result.append(forward_expert(x[i], prob_gate[i], i))
-        result = torch.vstack(result)
-        result = result[order.argsort(0)]  # restore original order
-
-        return result, balance_loss, gate_load
-
     def _forward_gate_sentence_post(self, x, attention_mask):
         """
             x: query_attention_output; torch.Size([bz, 32, 768])
@@ -174,13 +126,17 @@ class MoELayer(nn.Module):
         # importance loss
         importance_loss = self._importance_auxiliary_loss(prob_gate)
 
-        # output_average = candidate_output.sum(2) / candidate_attn_mask.unsqueeze(-1).sum(2) # torch.Size([num_expert, bz, 768])
-        # output_average = torch.permute(output_average, (1, 0, 2)) # torch.Size([bz, num_expert, 768])
-        # logits_gate = self.gate(output_average) # torch.Size([bz, num_experts, 1])
-
         prob_gate_topk = torch.zeros_like(prob_gate)
         prob_gate_topk.scatter_(1, gate, topk_values)
-        prob_gate_normalized = prob_gate_topk / prob_gate_topk.sum(dim=1, keepdim=True) # torch.Size([bz, num_expert])
+
+        if self.weight_type == 'average':
+            # torch.Size([bz, num_expert]) 未选中的expert prob_gate_norm为0
+            prob_gate_normalized = prob_gate_topk / prob_gate_topk.sum(dim=1, keepdim=True) 
+        elif self.weight_type == 'raw_prob':
+            prob_gate_normalized = prob_gate_topk
+        elif self.weight_type == 'softmax_norm':
+            prob_gate_normalized = F.softmax(prob_gate_topk, dim=-1) # torch.Size([bz, num_expert])
+
         candidate_output_ad = torch.permute(candidate_output, (1, 0, 2, 3)) # torch.Size([bz, num_expert, 32, 768])
         results = prob_gate_normalized.unsqueeze(-1).unsqueeze(-1) * candidate_output_ad # torch.Size([bz, num_expert, 32, 768])
         moe_result = torch.sum(results, dim=1) # torch.Size([bz, 32, 768])
@@ -188,6 +144,46 @@ class MoELayer(nn.Module):
 
         return moe_result, (balance_loss+importance_loss), prob_gate_normalized
 
+    def router(self, x, attention_mask):
+        # Prepare input x
+        attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
+        x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
+        x_average = torch.mean(x_masked, dim=1) # torch.Size([bz, 768])
+        
+        # Forward Gate
+        # logits_gate: [bz, num_experts]
+        logits_gate = self.gate(x_average)
+
+        # Probabilities for each sample of what expert it should be sent to.
+        # prob_gate: [bz, num_experts]
+        prob_gate = F.softmax(logits_gate, dim=-1)
+
+        # Get Top-K experts for each sample
+        # gate: [bz, topk]
+        # select_prob_gate: [bz, topk]
+        select_prob_gate, gate = torch.topk(prob_gate, self.topk, dim=1)
+
+        # Reshap Prob_gate & Gate
+        # expert_mask: [batch_size, topk, num_experts]
+        # expert_gate: [batch_size, topk, num_experts]
+        # combine_tensor: [batch_size, num_experts]
+        expert_mask = F.one_hot(gate, self.num_experts) 
+        expert_gate = select_prob_gate.unsqueeze(-1) * expert_mask
+        combine_tensor = torch.sum(expert_gate, dim=1) 
+
+        # Calculate Balancing Loss
+        if self.use_balance_loss:
+            num_sentences = F.one_hot(gate, self.num_experts).sum(1).gt(0).sum(0) # 每个expert被分配的样本数 torch.Size([num_expert])
+            balance_loss = self._balancing_loss(prob_gate, num_sentences)
+        else:
+            balance_loss = 0.0
+
+        # Calculate Importance Loss
+        importance_loss = self._importance_auxiliary_loss(prob_gate)
+
+        # import pdb; pdb.set_trace()
+
+        return expert_mask, combine_tensor, balance_loss, importance_loss
 
     def _forward_gate_sentence(self, x, attention_mask):
         """
@@ -200,81 +196,37 @@ class MoELayer(nn.Module):
             the values of extended_attention_mask are -0.0 or -10000
             it should be adjust to 1/0 version to be processed by experts
         """
-        attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
-        x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
-        x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz, 768])
-        logits_gate = self.gate(x_average) # torch.Size([bz, num_experts])
-        prob_gate = F.softmax(logits_gate, dim=-1) # torch.Size([bz, num_experts])
-        select_prob_gate, gate = torch.topk(prob_gate, self.topk, dim=1) # gate, 每个样本被分配的expert: torch.Size([bz, topk])
+        # Forward Router
+        expert_mask, combine_tensor, balance_loss, importance_loss = self.router(x, attention_mask)
+        
+        # Forward Expert FFN
+        result = []
+        for expert_idx in range(self.num_experts):
+            output_x = self.experts[expert_idx].forward(x)
+            result.append(output_x.unsqueeze(0))
+        expert_output = torch.cat(result).permute(1,0,2,3) # torch.Size([batch_size, num_expert, num_tokens, hidden_states])
 
-        # 这里用l2 norm 去加权
-        if self.weight_type == 'l2_norm':
-            normalized_tensor = torch.nn.functional.normalize(select_prob_gate, p=2, dim=0) # L2 Normalization  torch.Size([bz, topk])
-        elif self.weight_type == 'average':
-            normalized_tensor = select_prob_gate / select_prob_gate.sum(dim=1, keepdim=True)
+        # multiply outputs of experts by the routing probability
+        if self.weight_type == 'raw_prob':
+            expert_outputs_combined = expert_output * combine_tensor.unsqueeze(-1).unsqueeze(-1) # torch.Size([batch_size, num_expert, num_tokens, hidden_states])
+        elif self.weight_type == 'no_prob':
+            combine_index = torch.sum(expert_mask, dim=1)
+            expert_outputs_combined = expert_output * combine_index.unsqueeze(-1).unsqueeze(-1) # torch.Size([batch_size, num_expert, num_tokens, hidden_states])
 
-        num_sentences = F.one_hot(gate, self.num_experts).sum(1).gt(0).sum(0) # 每个expert被分配的样本数 torch.Size([num_expert])
-        gate_load = num_sentences.clone()
+        outputs = torch.sum(expert_outputs_combined, dim=1) # torch.Size([batch_size, num_tokens, hidden_states])
 
-        # load balancing loss
-        if self.use_balance_loss:
-            balance_loss = self._balancing_loss(prob_gate, num_sentences)
-        else:
-            balance_loss = 0.0
+        # import pdb; pdb.set_trace()
 
-        # importance loss
-        importance_loss = self._importance_auxiliary_loss(prob_gate)
-
-        # forward experts
-        def forward_expert(input_x, expert_idx):
-            input_x = self.experts[expert_idx].forward(input_x)
-            return input_x
-
-        result_lst = list()
-        for i in range(self.topk):
-            # top1、top2... 分别为一组，进行gate分组之后过expert，然后乘以概率后相加
-            tmp_gate = gate[:,i]
-            tmp_prob = normalized_tensor[:,i].unsqueeze(-1).unsqueeze(-1)
-            order = tmp_gate.argsort(0)
-            num_sentences_t = F.one_hot(tmp_gate, self.num_experts).gt(0).sum(0)
-            x1 = x[order]  # reorder according to expert number
-            x1 = x1.split(num_sentences_t.tolist(), dim=0)  # a list of length self.num_experts
-
-            result = []
-            for i in range(self.num_experts):
-                if x1[i].size(0) > 0:
-                    result.append(forward_expert(x1[i], i))
-            result = torch.vstack(result)
-            result = result[order.argsort(0)]  # restore original order
-            # result_lst.append(result * tmp_prob) # result * prob
-            result_lst.append(result) # result * prob # add_1212
-
-        moe_result = sum(result_lst)
-        # import pdb;pdb.set_trace()
-        return moe_result, (balance_loss+importance_loss), gate
-
-    def _forward_sentence_single_expert(self, x, attention_mask):
-        x_masked = x * attention_mask.unsqueeze(-1)
-        x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1)
-        logits_gate = self.gate(x_average)
-        prob_gate = F.softmax(logits_gate, dim=-1)
-        gate = torch.argmax(prob_gate, dim=-1)
-
-        gate_load = F.one_hot(gate, self.num_experts).gt(0).sum(0)
-        x = self.experts[gate.cpu().item()].forward(x)
-        return x, 0.0, gate_load
+        return outputs, (balance_loss+importance_loss), combine_tensor
 
     def forward(self, x, attention_mask):
         if self.route_method == "gate-token":
             x, balance_loss, gate_load = self._forward_gate_token(x)
         elif self.route_method == "gate-sentence":
-            if x.size(0) == 1:
-                x, balance_loss, gate_load = self._forward_sentence_single_expert(x, attention_mask)
-            else:
-                x, balance_loss, gate_load = self._forward_gate_sentence(x, attention_mask)
+            x, balance_loss, gate_load = self._forward_gate_sentence(x, attention_mask)
         elif self.route_method == "gate-sentence-post":
             x, balance_loss, gate_load = self._forward_gate_sentence_post(x, attention_mask)
         else:
             raise KeyError("Routing method not supported.")
-
+        # import pdb; pdb.set_trace()
         return x, balance_loss, gate_load

minigpt4/models/moe/moe_layer_backup.py

@@ -0,0 +1,330 @@
+import copy
+import pickle
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class MoELayer(nn.Module):
+    def __init__(self, hidden_size, expert, num_experts, route_method, topk=1, use_balance_loss=True, weight_type='l2_norm'):
+        # remove hash list
+        nn.Module.__init__(self)
+        self.num_experts = num_experts
+        self.experts = nn.ModuleList([copy.deepcopy(expert) for i in range(num_experts)])
+        self.route_method = route_method
+        self.topk = topk
+        self.use_balance_loss = use_balance_loss
+        self.weight_type = weight_type
+
+        if route_method in ["gate-token", "gate-sentence"]:
+            self.gate = nn.Linear(hidden_size, num_experts, bias=False).float()
+        elif route_method in ["gate-sentence-post"]:
+            gate = nn.Linear(hidden_size, 1, bias=False).float()
+            # self.gates = nn.ModuleList([copy.deepcopy(gate) for i in range(num_experts)])
+            self.gate = gate
+        else:
+            raise KeyError("Routing method not supported.")
+
+    def _balancing_loss(self, prob_gate, num_tokens):
+        # From MOEBERT
+        # compute the load balancing loss
+        # prob_gate，是 [bz, num_expert]，每个样本被分配给每个expert的概率
+        # 等价于 VMOE 中 _gshard_auxiliary_loss
+        P = prob_gate.mean(0) # torch.Size([num_expert]) 每个expert被分配到样本的平均概率
+        temp = num_tokens.float()
+        f = temp / temp.sum(0, keepdim=True) # 每个expert被分配的sample比例
+        balance_loss = self.num_experts * torch.sum(P * f) 
+        return balance_loss
+
+    def _importance_auxiliary_loss(self, prob_gate):
+        # From VMOE
+        # _importance_auxiliary_loss
+        axis = tuple(range(prob_gate.ndim - 1))  # All except last.
+        importance_per_expert = torch.sum(prob_gate, dim=axis)
+        std_importance_per_expert = torch.std(importance_per_expert)
+        mean_importance_per_expert = torch.mean(importance_per_expert)
+        # Compute coefficient of variation (i.e. std/mean) squared.
+        return (std_importance_per_expert / mean_importance_per_expert)**2
+
+    def _forward_gate_token(self, x):
+        bsz, seq_len, dim = x.size()
+
+        x = x.view(-1, dim)
+        logits_gate = self.gate(x)
+        prob_gate = F.softmax(logits_gate, dim=-1)
+        gate = torch.argmax(prob_gate, dim=-1)
+
+        order = gate.argsort(0)
+        num_tokens = F.one_hot(gate, self.num_experts).gt(0).sum(0)
+        gate_load = num_tokens.clone()
+        x = x[order]  # reorder according to expert number
+        x = x.split(num_tokens.tolist(), dim=0)  # a list of length self.num_experts
+
+        # compute the load balancing loss
+        P = prob_gate.mean(0)
+        temp = num_tokens.float()
+        f = temp / temp.sum(0, keepdim=True)
+        balance_loss = self.num_experts * torch.sum(P * f)
+
+        prob_gate = prob_gate.gather(dim=1, index=gate.unsqueeze(1))
+        prob_gate = prob_gate[order]
+        prob_gate = prob_gate.split(num_tokens.tolist(), dim=0)
+
+        def forward_expert(input_x, prob_x, expert_idx):
+            input_x = self.experts[expert_idx].forward(input_x)
+            input_x = input_x * prob_x
+            return input_x
+
+        x = [forward_expert(x[i], prob_gate[i], i) for i in range(self.num_experts)]
+        x = torch.vstack(x)
+        x = x[order.argsort(0)]  # restore original order
+        x = x.view(bsz, seq_len, dim)
+
+        return x, balance_loss, gate_load
+
+    def _forward_gate_sentence_top1_raw(self, x, attention_mask):
+        """
+            x: query_attention_output , torch.Size([bz, 32, 768])
+            attention_mask: torch.ones([bz, 32])
+            
+            ### Notice:
+            the raw version of expert_attention_mask is the extended_attention_mask, 
+            which will be add to attention_score directly
+            the values of extended_attention_mask are -0.0 or -10000
+            it should be adjust to 1/0 version to be processed by experts
+        """
+        attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
+        x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
+        x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz, 768])
+        logits_gate = self.gate(x_average) # torch.Size([bz, num_experts])
+        prob_gate = F.softmax(logits_gate, dim=-1) # torch.Size([bz, num_experts])
+        gate = torch.argmax(prob_gate, dim=-1) # torch.Size([bz])
+
+        order = gate.argsort(0)
+        num_sentences = F.one_hot(gate, self.num_experts).gt(0).sum(0)
+        gate_load = num_sentences.clone()
+        x = x[order]  # reorder according to expert number
+        x = x.split(num_sentences.tolist(), dim=0)  # a list of length self.num_experts
+
+        # compute the load balancing loss
+        P = prob_gate.mean(0)
+        temp = num_sentences.float()
+        f = temp / temp.sum(0, keepdim=True)
+        balance_loss = self.num_experts * torch.sum(P * f)
+
+        prob_gate = prob_gate.gather(dim=1, index=gate.unsqueeze(1))
+        prob_gate = prob_gate[order]
+        prob_gate = prob_gate.split(num_sentences.tolist(), dim=0)
+
+        def forward_expert(input_x, prob_x, expert_idx):
+            input_x = self.experts[expert_idx].forward(input_x)
+            input_x = input_x * prob_x.unsqueeze(-1)
+            return input_x
+
+        result = []
+        for i in range(self.num_experts):
+            if x[i].size(0) > 0:
+                result.append(forward_expert(x[i], prob_gate[i], i))
+        result = torch.vstack(result)
+        result = result[order.argsort(0)]  # restore original order
+
+        return result, balance_loss, gate_load
+
+    def _forward_gate_sentence_post(self, x, attention_mask):
+        """
+            x: query_attention_output; torch.Size([bz, 32, 768])
+            attention_mask: torch.ones([bz, 32])
+            bz = 4
+            x = torch.randn(bz,32,768)
+            attention_mask = torch.ones([bz, 32])
+
+        """
+        attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
+        x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
+        
+        def forward_expert(input_x, expert_idx):
+            # input_x += torch.randn(4,32,768)
+            # return input_x
+            output_x = self.experts[expert_idx].forward(input_x)
+            return output_x
+
+        outputs = list()
+        logits_gate_lst = list()
+        for expert_idx in range(self.num_experts):
+            output_x = forward_expert(x_masked, expert_idx)
+            outputs.append(output_x.unsqueeze(0))
+            
+            output_x_aver = output_x.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz, 768])
+            # gate_acore = self.gates[expert_idx](output_x_aver)
+            gate_score = self.gate(output_x_aver)
+            logits_gate_lst.append(gate_score)
+
+        candidate_output = torch.cat(outputs) # torch.Size([num_expert, bz, 32, 768])
+        logits_gate = torch.cat(logits_gate_lst,dim=1)# torch.Size([bz, num_expert])
+        prob_gate = F.softmax(logits_gate, dim=-1) # torch.Size([bz, num_experts])
+        topk_values, gate = torch.topk(prob_gate, self.topk, dim=1) # gate, 每个样本被分配的expert: torch.Size([bz, topk])
+        num_sentences = F.one_hot(gate, self.num_experts).sum(1).gt(0).sum(0) # 每个expert被分配的样本数 torch.Size([num_expert])
+        gate_load = num_sentences.clone()
+
+        # load balancing loss
+        if self.use_balance_loss:
+            balance_loss = self._balancing_loss(prob_gate, num_sentences)
+        else:
+            balance_loss = 0.0
+
+        # importance loss
+        importance_loss = self._importance_auxiliary_loss(prob_gate)
+
+        # output_average = candidate_output.sum(2) / candidate_attn_mask.unsqueeze(-1).sum(2) # torch.Size([num_expert, bz, 768])
+        # output_average = torch.permute(output_average, (1, 0, 2)) # torch.Size([bz, num_expert, 768])
+        # logits_gate = self.gate(output_average) # torch.Size([bz, num_experts, 1])
+
+        prob_gate_topk = torch.zeros_like(prob_gate)
+        prob_gate_topk.scatter_(1, gate, topk_values)
+
+        if self.weight_type == 'average':
+            # torch.Size([bz, num_expert]) 未选中的expert prob_gate_norm为0
+            prob_gate_normalized = prob_gate_topk / prob_gate_topk.sum(dim=1, keepdim=True) 
+        elif self.weight_type == 'raw_prob':
+            prob_gate_normalized = prob_gate_topk
+        elif self.weight_type == 'softmax_norm':
+            prob_gate_normalized = F.softmax(prob_gate_topk, dim=-1) # torch.Size([bz, num_expert])
+
+        candidate_output_ad = torch.permute(candidate_output, (1, 0, 2, 3)) # torch.Size([bz, num_expert, 32, 768])
+        results = prob_gate_normalized.unsqueeze(-1).unsqueeze(-1) * candidate_output_ad # torch.Size([bz, num_expert, 32, 768])
+        moe_result = torch.sum(results, dim=1) # torch.Size([bz, 32, 768])
+        # import pdb;pdb.set_trace()
+
+        return moe_result, (balance_loss+importance_loss), prob_gate_normalized
+
+    # def _forward_gate_sentence(self, x, attention_mask):
+
+    #     attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
+    #     x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
+    #     x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1)
+    #     logits_gate = self.gate(x_average)
+    #     prob_gate = F.softmax(logits_gate, dim=-1)
+    #     gate = torch.argmax(prob_gate, dim=-1)
+
+    #     order = gate.argsort(0)
+    #     num_sentences = F.one_hot(gate, self.num_experts).gt(0).sum(0)
+    #     gate_load = num_sentences.clone()
+    #     x = x[order]  # reorder according to expert number
+    #     x = x.split(num_sentences.tolist(), dim=0)  # a list of length self.num_experts
+
+    #     # compute the load balancing loss
+    #     P = prob_gate.mean(0)
+    #     temp = num_sentences.float()
+    #     f = temp / temp.sum(0, keepdim=True)
+    #     balance_loss = self.num_experts * torch.sum(P * f)
+
+    #     prob_gate = prob_gate.gather(dim=1, index=gate.unsqueeze(1))
+    #     prob_gate = prob_gate[order]
+    #     prob_gate = prob_gate.split(num_sentences.tolist(), dim=0)
+
+    #     def forward_expert(input_x, prob_x, expert_idx):
+    #         input_x = self.experts[expert_idx].forward(input_x)
+    #         input_x = input_x * prob_x.unsqueeze(-1)
+    #         return input_x
+
+    #     result = []
+    #     for i in range(self.num_experts):
+    #         if x[i].size(0) > 0:
+    #             result.append(forward_expert(x[i], prob_gate[i], i))
+    #     result = torch.vstack(result)
+    #     result = result[order.argsort(0)]  # restore original order
+
+    #     return result, balance_loss, gate_load
+
+    def _forward_gate_sentence(self, x, attention_mask):
+        """
+            x: query_attention_output , torch.Size([bz, 32, 768])
+            attention_mask: torch.ones([bz, 32])
+            
+            ### Notice:
+            the raw version of expert_attention_mask is the extended_attention_mask, 
+            which will be add to attention_score directly
+            the values of extended_attention_mask are -0.0 or -10000
+            it should be adjust to 1/0 version to be processed by experts
+        """
+        attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
+        x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
+        x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz, 768])
+        logits_gate = self.gate(x_average) # torch.Size([bz, num_experts])
+        prob_gate = F.softmax(logits_gate, dim=-1) # torch.Size([bz, num_experts])
+        select_prob_gate, gate = torch.topk(prob_gate, self.topk, dim=1) # gate, 每个样本被分配的expert: torch.Size([bz, topk])
+
+        # 这里用l2 norm 去加权
+        if self.weight_type == 'l2_norm':
+            # actually neigther dim=0 nor dim=1 is right
+            normalized_tensor = torch.nn.functional.normalize(select_prob_gate, p=2, dim=1) # L2 Normalization  torch.Size([bz, topk])
+        elif self.weight_type == 'l2_norm_0':
+            normalized_tensor = torch.nn.functional.normalize(select_prob_gate, p=2, dim=0) # L2 Normalization  torch.Size([bz, topk])
+        elif self.weight_type == 'average':
+            normalized_tensor = select_prob_gate / select_prob_gate.sum(dim=1, keepdim=True)
+        elif self.weight_type == 'raw_prob':
+            normalized_tensor = select_prob_gate
+
+        num_sentences = F.one_hot(gate, self.num_experts).sum(1).gt(0).sum(0) # 每个expert被分配的样本数 torch.Size([num_expert])
+        gate_load = num_sentences.clone()
+
+        # load balancing loss
+        if self.use_balance_loss:
+            balance_loss = self._balancing_loss(prob_gate, num_sentences)
+        else:
+            balance_loss = 0.0
+
+        # importance loss
+        importance_loss = self._importance_auxiliary_loss(prob_gate)
+
+        # forward experts
+        def forward_expert(input_x, expert_idx):
+            input_x = self.experts[expert_idx].forward(input_x)
+            return input_x
+
+        result_lst = list()
+        for i in range(self.topk):
+            # top1、top2... 分别为一组，进行gate分组之后过expert，然后乘以概率后相加
+            tmp_gate = gate[:,i]
+            tmp_prob = normalized_tensor[:,i].unsqueeze(-1).unsqueeze(-1)
+            order = tmp_gate.argsort(0)
+            num_sentences_t = F.one_hot(tmp_gate, self.num_experts).gt(0).sum(0)
+            x1 = x[order]  # reorder according to expert number
+            x1 = x1.split(num_sentences_t.tolist(), dim=0)  # a list of length self.num_experts
+
+            result = []
+            for i in range(self.num_experts):
+                if x1[i].size(0) > 0:
+                    result.append(forward_expert(x1[i], i))
+            result = torch.vstack(result)
+            result = result[order.argsort(0)]  # restore original order
+            result_lst.append(result * tmp_prob) # result * prob
+            # result_lst.append(result) # result * prob # add_1212
+
+        moe_result = sum(result_lst)
+        return moe_result, (balance_loss+importance_loss), gate
+
+    def _forward_sentence_single_expert(self, x, attention_mask):
+        x_masked = x * attention_mask.unsqueeze(-1)
+        x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1)
+        logits_gate = self.gate(x_average)
+        prob_gate = F.softmax(logits_gate, dim=-1)
+        gate = torch.argmax(prob_gate, dim=-1)
+
+        gate_load = F.one_hot(gate, self.num_experts).gt(0).sum(0)
+        x = self.experts[gate.cpu().item()].forward(x)
+        return x, 0.0, gate_load
+
+    def forward(self, x, attention_mask):
+        if self.route_method == "gate-token":
+            x, balance_loss, gate_load = self._forward_gate_token(x)
+        elif self.route_method == "gate-sentence":
+            if x.size(0) == 1:
+                x, balance_loss, gate_load = self._forward_sentence_single_expert(x, attention_mask)
+            else:
+                x, balance_loss, gate_load = self._forward_gate_sentence(x, attention_mask)
+        elif self.route_method == "gate-sentence-post":
+            x, balance_loss, gate_load = self._forward_gate_sentence_post(x, attention_mask)
+        else:
+            raise KeyError("Routing method not supported.")
+        # import pdb; pdb.set_trace()
+        return x, balance_loss, gate_load

minigpt4/models/moe/prompt_moe.py

@@ -92,7 +92,6 @@ class PrePromptMoE(PromptMoEBase):
         self.topk = topk
         if route_method in ["gate-token", "gate-single-token", "gate-sentence"]:
             self.gate = nn.Linear(hidden_size, num_experts, bias=False).float()
-            print(self.gate)
         else:
             raise KeyError("Routing method not supported.")
 

minigpt4/models/moe/route_moe_layer.py

@@ -5,7 +5,7 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 class RouteMoELayer(nn.Module):
-    def __init__(self, hidden_size, expert, num_experts, num_beams=2, layer_judge=None, route_method="pre-route"):
+    def __init__(self, hidden_size, expert, num_experts, num_beams=2, layer_judge=None, route_method="pre-route", weight_type="ffn_prob"):
         # remove hash list
         nn.Module.__init__(self)
         self.num_experts = num_experts
@@ -13,6 +13,7 @@ class RouteMoELayer(nn.Module):
         self.num_beams = num_beams
         self.hidden_size = hidden_size
         self.layer_judge = layer_judge
+        self.weight_type = weight_type
 
         self.route_method = route_method
         if self.route_method == "pre-route":
@@ -22,6 +23,17 @@ class RouteMoELayer(nn.Module):
             self.gate = gate
             # self.gates = nn.ModuleList([copy.deepcopy(gate) for i in range(num_experts)])
 
+    def _importance_auxiliary_loss(self, prob_gate):
+        # From VMOE
+        # _importance_auxiliary_loss
+        axis = tuple(range(prob_gate.ndim - 1))  # All except last.
+        importance_per_expert = torch.sum(prob_gate, dim=axis)
+        std_importance_per_expert = torch.std(importance_per_expert)
+        mean_importance_per_expert = torch.mean(importance_per_expert)
+        # Compute coefficient of variation (i.e. std/mean) squared.
+        return (std_importance_per_expert / mean_importance_per_expert)**2
+
+
     def forward_gate(self, x):
         """
             x : torch.Size([bz*num_beams, 32, 768]) or torch.Size([bz, 32, 768]) 
@@ -29,7 +41,8 @@ class RouteMoELayer(nn.Module):
         """
         attention_mask = torch.ones(x.shape[0], x.shape[1]).to(x.device)
         x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz*num_beams, 32, 768])
-        x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz*num_beams, 768])
+        # x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz*num_beams, 768])
+        x_average = torch.mean(x_masked, dim=1) # torch.Size([bz*num_beams, 768])
         logits_gate = self.gate(x_average) # torch.Size([bz*num_beams, num_experts])
         prob_gate = F.softmax(logits_gate, dim=-1) #  torch.Size([bz*num_beams, num_experts])
         return prob_gate
@@ -42,7 +55,7 @@ class RouteMoELayer(nn.Module):
             topk_values, gate = torch.topk(current_scores, self.num_beams, dim=1) # gate, 每个样本被分配的expert: torch.Size([bz, topk])
             beam_scores = topk_values.view(self.num_beams * batch_size) # torch.Size([bz * num_beams])
             expert_route = gate.view(self.num_beams * batch_size).unsqueeze(1) # torch.Size([bz * num_beams,1])
-            beam_idx = None
+            beam_idx = torch.tensor(range(self.num_beams * batch_size))
         else:
             if self.layer_judge=='first' and self.route_method == 'post-route':
                 batch_size = batch_size
@@ -89,54 +102,63 @@ class RouteMoELayer(nn.Module):
 
         return beam_scores, expert_route, beam_idx
     
-       
-    def forward_expert_ffn(self, x, expert_select, beam_scores):
+    def forward_expert_ffn(self, x, expert_select, current_scores):
         """
             x_repeat : [bz*num_beams, 32,768]
             expert_select : [bz*num_beams]
+            current_scores : [bz*num_beams, num_experts] / [bz, num_experts]
         """
-        # add_1212 l2_normalization
-        # normalized_tensor = torch.nn.functional.normalize(beam_scores, p=2, dim=0) # L2 Normalization  torch.Size([bz, topk])
+        # add_1228 l2_normalization
+        # normalized_tensor = torch.nn.functional.normalize(current_scores, p=2, dim=0) # L2 Normalization  torch.Size([bz, topk])
         # tmp_prob = normalized_tensor.unsqueeze(-1).unsqueeze(-1)
-
+        # import pdb;pdb.set_trace()
         outputs = list()
-        for i in range(x.shape[0]):
-            output_x = self.experts[expert_select[i]].forward(x[i])
-            outputs.append(output_x.unsqueeze(0))
-        candidate_output = torch.cat(outputs) 
-
-        # candidate_output = candidate_output * tmp_prob
-        return candidate_output # torch.Size([bz*num_beams, 32, 768])
-
+        for i  in range(self.num_experts):
+            output_x = self.experts[i].forward(x)
+            outputs.append(output_x.unsqueeze(1))
+        candidate_output = torch.cat(outputs, dim=1) 
+        expert_select_matrix = F.one_hot(expert_select, self.num_experts)
+        if self.weight_type == 'ffn_prob':
+            tmp_prob = current_scores * expert_select_matrix
+            candidate_output = candidate_output * tmp_prob.unsqueeze(-1).unsqueeze(-1)
+        else:
+            candidate_output = candidate_output * expert_select_matrix.unsqueeze(-1).unsqueeze(-1)
+        output = torch.sum(candidate_output, dim=1)
+        # import pdb;pdb.set_trace()
+        return output # torch.Size([bz*num_beams, 32, 768])
 
     def forward_pre_route(self, x, beam_scores, expert_route, use_log=True):
         
-        current_scores = self.forward_gate(x) # [bz*num_beams, 5]
+        current_scores = self.forward_gate(x) # [bz, num_beams] / [bz*num_beams, num_beams]
+
+        importance_loss = self._importance_auxiliary_loss(current_scores)
 
         if use_log:
             current_scores_log = torch.log(current_scores) # 取log之后可以直接相加
         else:
             current_scores_log = current_scores
-
+        # import pdb;pdb.set_trace()
         batch_size, num_tokens = x.shape[0], x.shape[1]
         beam_scores, expert_route, beam_idx = self.beam_search(current_scores_log, beam_scores, expert_route, batch_size)
-
         current_expert_select = expert_route[:,-1]
 
         if self.layer_judge=='first': # expand first dim to batch_size * num_beams
             replicated_tensor = x.unsqueeze(1).expand(batch_size, self.num_beams, num_tokens, self.hidden_size)
             x = replicated_tensor.contiguous().view(-1, num_tokens, self.hidden_size) # [bz*num_beams, 32,768]
+            current_scores_t = current_scores.unsqueeze(1).expand(batch_size, self.num_beams, self.num_experts)
+            current_scores = current_scores_t.contiguous().view(-1, self.num_experts) # [bz*num_beams, num_experts]
 
-        candidate_output = self.forward_expert_ffn(x, current_expert_select, beam_scores) # [bz*num_beams, 32,768]
-        
-        return candidate_output, beam_scores, expert_route, beam_idx
+        input_x = x[beam_idx]
+        candidate_output = self.forward_expert_ffn(input_x, current_expert_select, current_scores) # [bz*num_beams, 32,768]
+        # import pdb;pdb.set_trace()
+        return candidate_output, beam_scores, expert_route, beam_idx, importance_loss
 
 
     def forward_post_route(self, x, beam_scores, expert_route, use_log=True):
         
         attention_mask = torch.ones(x.shape[0], x.shape[1]).to(x.device)
         x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
-        
+
         def forward_expert(input_x, expert_idx):
             output_x = self.experts[expert_idx].forward(input_x)
             return output_x
@@ -145,12 +167,14 @@ class RouteMoELayer(nn.Module):
         logits_gate_lst = list()
         for expert_idx in range(self.num_experts):
             output_x = forward_expert(x_masked, expert_idx)
-            outputs.append(output_x.unsqueeze(0))
-            output_x_aver = output_x.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz*num_beam, 768])
+            # output_x_aver = output_x.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz*num_beam, 768])
+            output_x_aver = torch.mean(output_x, dim=1)
             # gate_score = self.gates[expert_idx](output_x_aver)
             gate_score = self.gate(output_x_aver)
             logits_gate_lst.append(gate_score)
-        candidate_output = torch.cat(outputs) # torch.Size([num_expert, bz*num_beam, 32, 768])
+            outputs.append(output_x.unsqueeze(0))
+
+        candidate_output_raw = torch.cat(outputs) # torch.Size([num_expert, bz*num_beam, 32, 768])
         logits_gate = torch.cat(logits_gate_lst,dim=1)# torch.Size([bz*num_beam, num_expert])
         current_scores = F.softmax(logits_gate, dim=-1) # torch.Size([bz*num_beam, num_experts])
 
@@ -159,24 +183,33 @@ class RouteMoELayer(nn.Module):
         else:
             current_scores_log = current_scores
         
-        batch_size = x.shape[0] # bz*num_beam
+        # importance loss
+        importance_loss = self._importance_auxiliary_loss(current_scores)
+        
+        batch_size, num_tokens = x.shape[0], x.shape[1] # bz*num_beam
         beam_scores, expert_route, beam_idx = self.beam_search(current_scores_log, beam_scores, expert_route, batch_size)
         # beam_scores torch.Size([bz*num_beam])
         # expert_route torch.Size([bz*num_beam, layer_n])
         current_select_expert = expert_route[:,-1]
+        # current_select_expert torch.Size([bz*num_beam, 1])
         
-        output = list()
-        for i in range(beam_idx.shape[0]):
-            b_idx = beam_idx[i]
-            ex_idx = current_select_expert[i]
-            ex_out = candidate_output[ex_idx, b_idx, :,:]
-            output.append(ex_out.unsqueeze(0))
-
-        final_output = torch.concat(output, dim=0)
-
-        return final_output, beam_scores, expert_route, beam_idx
-
-
+        if self.layer_judge == 'first':
+            replicated_tensor = candidate_output_raw.unsqueeze(2).expand(self.num_experts, batch_size, self.num_beams, num_tokens, self.hidden_size)
+            candidate_output_raw = replicated_tensor.contiguous().view(self.num_experts, -1, num_tokens, self.hidden_size) # [bz*num_beams, 32,768]
+            current_scores_t = current_scores.unsqueeze(1).expand(batch_size, self.num_beams, self.num_experts)
+            current_scores = current_scores_t.contiguous().view(-1, self.num_experts) # [bz*num_beams, num_experts]
+        
+        candidate_output = candidate_output_raw.permute(1, 0, 2, 3)[beam_idx] # torch.Size([8, 2, 32, 768])
+        expert_select_matrix = F.one_hot(current_select_expert, self.num_experts)
+        if self.weight_type == 'ffn_prob':
+            tmp_prob = current_scores[beam_idx] * expert_select_matrix
+            output = candidate_output * tmp_prob.unsqueeze(-1).unsqueeze(-1)
+        else:
+            output = candidate_output * expert_select_matrix.unsqueeze(-1).unsqueeze(-1)
+        final_output = torch.sum(output, dim=1)
+        
+        return final_output, beam_scores, expert_route, beam_idx, importance_loss
+    
     def forward(self, x, attention_mask, beam_scores, expert_route, use_log=True):
         """
             if first_layer: x [bz, 32, 768]
@@ -184,11 +217,11 @@ class RouteMoELayer(nn.Module):
         
         """
         if self.route_method == 'pre-route':
-            candidate_output, beam_scores, expert_route, beam_idx = self.forward_pre_route(x, beam_scores, expert_route, use_log=True)
+            candidate_output, beam_scores, expert_route, beam_idx, importance_loss = self.forward_pre_route(x, beam_scores, expert_route, use_log=True)
         elif self.route_method == "post-route":
-            candidate_output, beam_scores, expert_route, beam_idx = self.forward_post_route(x, beam_scores, expert_route, use_log=True)
+            candidate_output, beam_scores, expert_route, beam_idx, importance_loss = self.forward_post_route(x, beam_scores, expert_route, use_log=True)
 
-        return candidate_output, beam_scores, expert_route, beam_idx
+        return candidate_output, beam_scores, expert_route, beam_idx, importance_loss
 
 
 

minigpt4/models/moe/test_moe_layer.py

@@ -0,0 +1,294 @@
+import copy
+import pickle
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+import copy
+import pickle
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class MoELayer(nn.Module):
+    def __init__(self, hidden_size, expert, num_experts, route_method, topk=1, use_balance_loss=True, weight_type='raw_prob, topk(softmax)'):
+        # remove hash list
+        nn.Module.__init__(self)
+        self.num_experts = num_experts
+        self.experts = nn.ModuleList([copy.deepcopy(expert) for i in range(num_experts)])
+        self.route_method = route_method
+        self.topk = topk
+        self.use_balance_loss = use_balance_loss
+        self.weight_type = weight_type
+
+        if route_method in ["gate-token", "gate-sentence"]:
+            self.gate = nn.Linear(hidden_size, num_experts, bias=False).float()
+        elif route_method in ["gate-sentence-post"]:
+            gate = nn.Linear(hidden_size, 1, bias=False).float()
+            # self.gates = nn.ModuleList([copy.deepcopy(gate) for i in range(num_experts)])
+            self.gate = gate
+        else:
+            raise KeyError("Routing method not supported.")
+
+    def _balancing_loss(self, prob_gate, num_tokens):
+        # From MOEBERT
+        # compute the load balancing loss
+        # prob_gate，是 [bz, num_expert]，每个样本被分配给每个expert的概率
+        # 等价于 VMOE 中 _gshard_auxiliary_loss
+        P = prob_gate.mean(0) # torch.Size([num_expert]) 每个expert被分配到样本的平均概率
+        temp = num_tokens.float()
+        f = temp / temp.sum(0, keepdim=True) # 每个expert被分配的sample比例
+        balance_loss = self.num_experts * torch.sum(P * f) 
+        return balance_loss
+
+    def _importance_auxiliary_loss(self, prob_gate):
+        # From VMOE
+        # _importance_auxiliary_loss
+        axis = tuple(range(prob_gate.ndim - 1))  # All except last.
+        importance_per_expert = torch.sum(prob_gate, dim=axis)
+        std_importance_per_expert = torch.std(importance_per_expert)
+        mean_importance_per_expert = torch.mean(importance_per_expert)
+        # Compute coefficient of variation (i.e. std/mean) squared.
+        return (std_importance_per_expert / mean_importance_per_expert)**2
+
+    def _forward_gate_token(self, x):
+        bsz, seq_len, dim = x.size()
+
+        x = x.view(-1, dim)
+        logits_gate = self.gate(x)
+        prob_gate = F.softmax(logits_gate, dim=-1)
+        gate = torch.argmax(prob_gate, dim=-1)
+
+        order = gate.argsort(0)
+        num_tokens = F.one_hot(gate, self.num_experts).gt(0).sum(0)
+        gate_load = num_tokens.clone()
+        x = x[order]  # reorder according to expert number
+        x = x.split(num_tokens.tolist(), dim=0)  # a list of length self.num_experts
+
+        # compute the load balancing loss
+        P = prob_gate.mean(0)
+        temp = num_tokens.float()
+        f = temp / temp.sum(0, keepdim=True)
+        balance_loss = self.num_experts * torch.sum(P * f)
+
+        prob_gate = prob_gate.gather(dim=1, index=gate.unsqueeze(1))
+        prob_gate = prob_gate[order]
+        prob_gate = prob_gate.split(num_tokens.tolist(), dim=0)
+
+        def forward_expert(input_x, prob_x, expert_idx):
+            input_x = self.experts[expert_idx].forward(input_x)
+            input_x = input_x * prob_x
+            return input_x
+
+        x = [forward_expert(x[i], prob_gate[i], i) for i in range(self.num_experts)]
+        x = torch.vstack(x)
+        x = x[order.argsort(0)]  # restore original order
+        x = x.view(bsz, seq_len, dim)
+
+        return x, balance_loss, gate_load
+
+    def _forward_gate_sentence_post(self, x, attention_mask):
+        """
+            x: query_attention_output; torch.Size([bz, 32, 768])
+            attention_mask: torch.ones([bz, 32])
+            bz = 4
+            x = torch.randn(bz,32,768)
+            attention_mask = torch.ones([bz, 32])
+
+        """
+        # Prepare Input x
+        attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
+        x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
+        
+        # FeedForward(x) & Forward Gate
+        outputs = list()
+        logits_gate_lst = list()
+        for expert_idx in range(self.num_experts):
+            output_x = self.experts[expert_idx].forward(x_masked)
+            outputs.append(output_x.unsqueeze(0))
+            
+            output_x_aver = torch.mean(output_x, dim=1)
+            # gate_acore = self.gates[expert_idx](output_x_aver)
+            gate_score = self.gate(output_x_aver)
+            logits_gate_lst.append(gate_score)
+        candidate_output = torch.cat(outputs) # torch.Size([num_expert, bz, 32, 768])
+        logits_gate = torch.cat(logits_gate_lst,dim=1)# torch.Size([bz, num_expert])
+
+        # Probabilities for each sample of what expert it should be sent to.
+        prob_gate = F.softmax(logits_gate, dim=-1) # torch.Size([bz, num_experts])
+        if 'softmax(topk)' in self.weight_type:
+            prob_gate1, gate = torch.topk(logits_gate, self.topk, dim=1)
+            select_prob_gate = F.softmax(prob_gate1, dim=-1)
+        else:
+            select_prob_gate, gate = torch.topk(prob_gate, self.topk, dim=1) # gate, 每个样本被分配的expert: torch.Size([bz, topk])
+
+        # Calculate Balancing Loss
+        if self.use_balance_loss:
+            num_sentences = F.one_hot(gate, self.num_experts).sum(1).gt(0).sum(0) # 每个expert被分配的样本数 torch.Size([num_expert])
+            balance_loss = self._balancing_loss(prob_gate, num_sentences)
+        else:
+            balance_loss = 0.0
+        # Calculate Importance Loss
+        importance_loss = self._importance_auxiliary_loss(prob_gate)
+
+        # Reshap Prob_gate & Gate
+        # expert_mask: [batch_size, topk, num_experts]
+        # expert_gate: [batch_size, topk, num_experts]
+        # combine_tensor: [batch_size, num_experts]
+        expert_mask = F.one_hot(gate, self.num_experts) 
+        expert_gate = select_prob_gate.unsqueeze(-1) * expert_mask
+        combine_tensor = torch.sum(expert_gate, dim=1) 
+        # combine_tensor = torch.zeros_like(prob_gate)
+        # combine_tensor.scatter_(1, gate, select_prob_gate) # 等价操作，但可能不可导
+
+        candidate_output_ad = torch.permute(candidate_output, (1, 0, 2, 3)) # torch.Size([bz, num_expert, 32, 768])
+        results = candidate_output_ad * combine_tensor.unsqueeze(-1).unsqueeze(-1) # torch.Size([bz, num_expert, 32, 768])
+        outputs = torch.sum(results, dim=1) # torch.Size([bz, 32, 768])
+        import pdb;pdb.set_trace()
+
+        return outputs, (balance_loss+importance_loss), combine_tensor
+
+    def pre_router(self, x, attention_mask):
+        # Prepare input x
+        attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
+        x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
+        x_average = torch.mean(x_masked, dim=1) # torch.Size([bz, 768])
+        
+        # Forward Gate
+        # logits_gate: [bz, num_experts]
+        logits_gate = self.gate(x_average)
+
+        # Probabilities for each sample of what expert it should be sent to.
+        # prob_gate: [bz, num_experts]
+        prob_gate = F.softmax(logits_gate, dim=-1)
+        
+        if 'softmax(topk)' in self.weight_type:
+            prob_gate1, gate = torch.topk(logits_gate, self.topk, dim=1)
+            select_prob_gate = F.softmax(prob_gate1, dim=-1)
+        else:
+            # topk(softmax)
+            # Get Top-K experts for each sample
+            # gate: [bz, topk]
+            # select_prob_gate: [bz, topk]
+            select_prob_gate, gate = torch.topk(prob_gate, self.topk, dim=1)
+
+        # Reshap Prob_gate & Gate
+        # expert_mask: [batch_size, topk, num_experts]
+        # expert_gate: [batch_size, topk, num_experts]
+        # combine_tensor: [batch_size, num_experts]
+        expert_mask = F.one_hot(gate, self.num_experts) 
+        expert_gate = select_prob_gate.unsqueeze(-1) * expert_mask
+        combine_tensor = torch.sum(expert_gate, dim=1) 
+
+        # Calculate Balancing Loss
+        if self.use_balance_loss:
+            num_sentences = F.one_hot(gate, self.num_experts).sum(1).gt(0).sum(0) # 每个expert被分配的样本数 torch.Size([num_expert])
+            balance_loss = self._balancing_loss(prob_gate, num_sentences)
+        else:
+            balance_loss = 0.0
+
+        # Calculate Importance Loss
+        importance_loss = self._importance_auxiliary_loss(prob_gate)
+
+        import pdb; pdb.set_trace()
+
+        return expert_mask, combine_tensor, balance_loss, importance_loss
+
+    def _forward_gate_sentence(self, x, attention_mask):
+        """
+            x: query_attention_output , torch.Size([bz, 32, 768])
+            attention_mask: torch.ones([bz, 32])
+            
+            ### Notice:
+            the raw version of expert_attention_mask is the extended_attention_mask, 
+            which will be add to attention_score directly
+            the values of extended_attention_mask are -0.0 or -10000
+            it should be adjust to 1/0 version to be processed by experts
+        """
+        # Forward Router
+        expert_mask, combine_tensor, balance_loss, importance_loss = self.pre_router(x, attention_mask)
+        
+        # Forward Expert FFN
+        result = []
+        for expert_idx in range(self.num_experts):
+            output_x = self.experts[expert_idx].forward(x)
+            result.append(output_x.unsqueeze(0))
+        expert_output = torch.cat(result).permute(1,0,2,3) # torch.Size([batch_size, num_expert, num_tokens, hidden_states])
+
+        # multiply outputs of experts by the routing probability
+        expert_outputs_combined = expert_output * combine_tensor.unsqueeze(-1).unsqueeze(-1) # torch.Size([batch_size, num_expert, num_tokens, hidden_states])
+        outputs = torch.sum(expert_outputs_combined, dim=1) # torch.Size([batch_size, num_tokens, hidden_states])
+
+        import pdb; pdb.set_trace()
+
+        return outputs, (balance_loss+importance_loss), combine_tensor
+
+
+    def forward(self, x, attention_mask):
+        if self.route_method == "gate-token":
+            x, balance_loss, gate_load = self._forward_gate_token(x)
+        elif self.route_method == "gate-sentence":
+            x, balance_loss, gate_load = self._forward_gate_sentence(x, attention_mask)
+        elif self.route_method == "gate-sentence-post":
+            x, balance_loss, gate_load = self._forward_gate_sentence_post(x, attention_mask)
+        else:
+            raise KeyError("Routing method not supported.")
+        # import pdb; pdb.set_trace()
+        return x, balance_loss, gate_load
+
+if __name__ == '__main__':
+
+    import sys
+    sys.path.append("/mnt/pfs-guan-ssai/nlu/wanghanzi/multimodal/PromptMoE")
+    from minigpt4.models.QformerRouteMoE import BertConfig
+    from minigpt4.models.QformerRouteMoE import FeedForward
+    from minigpt4.models.moe.utils import (
+        moe_layer_judge,
+    )
+
+    vision_width = 1408
+    cross_attention_freq = 2
+    num_query_token = 32
+    # init_QformerMoE
+    config = BertConfig.from_pretrained("/mnt/pfs-guan-ssai/nlu/wanghanzi/models/bert-base-uncased")
+    config.encoder_width = vision_width
+    # insert cross-attention layer every other block
+    config.add_cross_attention = True
+    config.cross_attention_freq = cross_attention_freq
+    config.query_length = num_query_token
+    config.moebert_expert_num = 3
+    config.moebert_num_beams = 2
+    config.moebert_route_method = 'gate-sentence-post'
+    config.moe_topk = 1
+    config.use_balance_loss = False
+    # config.moe_weight_type = 'raw_prob, softmax(topk)'
+    config.moe_weight_type = 'raw_prob, topk(softmax)'
+
+    batch_size = 4
+    x2 = torch.randn(batch_size, 32, 768)
+    beam_scores, expert_route = None, None
+
+    for layer_num in [6, 8, 10]:
+        layer_judge = moe_layer_judge(layer_num)
+        ffn = FeedForward(config)
+        gate = nn.Linear(768, config.moebert_expert_num, bias=False).float()
+
+        experts_moe = MoELayer(
+                hidden_size=config.hidden_size,
+                expert=ffn,
+                num_experts=config.moebert_expert_num,
+                route_method=config.moebert_route_method,
+                topk=config.moe_topk,
+                use_balance_loss=config.use_balance_loss,
+                weight_type=config.moe_weight_type,
+            )
+        attn_mask = torch.ones([batch_size, 32])
+        layer_output = experts_moe(x2, attn_mask)
+        hidden_states3, aux_loss, combine_tensor = layer_output
+        
+        print(combine_tensor)
+        print(aux_loss)
+        x2 = hidden_states3
+
+        print("------------------------------------")
+        import pdb; pdb.set_trace()

minigpt4/models/moe/utils.py

@@ -19,15 +19,33 @@ def use_experts(layer_idx):
     else:
         return False
 
+def use_experts_route(layer_idx):
+    # if layer_idx % 2 == 0:
+    # use moe_ffn after cross_attns
+    # if int(layer_idx) in [0,2,4,6,8,10]:
+    if int(layer_idx) in [6,7,8,9,10,11]:
+        return True
+    else:
+        return False
+
 def moe_layer_judge(layer_idx):
     if layer_idx == 6:
         return 'first'
-    elif layer_idx == 8:
+    elif layer_idx in [7,8,9,10]:
         return 'mid'
-    elif layer_idx == 10:
+    elif layer_idx == 11:
         return 'last'
     else:
         return None
+    
+    # if layer_idx == 0:
+    #     return 'first'
+    # elif layer_idx in [2,4,6,8]:
+    #     return 'mid'
+    # elif layer_idx == 10:
+    #     return 'last'
+    # else:
+    #     return None
 
 def process_ffn(model):
     if model.config.model_type == "bert":

minigpt4/projects/qformer_moe_post_vicuna/train/mix_qformer_moe_post_blip2_vicuna7b_data_balance.yaml

@@ -10,7 +10,6 @@ model:
   load_finetuned: False
   vit_model: eva_clip_g
   pretrained: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/blip2/blip2_vicuna7b/blip2_pretrained_vicuna7b.pth"
-  # finetuned: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe/mix_coco_gqa_balance_raw_QformerMoE_Post_train_qf_train_qt_aver_weight_5ex_top1_1loss_textinqf_epo3_s42_1201/20231201184/checkpoint_best.pth"
   finetuned: ""
   q_former_model: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/blip2/blip2_vicuna7b/blip2_pretrained_vicuna7b.pth"
 
@@ -38,7 +37,7 @@ model:
 
   # moe
   use_moeqformer: True
-  moebert_expert_num: 5
+  moebert_expert_num: 3
   moebert_route_method: "gate-sentence-post"
   moebert_load_balance: 0
   moe_topk: 1
@@ -110,6 +109,7 @@ run:
   max_epoch: 1
   num_workers: 4
   warmup_steps: 600
+  iters_per_epoch: 1000
 
   seed: 42
   output_dir: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe/mix_coco_gqa_balance_raw_QformerMoE_Post_train_qf_train_qt_aver_weight_5ex_top1_1loss_textinqf_epo3_s42_1201/"

minigpt4/projects/qformer_moe_route_vicuna/eval/mix_vqa_coco_vicuna_eval.yaml

@@ -10,7 +10,7 @@ model:
   load_finetuned: True
   vit_model: eva_clip_g
   pretrained: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/blip2/blip2_vicuna7b/blip2_pretrained_vicuna7b.pth"
-  finetuned: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe_route/mix_coco_gqa_1610k_raw_QformerMoE_Route_Post_linear_gate_3ex_3beam_1loss_top3layer_log_textinqf_epo3_1216/20231216155/checkpoint_best.pth"
+  finetuned: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe_route/mix_coco_gqa_1610k_raw_QformerMoE_Route_Post_ffn_prob_linear_gate_2ex_2beam_1gate_2loss_5e5lr_top6layer_textinqf_epo8_0112/20240112212/checkpoint_best.pth"
   q_former_model: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/blip2/blip2_vicuna7b/blip2_pretrained_vicuna7b.pth"
 
   # vit encoder
@@ -38,10 +38,12 @@ model:
   # moe
   use_moeqformer: True
   use_route_moe: True
-  moebert_expert_num: 3
-  moebert_num_beams: 3
   moebert_route_method: "post-route"
-  gate_save_path: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe_route/route_save/mix_coco_gqa_balance_raw_QformerMoE_Route_linear_gate_5ex_2beam_1loss_textinqf_epo5_toplayer3_1209_eval_latest1/"
+  moebert_load_balance: 0
+  moebert_expert_num: 2
+  moebert_num_beams: 2
+  moe_weight_type: 'ffn_prob'
+  gate_save_path: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe_route/route_save/mix_coco_gqa_1610k_raw_QformerMoE_Route_Post_ffn_prob_linear_gate_2ex_2beam_1gate_2loss_5e5lr_top6layer_textinqf_epo8_0112/"
 
 datasets:
   gqa:
@@ -81,19 +83,20 @@ run:
   task: instruction_tuning
   # optimizer
   lr_sched: "linear_warmup_cosine_lr"
-  init_lr: 2e-5
+  init_lr: 5e-5
   min_lr: 1e-6
   warmup_lr: 1e-6
   log_freq: 5
   save_freq: 1500
 
   weight_decay: 0.05
-  max_epoch: 5
+  max_epoch: 10
   num_workers: 4
   warmup_steps: 600
+  iters_per_epoch: 3000
 
   seed: 42
-  output_dir: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe_route/eval/mix_coco_gqa_balance_raw_QformerMoE_Route_linear_gate_5ex_2beam_1loss_textinqf_epo5_toplayer3_1209/"
+  output_dir: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe_route/eval/mix_coco_gqa_1610k_raw_QformerMoE_Route_Post_ffn_prob_linear_gate_2ex_2beam_1gate_2loss_5e5lr_top6layer_textinqf_epo8_0112/"
   
   amp: True
   resume_ckpt_path: null

minigpt4/projects/qformer_moe_route_vicuna/train/mix_qformer_moe_route_blip2_vicuna7b_data_balance.yaml

@@ -38,10 +38,12 @@ model:
   # moe
   use_moeqformer: True
   use_route_moe: True
+  moebert_route_method: "post-route"
+  moebert_load_balance: 0
   moebert_expert_num: 3
   moebert_num_beams: 3
-  moebert_route_method: "post-route"
-  gate_save_path: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe/route_save/mix_coco_gqa_balance_raw_QformerMoE_Route_linear_gate_5ex_2beam_1loss_textinqf_epo5_toplayer3_1209/"
+  moe_weight_type: 'ffn_prob'
+  use_balance_loss: False
 
 datasets:
   gqa: # train: 943000, 12578, 12578)
@@ -97,19 +99,20 @@ run:
   task: instruction_tuning
   # optimizer
   lr_sched: "linear_warmup_cosine_lr"
-  init_lr: 2e-5
+  init_lr: 5e-5
   min_lr: 1e-6
   warmup_lr: 1e-6
   log_freq: 5
   save_freq: 1500
 
   weight_decay: 0.05
-  max_epoch: 5
+  max_epoch: 8
   num_workers: 4
   warmup_steps: 600
+  iters_per_epoch: 5000
 
   seed: 42
-  output_dir: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe_route/mix_coco_gqa_balance_raw_QformerMoE_Route_linear_gate_5ex_2beam_1loss_textinqf_epo5_toplayer3_1209/"
+  output_dir: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe_route/mix_coco_gqa_1610k_raw_QformerMoE_Route_Post_ffn_prob_linear_1gate_3ex_3beam_1loss_5e5lr_top6layer_textinqf_epo8_0117/"
   
   amp: True
   resume_ckpt_path: null

minigpt4/projects/qformer_moe_route_vicuna/train/mix_qformer_moe_route_blip2_vicuna7b_data_balance_1220.yaml

@@ -0,0 +1,129 @@
+ # Copyright (c) 2022, salesforce.com, inc.
+ # All rights reserved.
+ # SPDX-License-Identifier: BSD-3-Clause
+ # For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+# 0107test
+
+model:
+  arch: blip2_vicuna_instruct
+  model_type: vicuna7b_pretrain
+  load_pretrained: True
+  load_finetuned: False
+  vit_model: eva_clip_g
+  pretrained: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/blip2/blip2_vicuna7b/blip2_pretrained_vicuna7b.pth"
+  # finetuned: ""
+  q_former_model: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/blip2/blip2_vicuna7b/blip2_pretrained_vicuna7b.pth"
+
+  # vit encoder
+  image_size: 224
+  drop_path_rate: 0
+  use_grad_checkpoint: False
+  vit_precision: "fp16"
+
+  # Q-Former
+  num_query_token: 32
+  qformer_text_input: True
+
+  # vicuna
+  llm_model: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/vicuna-7b-v1.1"
+  prompt: ""
+  max_txt_len: 256
+  max_output_txt_len: 256
+
+  # freeze
+  freeze_vit: True
+  freeze_llm: True
+  freeze_qformer: False
+  freeze_t5_proj: False
+
+  # moe
+  use_moeqformer: True
+  use_route_moe: True
+  moebert_route_method: "post-route"
+  moebert_load_balance: 0
+  moebert_expert_num: 2
+  moebert_num_beams: 2
+  moe_weight_type: 'ffn_prob'
+  use_balance_loss: False
+  # gate_save_path: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe/route_save/mix_coco_gqa_balance_raw_QformerMoE_Route_linear_gate_5ex_2beam_1loss_textinqf_epo5_toplayer3_1209/"
+
+datasets:
+  # gqa: # train: 943000, 12578, 12578)
+  #   type: balanced_sft_raw
+  #   batch_size: 1
+  #   vis_processor:
+  #     train:
+  #       name: "blip2_image_train"
+  #       image_size: 224
+  #     eval:
+  #       name: "blip2_image_eval"
+  #       image_size: 224
+  #   text_processor:
+  #     train:
+  #       name: "blip_caption"
+  #     eval:
+  #       name: "blip_caption"
+  #   sample_ratio: 10
+  
+  ok_vqa: # train, valid (9009, 5046)
+    batch_size: 1
+    vis_processor:
+      train:
+        name: "blip2_image_train"
+        image_size: 224
+      eval:
+        name: "blip2_image_eval"
+        image_size: 224
+    text_processor:
+      train:
+        name: "blip_caption"
+      eval:
+        name: "blip_caption"
+    sample_ratio: 1
+  
+  # coco_vqa:    # 658104
+  #   batch_size: 1
+  #   vis_processor:
+  #     train:
+  #       name: "blip2_image_train"
+  #       image_size: 224
+  #     eval:
+  #       name: "blip2_image_eval"
+  #       image_size: 224
+  #   text_processor:
+  #     train:
+  #       name: "blip_caption"
+  #     eval:
+  #       name: "blip_caption"
+  #   sample_ratio: 9
+
+run:
+  task: instruction_tuning
+  # optimizer
+  lr_sched: "linear_warmup_cosine_lr"
+  init_lr: 2e-5
+  min_lr: 1e-6
+  warmup_lr: 1e-6
+  log_freq: 5
+  save_freq: 1500
+
+  weight_decay: 0.05
+  max_epoch: 5
+  num_workers: 4
+  warmup_steps: 600
+
+  seed: 42
+  output_dir: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe_route/mix_coco_gqa_balance_raw_QformerMoE_Route_linear_gate_5ex_2beam_1loss_textinqf_epo5_toplayer3_1209/"
+  
+  amp: True
+  resume_ckpt_path: null
+
+  evaluate: False 
+  train_splits: ["train"]
+  valid_splits: ["val"]
+  # test_splits: ["val"]
+
+  device: "cuda"
+  world_size: 1
+  dist_url: "env://"
+  distributed: True

minigpt4/projects/qformer_moe_vicuna/eval/vqa_benchmark_evaluation.yaml

@@ -10,7 +10,7 @@ model:
   load_finetuned: True
   vit_model: eva_clip_g
   pretrained: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/blip2/blip2_vicuna7b/blip2_pretrained_vicuna7b.pth"
-  finetuned: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe_route/mix_1048k_raw_QformerMoE_Route_Post_NoNorm_5ex_2beam_1loss_top3layer_textinqf_epo6_1215/20231216161/checkpoint_best.pth"
+  finetuned: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe_route/mix_coco_gqa_1610k_raw_QformerMoE_Route_Post_ffn_prob_linear_gate_2ex_2beam_1gate_1loss_5e5lr_top6layer_textinqf_epo8_0111/20240111145/checkpoint_best.pth"
   q_former_model: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/blip2/blip2_vicuna7b/blip2_pretrained_vicuna7b.pth"
 
   # vit encoder
@@ -39,8 +39,11 @@ model:
   use_moeqformer: True
   use_route_moe: True
   moebert_route_method: "post-route"
-  moebert_expert_num: 5
+  moebert_load_balance: 0
+  moebert_expert_num: 2
   moebert_num_beams: 2
+  moe_weight_type: 'ffn_prob'
+  use_balance_loss: False
 
 datasets:
   ok_vqa: # train, valid (9009, 5046)
@@ -78,7 +81,7 @@ evaluation_datasets:
 run:
   task: instruction_tuning
   name: vqa_benchmark_evaluation
-  save_path: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe/eval/benchmarks/mix_1048k_raw_QformerMoE_Route_Post_NoNorm_5ex_2beam_1loss_top3layer_textinqf_epo6_1215/"
+  save_path: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe_route/eval/benchmarks/mix_coco_gqa_1610k_raw_QformerMoE_Route_Post_ffn_prob_linear_gate_2ex_2beam_1gate_1loss_5e5lr_top6layer_textinqf_epo8_0111/"
   seed: 42
 
   

minigpt4/projects/qformer_moe_vicuna/train/mix_qformer_moe_blip2_vicuna7b_data_3ex3beam_0112.yaml

@@ -0,0 +1,131 @@
+ # Copyright (c) 2022, salesforce.com, inc.
+ # All rights reserved.
+ # SPDX-License-Identifier: BSD-3-Clause
+ # For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+
+model:
+  arch: blip2_vicuna_instruct
+  model_type: vicuna7b_pretrain
+  load_pretrained: True
+  load_finetuned: False
+  vit_model: eva_clip_g
+  pretrained: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/blip2/blip2_vicuna7b/blip2_pretrained_vicuna7b.pth"
+  # finetuned: ""
+  q_former_model: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/blip2/blip2_vicuna7b/blip2_pretrained_vicuna7b.pth"
+
+  # vit encoder
+  image_size: 224
+  drop_path_rate: 0
+  use_grad_checkpoint: False
+  vit_precision: "fp16"
+
+  # Q-Former
+  num_query_token: 32
+  qformer_text_input: True
+
+  # vicuna7b
+  llm_model: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/vicuna-7b-v1.1"
+  prompt: ""
+  max_txt_len: 256
+  max_output_txt_len: 256
+
+  # freeze
+  freeze_vit: True
+  freeze_llm: True
+  freeze_qformer: False
+  freeze_t5_proj: False
+
+  # moe
+  use_moeqformer: True
+  use_route_moe: True
+  moebert_route_method: "post-route"
+  moebert_load_balance: 0.05
+  moebert_expert_num: 3
+  moebert_num_beams: 3
+  moe_weight_type: 'ffn_prob'
+  use_balance_loss: False
+
+datasets:
+  gqa: # train: 943000, 12578, 12578)
+    type: balanced_sft_raw
+    # batch_size: 16
+    batch_size: 32
+    vis_processor:
+      train:
+        name: "blip2_image_train"
+        image_size: 224
+      eval:
+        name: "blip2_image_eval"
+        image_size: 224
+    text_processor:
+      train:
+        name: "blip_caption"
+      eval:
+        name: "blip_caption"
+    sample_ratio: 50
+  
+  ok_vqa: # train, valid (9009, 5046)
+    # batch_size: 16
+    batch_size: 32
+    vis_processor:
+      train:
+        name: "blip2_image_train"
+        image_size: 224
+      eval:
+        name: "blip2_image_eval"
+        image_size: 224
+    text_processor:
+      train:
+        name: "blip_caption"
+      eval:
+        name: "blip_caption"
+    sample_ratio: 8  
+  
+  coco_vqa:    # 658104
+    # batch_size: 16
+    batch_size: 32
+    vis_processor:
+      train:
+        name: "blip2_image_train"
+        image_size: 224
+      eval:
+        name: "blip2_image_eval"
+        image_size: 224
+    text_processor:
+      train:
+        name: "blip_caption"
+      eval:
+        name: "blip_caption"
+    sample_ratio: 15
+
+run:
+  task: instruction_tuning
+  # optimizer
+  lr_sched: "linear_warmup_cosine_lr"
+  # init_lr: 2e-5
+  init_lr: 5e-5
+  min_lr: 1e-6
+  warmup_lr: 1e-6
+  log_freq: 5
+  save_freq: 1500
+
+  weight_decay: 0.05
+  max_epoch: 8
+  num_workers: 4
+  warmup_steps: 600
+  iters_per_epoch: 5000
+
+  seed: 42
+  output_dir: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe_route/mix_coco_gqa_1610k_raw_QformerMoE_Route_Post_ffn_prob_linear_gate_3ex_3beam_1gate_2loss_5e5lr_top6layer_textinqf_epo8_0112/"
+  
+  amp: True
+  resume_ckpt_path: null
+
+  evaluate: False 
+  train_splits: ["train"]
+  valid_splits: ["val"]
+
+  device: "cuda"
+  world_size: 1
+  dist_url: "env://"
+  distributed: True

minigpt4/projects/qformer_moe_vicuna/train/mix_qformer_moe_blip2_vicuna7b_data_balance.yaml

@@ -37,19 +37,19 @@ model:
 
   # moe
   use_moeqformer: True
-  moebert_expert_num: 5
+  moebert_expert_num: 3
   moebert_route_method: "gate-sentence"
   moebert_load_balance: 0
   moe_topk: 1
   use_balance_loss: False
-  moe_weight_type: 'l2_norm'
-  gate_save_path: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe/gate_save/mix_coco_gqa_balance_raw_QformerMoE_train_qf_train_qt_linear_gate_5ex_top1_1loss_textinqf_training_epo5_toplayer3_1206/"
+  moe_weight_type: 'raw_prob'
+  # gate_save_path: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe/gate_save/mix_coco_gqa_balance_raw_QformerMoE_train_qf_train_qt_linear_gate_5ex_top1_1loss_textinqf_training_epo5_toplayer3_1206/"
 
 
 datasets:
   gqa: # train: 94254
     type: balanced_sft_raw_part
-    batch_size: 32
+    batch_size: 1
     vis_processor:
       train:
         name: "blip2_image_train"
@@ -65,7 +65,7 @@ datasets:
     sample_ratio: 50
   
   ok_vqa: # train, valid (9009, 5046
-    batch_size: 32
+    batch_size: 1
     vis_processor:
       train:
         name: "blip2_image_train"
@@ -80,22 +80,22 @@ datasets:
         name: "blip_caption"
     sample_ratio: 8
   
-  coco_vqa:    # 214352 vqa_val
-    type: vqa_v2_part
-    batch_size: 32
-    vis_processor:
-      train:
-        name: "blip2_image_train"
-        image_size: 224
-      eval:
-        name: "blip2_image_eval"
-        image_size: 224
-    text_processor:
-      train:
-        name: "blip_caption"
-      eval:
-        name: "blip_caption"
-    sample_ratio: 15
+  # coco_vqa:    # 214352 vqa_val
+  #   type: vqa_v2_part
+  #   batch_size: 1
+  #   vis_processor:
+  #     train:
+  #       name: "blip2_image_train"
+  #       image_size: 224
+  #     eval:
+  #       name: "blip2_image_eval"
+  #       image_size: 224
+  #   text_processor:
+  #     train:
+  #       name: "blip_caption"
+  #     eval:
+  #       name: "blip_caption"
+  #   sample_ratio: 15
 
 run:
   task: instruction_tuning
@@ -108,12 +108,13 @@ run:
   save_freq: 1500
 
   weight_decay: 0.05
-  max_epoch: 5
+  max_epoch: 1
   num_workers: 4
   warmup_steps: 600
+  iters_per_epoch: 1000
 
   seed: 42
-  output_dir: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe/mix_coco_gqa_balance_raw_QformerMoE_train_qf_train_qt_linear_gate_5ex_top1_1loss_textinqf_training_epo5_toplayer3_1206/"
+  output_dir: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe/mix_coco_gqa_balance_raw_QformerMoE_train_qf_train_qt_linear_gate_5ex_top1_1loss_textinqf_training_epo5_toplayer3_1220_test/"
   
   amp: True
   resume_ckpt_path: null

minigpt4/projects/qformer_moe_vicuna/train/mix_qformer_moe_blip2_vicuna7b_data_raw_0112.yaml

@@ -0,0 +1,125 @@
+ # Copyright (c) 2022, salesforce.com, inc.
+ # All rights reserved.
+ # SPDX-License-Identifier: BSD-3-Clause
+ # For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+
+model:
+  arch: blip2_vicuna_instruct
+  model_type: vicuna7b_pretrain
+  load_pretrained: True
+  load_finetuned: False
+  vit_model: eva_clip_g
+  pretrained: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/blip2/blip2_vicuna7b/blip2_pretrained_vicuna7b.pth"
+  # finetuned: ""
+  q_former_model: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/blip2/blip2_vicuna7b/blip2_pretrained_vicuna7b.pth"
+
+  # vit encoder
+  image_size: 224
+  drop_path_rate: 0
+  use_grad_checkpoint: False
+  vit_precision: "fp16"
+
+  # Q-Former
+  num_query_token: 32
+  qformer_text_input: True
+
+  # vicuna7b
+  llm_model: "/mnt/pfs-guan-ssai/nlu/wanghanzi/models/vicuna-7b-v1.1"
+  prompt: ""
+  max_txt_len: 256
+  max_output_txt_len: 256
+
+  # freeze
+  freeze_vit: True
+  freeze_llm: True
+  freeze_qformer: False
+  freeze_t5_proj: False
+
+  # moe
+  use_moeqformer: False
+  moebert_expert_num: 1
+  moebert_route_method: "gate-sentence"
+  moebert_load_balance: 0.05
+  moe_topk: 1
+
+datasets:
+  gqa: # train: 943000, 12578, 12578)
+    type: balanced_sft_raw
+    batch_size: 16
+    vis_processor:
+      train:
+        name: "blip2_image_train"
+        image_size: 224
+      eval:
+        name: "blip2_image_eval"
+        image_size: 224
+    text_processor:
+      train:
+        name: "blip_caption"
+      eval:
+        name: "blip_caption"
+    sample_ratio: 50
+  
+  ok_vqa: # train, valid (9009, 5046)
+    batch_size: 16
+    vis_processor:
+      train:
+        name: "blip2_image_train"
+        image_size: 224
+      eval:
+        name: "blip2_image_eval"
+        image_size: 224
+    text_processor:
+      train:
+        name: "blip_caption"
+      eval:
+        name: "blip_caption"
+    sample_ratio: 8  
+  
+  coco_vqa:    # 658104
+    batch_size: 16
+    vis_processor:
+      train:
+        name: "blip2_image_train"
+        image_size: 224
+      eval:
+        name: "blip2_image_eval"
+        image_size: 224
+    text_processor:
+      train:
+        name: "blip_caption"
+      eval:
+        name: "blip_caption"
+    sample_ratio: 15
+
+run:
+  task: instruction_tuning
+  # optimizer
+  lr_sched: "linear_warmup_cosine_lr"
+  # init_lr: 2e-5
+  init_lr: 5e-5
+  min_lr: 1e-6
+  warmup_lr: 1e-6
+  log_freq: 5
+  save_freq: 1500
+
+  weight_decay: 0.05
+  max_epoch: 8
+  num_workers: 4
+  warmup_steps: 600
+  iters_per_epoch: 5000
+
+  seed: 42
+  output_dir: "/mnt/pfs-guan-ssai/nlu/wanghanzi/experiments/blip2/vicuna7b/qformer_moe/mix_coco_gqa_1610k_raw_QformerMoE_train_qf_train_qt_1ex_top1_textinqf_epo8_lr5e5_seed42_0112/"
+  
+  amp: True
+  resume_ckpt_path: null
+
+  evaluate: False 
+  train_splits: ["train"]
+  valid_splits: ["val"]
+
+  device: "cuda"
+  world_size: 1
+  dist_url: "env://"
+  distributed: True

minigpt4/runners/runner_base.py

@@ -110,6 +110,7 @@ class RunnerBase:
                 else:
                     p_wd.append(p)
                 num_parameters += p.data.nelement()
+            # import pdb; pdb.set_trace() # 0107test
             logging.info("number of trainable parameters: %d" % num_parameters)
             optim_params = [
                 {

minigpt4/tasks/base_task.py

@@ -238,13 +238,17 @@ class BaseTask:
 
             with torch.cuda.amp.autocast(enabled=use_amp):
                 loss = self.train_step(model=model, samples=samples)
-
+            
             # after_train_step()
             if use_amp:
+                # torch.autograd.set_detect_anomaly(True)
+                # 反向传播时检测是否有异常值，定位code
+                # with torch.autograd.detect_anomaly():
                 scaler.scale(loss).backward()
             else:
                 loss.backward()
 
+            # import pdb; pdb.set_trace() # 0107test
             # update gradients every accum_grad_iters iterations
             if (i + 1) % accum_grad_iters == 0:
                 if use_amp:
@@ -252,6 +256,9 @@ class BaseTask:
                     scaler.update()                     
                 else:    
                     optimizer.step()
+                
+                # import pdb; pdb.set_trace()# 0107test
+
                 optimizer.zero_grad()
                 # if self.cfg.wandb_log:
                 # if self.cfg.run_cfg.wandb_log:

requirements.txt

@@ -44,4 +44,6 @@ wheel
 visualizer
 tensorboard
 kmeans_pytorch
-visual_genome
+visual_genome
+gpustat
+torchviz

setup.py

@@ -0,0 +1,36 @@
+"""
+ Copyright (c) 2022, salesforce.com, inc.
+ All rights reserved.
+ SPDX-License-Identifier: BSD-3-Clause
+ For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+"""
+
+from setuptools import setup, find_namespace_packages
+import platform
+
+DEPENDENCY_LINKS = []
+if platform.system() == "Windows":
+    DEPENDENCY_LINKS.append("https://download.pytorch.org/whl/torch_stable.html")
+
+
+def fetch_requirements(filename):
+    with open(filename) as f:
+        return [ln.strip() for ln in f.read().split("\n")]
+
+
+setup(
+    name="PromptMoE",
+    version="1.0.1",
+    author="Hanzi Wang",
+    description="PromptMoE & QformerMoE Based on LAVIS",
+    long_description=open("README.md", "r", encoding="utf-8").read(),
+    long_description_content_type="text/markdown",
+    keywords="Vision-Language, Multimodal, Image Captioning, Generative AI, Deep Learning, Library, PyTorch",
+    license="3-Clause BSD",
+    packages=find_namespace_packages(include="lavis.*"),
+    install_requires=fetch_requirements("requirements.txt"),
+    python_requires=">=3.7.0",
+    include_package_data=True,
+    dependency_links=DEPENDENCY_LINKS,
+    zip_safe=False,
+)

test.txt

@@ -0,0 +1,360 @@
+tmp_name =  [name for name, p in model.named_parameters() if (p.requires_grad and '10.expert' in name)]
+
+tmp =  [p for name, p in model.named_parameters() if (p.requires_grad and '10.expert' in name)]
+
+tensor([[-1.4032e-02,  3.7242e-03,  8.4997e-03, -3.4016e-03, -6.4855e-03,
+          4.3595e-02,  3.4423e-02, -8.6274e-03, -1.9702e-02,  9.1813e-03,
+          1.1643e-02,  2.3939e-02, -2.0908e-02,  3.4555e-03,  9.1636e-03,
+          1.5413e-02,  2.4148e-02, -1.0880e-03,  1.1193e-02, -1.3591e-02,
+          9.3484e-03,  1.5999e-02, -9.6086e-04,  3.8322e-02, -8.0687e-03,
+         -1.4056e-02,  3.9486e-02,  3.5167e-02, -9.3226e-03, -1.0493e-02,
+         -2.5795e-02, -9.7541e-03,  4.4437e-03,  7.7226e-03,  7.5210e-03,
+         -1.3526e-02, -5.0316e-03, -1.1149e-02,  6.0583e-03,  2.0564e-02,
+         -6.4477e-03,  1.4170e-02, -3.7847e-02,  1.1780e-02,  1.3321e-02,
+         -8.2501e-03, -1.0298e-02,  1.4805e-02, -1.2432e-02, -1.9159e-02,
+         -5.7095e-04, -3.8618e-02, -2.4230e-02, -1.4991e-03, -1.4114e-02,
+         -1.5365e-02,  1.5640e-02, -4.8623e-02, -2.9991e-02,  1.2796e-02,
+         -4.9917e-03,  2.3846e-03,  7.7368e-03,  1.2913e-02,  1.5300e-02,
+          8.5125e-03,  1.1582e-02,  8.1161e-03,  4.2259e-03,  7.6109e-03,
+         -2.0747e-02, -3.5099e-03,  2.2282e-02,  5.0493e-02, -1.7849e-02,
+         -3.7106e-02, -1.4944e-02, -1.4582e-02, -2.2458e-02, -4.6173e-05,
+         -8.1270e-03,  1.9037e-02, -2.0086e-02,  3.0980e-03, -9.3947e-03,
+          1.3054e-02,  2.3203e-02, -9.9304e-03, -2.6038e-02,  1.8679e-02,
+          9.2081e-03, -2.1770e-02, -1.6568e-03, -3.6503e-02,  2.0054e-02,
+          1.2886e-02, -1.8021e-02,  3.4457e-02, -1.3704e-02, -6.1498e-03,
+         -8.6769e-03,  1.5024e-02, -1.3875e-02,  1.7416e-02, -1.1178e-02,
+         -2.4088e-02, -1.7802e-02,  3.3326e-02, -1.1216e-02, -8.6330e-03,
+         -5.5359e-03, -1.1939e-02, -1.7777e-02, -2.8666e-02, -3.8280e-02,
+          4.2682e-02,  1.4946e-02,  9.6427e-03,  8.2754e-03, -1.0516e-03,
+          2.9560e-02,  2.4552e-03, -4.8354e-02,  1.5568e-02,  2.5881e-02,
+         -1.7354e-02, -3.1232e-02,  2.3683e-02, -2.3239e-02,  2.2966e-02,
+          5.6349e-03, -8.7595e-03,  1.5173e-02,  2.7660e-02, -4.3304e-03,
+         -2.5330e-02, -2.1795e-02,  1.6856e-02, -2.1587e-04,  2.3707e-02,
+         -2.3667e-02,  3.5378e-02, -7.9245e-03,  7.1029e-04, -3.2800e-02,
+         -1.5402e-03, -8.5634e-03, -1.1356e-02, -2.1935e-03, -1.8854e-02,
+         -1.9705e-03, -3.8333e-02,  2.9131e-02, -4.4470e-02, -2.0893e-03,
+          1.2937e-02, -1.7116e-02,  2.7778e-02,  1.0311e-02, -6.4017e-03,
+          3.7647e-02, -1.9953e-02, -5.3925e-03,  3.6978e-02, -1.5534e-02,
+          1.2241e-02,  1.3597e-02,  2.0703e-03,  2.4213e-03,  9.2604e-03,
+          6.6108e-03, -5.8213e-03,  9.8167e-03, -9.8300e-04, -1.0236e-02,
+          2.9581e-02,  1.0987e-02,  2.0046e-02, -1.0500e-02, -3.2221e-03,
+         -2.6303e-02,  1.3688e-02, -2.2529e-02, -5.7654e-03,  1.1784e-02,
+          1.6221e-02,  2.8743e-02,  5.7565e-03,  1.8129e-02,  1.5140e-02,
+         -1.1748e-02, -1.7528e-02,  4.7977e-02,  1.5568e-02,  4.7030e-04,
+          3.2757e-03,  1.6631e-02,  1.9986e-02, -7.3463e-03,  1.1435e-02,
+         -1.4739e-02, -3.2959e-03, -2.8770e-03,  2.9260e-02,  1.7007e-02,
+          3.0611e-02,  2.2102e-02, -3.3819e-02, -1.9403e-02,  2.5524e-02,
+          3.0738e-02, -1.9951e-02, -1.4553e-02, -1.5796e-02, -2.3143e-02,
+         -2.8826e-02,  2.4739e-02, -5.8602e-03,  4.1871e-02,  5.0821e-04,
+          3.3493e-02,  2.3524e-02,  2.3191e-02,  9.0416e-03,  3.3262e-02,
+         -1.6805e-02,  1.1545e-02, -1.7195e-02, -3.8696e-02, -8.4358e-04,
+         -8.1605e-03,  3.1372e-03,  1.0726e-03,  1.0865e-03,  1.0760e-02,
+         -5.2421e-03,  1.3039e-02,  3.6873e-04,  1.0464e-02, -1.1544e-02,
+         -2.2775e-02, -4.8439e-02, -1.0711e-02,  4.4236e-03,  2.0351e-02,
+          2.4479e-03, -1.9968e-02, -2.2941e-02, -2.0486e-02, -1.9528e-02,
+         -2.3176e-02, -3.2731e-03,  1.1789e-02,  2.0921e-02,  2.9809e-03,
+         -8.8507e-03, -3.5716e-02,  8.8418e-03,  5.3665e-05, -1.1288e-02,
+         -7.5571e-03,  2.1053e-02, -3.7381e-03, -4.0165e-03, -2.2628e-03,
+          3.7554e-03, -1.6597e-02,  7.6946e-03, -3.2689e-02,  2.2016e-02,
+          5.5122e-03,  4.5455e-02,  6.7586e-03,  1.5714e-02,  5.2125e-03,
+          3.9596e-03,  1.8134e-02,  1.5834e-03, -1.6239e-02, -1.3889e-02,
+         -2.3522e-02,  1.4738e-02,  5.5867e-03, -7.0727e-03, -2.8140e-03,
+          1.6849e-02, -3.1327e-02, -3.2443e-02,  4.7851e-03,  1.2980e-02,
+         -2.0014e-04, -9.9475e-03,  8.0657e-03,  1.9468e-02, -1.5774e-02,
+          1.7017e-02, -8.7196e-03, -4.0681e-03, -6.9754e-03, -2.2007e-02,
+         -6.6217e-03, -1.8219e-02,  4.2186e-02, -5.6621e-03, -9.3449e-03,
+         -1.1662e-02,  2.8700e-02, -9.0654e-03,  3.1569e-02, -2.9825e-03,
+         -3.8198e-02, -5.2723e-02, -4.8325e-02, -2.7871e-03,  5.1127e-03,
+          1.4511e-02,  9.3245e-03, -2.3339e-02, -8.6658e-03,  1.5276e-02,
+         -1.5823e-02, -3.4476e-03,  1.4601e-02,  6.3504e-03, -1.4307e-02,
+          2.2817e-02,  2.1998e-02,  1.7330e-02, -2.4448e-02,  4.0178e-03,
+          3.2280e-03, -1.2721e-02,  1.9661e-02,  7.5263e-03,  2.0245e-02,
+          4.5525e-02, -1.5658e-02, -4.0676e-02,  9.3160e-03,  1.1920e-02,
+         -1.9317e-02,  1.7848e-02, -5.8601e-03,  1.1786e-03,  8.3864e-03,
+         -1.8341e-02,  2.5985e-02, -1.1387e-02, -1.5069e-02, -2.8097e-02,
+          2.4966e-02,  1.4790e-02,  2.0424e-02, -1.3062e-02,  3.1314e-02,
+          1.7811e-02,  7.2393e-03,  1.4413e-02, -1.2746e-02,  3.1039e-02,
+         -1.1697e-02, -1.4826e-02, -8.8397e-03,  1.5157e-02, -1.5855e-02,
+         -1.8157e-03,  1.3024e-02, -1.8902e-03,  2.5212e-02, -3.4886e-02,
+          4.3029e-02, -4.0842e-02,  1.1362e-02, -1.4654e-02, -1.3337e-02,
+         -3.1832e-02,  3.6222e-03,  8.2804e-03, -1.4269e-02,  2.8399e-03,
+         -1.2008e-02,  2.4685e-02, -4.3070e-03,  6.3163e-03, -1.3517e-02,
+         -1.3807e-02,  2.4617e-02,  2.1453e-02,  4.7332e-03,  9.1636e-03,
+         -1.2881e-02,  1.9077e-02,  1.7571e-04, -5.2817e-03, -2.8821e-02,
+          5.8223e-03, -3.0979e-02,  2.4609e-02,  3.6666e-02, -1.0950e-02,
+          2.0421e-02, -2.6378e-03,  3.1825e-02, -9.6689e-04, -2.8398e-02,
+         -2.7513e-02,  1.6946e-02, -2.4110e-02, -1.3575e-02, -1.3443e-02,
+          8.4217e-03,  2.6754e-02, -2.3309e-03, -2.5086e-02,  1.1844e-02,
+          1.4152e-02,  1.2989e-02, -5.7336e-03,  4.7391e-03,  3.4106e-02,
+          1.0142e-02, -1.8029e-02, -1.5410e-04, -1.3548e-02,  9.1742e-03,
+         -3.0150e-02,  1.5666e-02,  4.3049e-03,  1.6273e-02,  2.0672e-02,
+         -1.2458e-02,  4.5496e-02,  3.2131e-02, -3.0967e-03,  2.1891e-02,
+          2.5524e-02, -1.1998e-02, -1.8866e-03, -1.0945e-02,  5.9930e-03,
+         -8.4233e-03, -8.9095e-03, -1.8261e-02,  1.9308e-02, -1.9728e-02,
+         -1.4216e-02,  1.4952e-02,  5.7355e-04, -2.4753e-02, -1.0948e-02,
+          1.0965e-02,  1.3607e-03,  3.4974e-02, -4.1396e-03,  2.5519e-02,
+          1.0364e-02, -1.5851e-02, -4.9224e-03,  1.0903e-02, -1.0523e-04,
+          3.1355e-02, -1.5105e-02,  5.6972e-03, -8.4078e-03, -1.9868e-02,
+          1.7186e-03,  2.9396e-02, -4.1439e-02,  1.4124e-02, -3.7745e-03,
+          3.3007e-02,  8.0368e-04,  8.5574e-03,  1.7269e-02,  1.1955e-02,
+          8.8142e-03, -1.3123e-02,  1.6817e-02, -1.5456e-02, -1.3868e-02,
+          2.4139e-02, -9.1566e-03, -1.8477e-02, -4.7972e-03, -6.8459e-03,
+          1.6818e-02,  3.1645e-03, -3.0901e-02, -5.6036e-03, -1.4758e-02,
+          2.0473e-02, -7.5411e-05,  2.0673e-03, -7.0061e-03,  9.5544e-03,
+          1.6600e-02, -1.7315e-02, -2.0168e-02, -5.3008e-03,  2.0206e-02,
+          2.4209e-03,  2.1205e-02, -8.9188e-03, -4.1350e-04, -1.0638e-02,
+          1.3705e-02,  9.5925e-05,  3.8877e-02,  3.2884e-02, -2.7730e-03,
+          1.0052e-02,  1.9311e-02,  1.1341e-02, -1.2988e-02, -1.7157e-02,
+          3.2095e-02, -1.8493e-02, -9.2551e-03, -2.6509e-03, -1.1130e-02,
+          1.6581e-02,  1.0216e-02,  1.3687e-02,  1.1860e-02, -3.0462e-03,
+         -1.2082e-02,  2.8502e-03, -1.2620e-02,  8.8330e-03,  1.7357e-02,
+          1.8383e-02, -2.3130e-02, -3.2654e-02,  1.2853e-02, -7.8144e-03,
+          1.9418e-04,  3.8635e-03,  4.9333e-02,  1.9350e-02, -2.0643e-02,
+          8.4650e-04,  5.0242e-02,  1.6576e-02, -8.9166e-03, -5.8805e-03,
+         -4.1484e-02,  9.3217e-03, -1.1292e-02, -8.7944e-03, -3.3190e-03,
+          5.7970e-03, -6.6078e-03, -2.4052e-02, -5.6347e-03,  8.4539e-03,
+          1.9250e-02,  7.9559e-03, -3.0055e-03, -3.0398e-04,  2.7007e-02,
+          3.1046e-03,  1.8332e-02,  5.5470e-03,  6.6815e-03,  1.1466e-02,
+          1.9738e-02,  1.2176e-02, -2.0220e-02,  8.6928e-03,  4.2451e-03,
+          4.4517e-03, -5.1524e-03,  1.0805e-02, -2.1935e-02, -1.7575e-02,
+         -1.2529e-02, -2.2191e-02, -1.0854e-02, -9.4462e-03, -2.9102e-02,
+          2.6752e-02, -1.0919e-02, -2.6724e-02,  8.3694e-04,  2.9832e-03,
+          1.4416e-02, -2.9906e-02,  2.3556e-02, -6.6624e-03,  2.6671e-02,
+         -3.6474e-02,  1.7237e-02, -2.5176e-02,  6.5560e-03, -2.6062e-02,
+         -2.3838e-02,  3.0629e-02,  2.5382e-02,  1.2302e-02, -1.1665e-02,
+         -7.0603e-03,  1.9931e-02,  2.3401e-02, -2.6047e-03, -2.7728e-02,
+         -1.7212e-02,  2.3061e-02, -2.5961e-02,  3.9764e-04, -2.9022e-02,
+         -1.5546e-03,  4.5519e-03,  2.3589e-02, -3.5005e-02,  4.1890e-03,
+         -1.5586e-02,  1.2389e-02, -2.1045e-02,  1.6377e-03, -1.1328e-02,
+          1.0195e-02,  6.4322e-03, -3.8431e-02,  2.2918e-02, -4.0123e-03,
+          6.6680e-02,  4.1135e-02, -1.5031e-02, -1.3550e-02, -2.2566e-02,
+         -2.3622e-03, -2.9323e-02,  2.1756e-02,  1.8399e-03, -4.2460e-03,
+         -1.5128e-03, -2.4731e-02,  1.8663e-02,  1.3469e-02, -1.3897e-02,
+          2.6399e-02, -8.0740e-03, -4.6753e-03,  3.9857e-02,  6.2364e-03,
+          2.2371e-03,  2.1501e-03,  5.9443e-02,  1.3574e-02,  7.6483e-03,
+         -6.2290e-03,  1.4324e-02,  1.2572e-02,  2.7331e-02, -6.0165e-03,
+         -5.9154e-03, -3.7000e-02,  1.4001e-02,  1.2869e-02, -2.8854e-02,
+         -9.4147e-03,  8.3965e-03, -1.4530e-03, -7.4215e-03,  9.0369e-03,
+         -2.4612e-02,  2.0625e-02,  2.2329e-02, -1.5216e-02,  1.4947e-03,
+         -3.6020e-02, -2.0702e-02, -4.0410e-02, -1.3157e-02, -1.5085e-02,
+          1.2911e-02, -2.7552e-02, -2.9781e-02, -4.7424e-03,  2.0521e-02,
+         -4.0043e-02, -4.8763e-02, -1.3175e-02,  2.6802e-02,  2.8869e-02,
+          6.5014e-03, -2.3213e-02,  1.4438e-02, -7.6318e-03, -1.9928e-03,
+          1.8509e-03,  2.9728e-03,  1.5225e-02, -2.9405e-03, -7.2875e-03,
+          2.9562e-05, -1.8661e-02,  9.1341e-03, -2.4919e-02,  2.9786e-02,
+          9.5186e-03,  1.5435e-02, -1.1080e-02,  1.1192e-02, -2.7315e-03,
+          6.9769e-05, -1.5392e-02,  4.9892e-03,  7.9857e-03,  2.0063e-02,
+         -2.0283e-02, -1.2596e-02, -4.1985e-04, -6.9686e-03, -5.4704e-02,
+         -1.9142e-02,  9.9706e-03,  2.3217e-02, -5.0579e-03, -4.9132e-02,
+          2.0023e-02, -2.6238e-02,  1.0709e-02,  2.1528e-02, -1.6390e-03,
+         -6.7829e-03,  1.3211e-02, -9.6793e-03,  1.3130e-02, -1.2878e-02,
+          1.7365e-02,  1.2509e-02,  1.2986e-03, -3.9292e-02,  9.5784e-03,
+         -8.0514e-03, -3.5619e-02, -3.2298e-02,  6.5933e-04,  9.9298e-03,
+          3.7268e-02, -3.4047e-02, -7.8385e-03,  2.3999e-02,  1.0386e-02,
+          1.7853e-02, -1.0122e-04,  5.2483e-04, -7.3150e-03,  1.0818e-02,
+          1.6245e-02, -3.5619e-02, -9.9190e-03,  4.0132e-03,  9.7788e-03,
+          2.7039e-02, -4.7858e-02, -2.0010e-02, -2.3702e-02,  7.8376e-04,
+         -2.5326e-02,  1.1698e-02, -1.3041e-02,  3.8634e-03,  9.3083e-03,
+          4.8204e-03,  3.9503e-02, -4.1356e-03]], requires_grad=True)
+model.Qformer.bert.encoder.layer[10].experts.gate.weight
+
+layer 11
+0:
+model.Qformer.bert.encoder.layer[11].output.dense.weight.grad
+model.Qformer.bert.encoder.layer[11].intermediate.dense.weight.grad
+
+nan:
+model.Qformer.bert.encoder.layer[11].attention.output.dense.weight.grad
+model.Qformer.bert.encoder.layer[11].attention.self.query.weight.grad
+model.Qformer.bert.encoder.layer[11].experts.intermediate_query.dense.weight.grad
+model.Qformer.bert.encoder.layer[11].experts.output_query.dense.weight.grad
+
+None:
+model.Qformer.bert.encoder.layer[11].intermediate_query.dense.weight.grad
+model.Qformer.bert.encoder.layer[11].output_query.dense.weight.grad
+
+layer 8
+0:
+model.Qformer.bert.encoder.layer[8].experts.experts[0].intermediate_query.dense.weight.grad
+model.Qformer.bert.encoder.layer[8].experts.experts[2].intermediate_query.dense.weight.grad
+model.Qformer.bert.encoder.layer[8].experts.experts[0].output_query.dense.weight.grad
+model.Qformer.bert.encoder.layer[8].experts.experts[2].output_query.dense.weight.grad
+
+nan:
+model.Qformer.bert.encoder.layer[8].experts.experts[1].intermediate_query.dense.weight.grad
+model.Qformer.bert.encoder.layer[8].experts.experts[1].output_query.dense.weight.grad
+(Qformer)model.Qformer.bert.encoder.layer[8].intermediate_query.dense.weight.grad
+
+None:
+model.Qformer.bert.encoder.layer[8].experts.gate.weight.grad == None
+model.Qformer.bert.encoder.layer[8].experts.gate.weight.requires_grad == True
+
+
+model.Qformer.bert.encoder.layer[6].experts.gate.weight
+Qformer.bert.encoder.layer.6.experts.gate.weight
+
+tensor([[-0.0089, -0.0123, -0.0168,  ..., -0.0072,  0.0295, -0.0167],
+        [ 0.0305,  0.0277, -0.0215,  ...,  0.0149,  0.0016, -0.0415],
+        [ 0.0199,  0.0151,  0.0237,  ...,  0.0007,  0.0023,  0.0167]],
+               requires_grad=True)
+
+tensor([[-0.0089, -0.0123, -0.0168,  ..., -0.0072,  0.0295, -0.0167],
+        [ 0.0305,  0.0277, -0.0215,  ...,  0.0149,  0.0016, -0.0415],
+        [ 0.0199,  0.0151,  0.0237,  ...,  0.0007,  0.0023,  0.0167]],
+       requires_grad=True)
+
+
+tensor([[ 4.5972e-02, -1.5231e-02, -6.9533e-03,  3.2431e-02, -7.9703e-03,
+          1.5567e-02,  2.9619e-03, -2.2609e-04,  1.8580e-02, -2.8783e-02,
+          1.3093e-02, -1.0594e-02,  1.1918e-02,  4.4701e-02,  2.0108e-02,
+         -1.1011e-03, -8.2449e-03,  8.8876e-03,  4.6096e-03,  2.3274e-02,
+         -9.2557e-03,  2.5704e-03,  1.8919e-02, -5.3251e-03, -3.2665e-03,
+         -3.2663e-02, -5.6756e-02, -2.3400e-02,  1.3674e-02, -6.6185e-03,
+          1.4429e-03,  1.2354e-02,  2.5934e-03,  2.1895e-02, -1.9793e-02,
+          1.5497e-03,  4.3056e-03, -4.0023e-02,  9.8740e-03,  3.8631e-03,
+         -1.2918e-02, -3.6782e-02, -9.8365e-03,  3.2182e-02,  2.3729e-02,
+          2.3509e-03,  1.8473e-02,  1.5583e-02, -1.1029e-02, -1.0738e-02,
+         -3.0278e-02, -9.8731e-03, -1.0500e-02,  7.9832e-05, -1.0345e-02,
+          8.2803e-03, -5.9923e-03, -1.2669e-02,  1.2065e-03,  7.5720e-03,
+         -1.9286e-02,  4.0070e-02,  3.6221e-03, -1.7486e-02,  2.1725e-02,
+         -3.3231e-02,  7.3948e-03, -1.0924e-02,  3.1448e-02,  1.2101e-02,
+          6.1737e-03, -2.0851e-02, -3.7964e-02,  8.0938e-03, -8.8967e-03,
+          2.5925e-02, -7.8063e-04,  8.6102e-03,  2.7370e-02,  1.2323e-02,
+          4.0606e-03,  3.9316e-02, -1.0837e-02, -2.6835e-03,  3.1941e-03,
+         -1.2017e-02, -2.3022e-02,  8.3533e-03, -2.2668e-02,  1.4438e-02,
+         -2.3664e-02,  4.5595e-02, -1.0962e-02,  1.7547e-02, -1.6739e-03,
+          1.2048e-02,  2.0544e-02,  2.8837e-02, -1.6736e-02,  2.1207e-02,
+          8.7612e-03,  2.8757e-02, -3.8561e-03,  8.4050e-03, -1.1503e-02,
+         -5.8332e-03,  1.5734e-02, -1.0773e-02,  7.5827e-03,  6.5794e-03,
+          2.4291e-02,  2.6811e-02,  1.1681e-02, -3.3246e-02,  4.5776e-03,
+         -9.0628e-04, -2.9400e-02,  4.2933e-03,  1.5885e-03,  5.5757e-02,
+          7.5518e-03,  1.0099e-02,  5.3507e-03, -3.0182e-02,  2.0830e-02,
+          1.0102e-02, -9.3074e-03,  3.1161e-02, -1.7800e-02, -4.4445e-03,
+         -3.1503e-02,  2.3028e-02,  8.3472e-03,  7.4444e-03,  1.8838e-02,
+         -1.1977e-02, -2.6713e-02,  1.1364e-02,  8.3522e-04,  3.3736e-03,
+          6.9425e-03, -2.0632e-02,  1.8155e-02, -2.1711e-02, -3.4703e-02,
+         -3.6268e-03, -4.8810e-03, -2.8142e-02, -1.5781e-02, -3.3166e-02,
+         -2.9910e-02, -9.7459e-03, -6.7474e-03,  1.7988e-02,  9.0176e-03,
+          1.9452e-02,  4.2009e-02,  1.7217e-02,  1.4959e-02, -1.6552e-02,
+         -3.8206e-03, -2.4889e-02,  7.7993e-03, -1.9285e-02, -1.9770e-02,
+          2.6936e-02, -5.0484e-03, -2.5117e-02, -2.3122e-02,  1.3754e-02,
+          1.6025e-02, -9.1569e-03, -2.0068e-02, -1.6013e-02, -2.1775e-02,
+         -2.4154e-02,  6.2840e-03, -1.3684e-02,  2.5378e-02, -1.3166e-02,
+         -1.2201e-02,  1.0011e-02, -8.2324e-03, -5.6623e-03, -1.0383e-02,
+         -1.6251e-02,  1.0723e-02, -3.0207e-03, -6.9374e-03, -2.3161e-03,
+         -2.0850e-03, -3.4216e-02,  3.3997e-02,  3.7444e-02, -3.4273e-02,
+          1.5051e-02, -9.5605e-03, -2.6979e-03,  1.8848e-02,  2.3090e-02,
+          1.9669e-02, -3.9656e-02,  1.0453e-02,  5.2222e-03, -7.2493e-03,
+          1.4122e-02,  5.6583e-04, -1.3991e-02,  4.0975e-02,  1.3947e-02,
+          4.6919e-03,  7.9121e-03,  2.6936e-02,  1.2338e-02,  1.9048e-02,
+          7.7740e-03, -6.4494e-03, -5.2965e-02,  8.1929e-03, -1.3503e-02,
+          3.7466e-03, -3.3504e-02, -8.1192e-03,  1.0463e-02, -2.1568e-02,
+          1.0076e-02, -1.3420e-02, -6.3353e-04,  7.4253e-03,  2.2281e-02,
+          5.2829e-03,  1.4102e-02,  1.4427e-02,  1.6331e-02, -2.3305e-04,
+         -4.4875e-02,  6.5300e-03,  2.4963e-02,  2.2141e-03,  3.9830e-02,
+          1.1405e-02,  8.6810e-03, -2.0404e-03, -1.8579e-03,  1.4765e-02,
+          5.4752e-03, -1.3364e-02, -1.3082e-03,  1.5873e-03,  1.9309e-02,
+          3.4367e-02,  1.8459e-02, -1.1323e-02, -1.8764e-02, -1.5370e-02,
+          3.6180e-03,  2.8253e-02, -1.6867e-03,  3.5884e-03, -2.1952e-02,
+         -1.5026e-02, -2.1070e-02, -1.2149e-02,  1.1162e-02, -3.0343e-02,
+         -4.1372e-02,  1.0880e-02,  2.2365e-02,  1.2896e-02,  2.9694e-02,
+         -8.4248e-03, -7.8876e-03, -6.7049e-03,  2.3700e-02,  4.7528e-03,
+         -7.8350e-03, -5.9220e-03,  3.8396e-02, -4.1598e-02, -2.3161e-03,
+          1.3419e-02,  7.1029e-03,  1.4195e-02, -1.1124e-02,  1.5812e-02,
+         -1.9789e-02, -2.3883e-02, -8.2788e-04,  1.4670e-02, -2.1482e-02,
+         -1.1182e-02, -1.6532e-02, -8.0637e-03, -3.7822e-02,  3.9402e-02,
+         -1.4097e-03, -7.6648e-03, -3.7156e-02,  2.5791e-02,  6.1038e-03,
+         -6.3429e-03,  3.2865e-03,  3.6277e-02,  9.4312e-03, -2.1003e-02,
+         -3.6885e-03,  1.7147e-02, -1.3079e-02, -4.9414e-02, -3.2066e-02,
+          1.4835e-02, -2.9742e-02,  1.8358e-02, -2.1733e-02,  3.0256e-03,
+          1.7825e-02,  1.1079e-02,  1.1619e-02, -2.3680e-02, -7.8721e-03,
+          2.4456e-03,  4.3608e-02, -4.5674e-03, -3.6818e-02,  3.3952e-02,
+          3.3108e-02, -3.1665e-03, -2.3468e-03,  1.5091e-02,  7.0856e-03,
+          1.1723e-02, -2.0713e-02, -6.9180e-03,  3.7929e-02,  3.7671e-03,
+          4.6663e-02,  9.5301e-03,  1.2638e-02, -6.5623e-03, -3.1771e-03,
+         -1.7568e-02,  1.8711e-03, -1.2310e-02,  2.1518e-02,  4.3408e-03,
+         -6.7171e-03, -5.0451e-03,  2.6870e-02, -1.9832e-02,  7.0422e-03,
+          1.1274e-02, -2.4637e-02, -4.8450e-03,  2.1892e-02, -2.6059e-02,
+          1.5605e-02, -1.1617e-02, -1.9273e-02, -8.6735e-04, -9.8002e-04,
+         -1.8553e-02,  2.1239e-02,  2.1078e-02, -1.2091e-02,  9.7025e-03,
+          1.3426e-02, -1.1710e-02, -2.2242e-03,  6.4133e-03, -1.4820e-02,
+          1.4682e-02,  3.0679e-02,  1.1526e-02,  1.0072e-02, -1.1572e-02,
+          2.6128e-02,  4.0879e-03, -1.7936e-02,  1.3715e-02, -2.3667e-02,
+          2.0419e-03, -1.6887e-02,  1.2595e-02, -2.1988e-02, -2.3777e-02,
+         -1.0399e-02,  2.4868e-03, -1.2265e-02, -1.8543e-02,  3.4672e-02,
+          2.1114e-02,  2.0523e-02,  7.6818e-03,  2.9282e-02, -5.9593e-03,
+         -2.8496e-02,  2.8482e-03,  3.6874e-04,  4.7455e-02, -2.9770e-02,
+         -2.0684e-02, -2.0749e-02, -5.7681e-02, -2.6175e-03, -2.4488e-02,
+         -5.2550e-03, -7.1191e-03,  3.8192e-02,  4.3438e-02,  5.4181e-03,
+          2.8392e-02,  1.9493e-02, -3.5262e-02,  1.4839e-02,  4.6481e-03,
+          1.7219e-02,  2.0160e-02,  4.9998e-03,  2.1316e-02, -8.7929e-04,
+         -2.1542e-02,  3.9816e-03,  1.5879e-02,  9.9231e-03,  1.3962e-02,
+         -5.3418e-03,  3.9857e-02,  2.0997e-02, -2.1291e-05,  1.8133e-02,
+         -1.2472e-02,  4.9437e-03, -1.5099e-02,  4.8860e-02,  6.1980e-03,
+          2.0197e-02,  1.3141e-04, -3.1087e-03, -2.2718e-03,  2.3804e-02,
+          6.0726e-03, -2.0485e-02, -2.0514e-02, -2.7679e-02, -3.0412e-02,
+         -1.7661e-02, -1.7462e-02,  7.5216e-03,  2.2238e-02,  1.1413e-03,
+          2.6647e-02, -2.3855e-02,  2.2652e-03, -4.3256e-03, -9.3274e-03,
+          2.5149e-02,  6.8432e-03,  4.2664e-03,  3.8221e-02,  7.7480e-03,
+          8.7203e-03, -1.2851e-03, -1.1325e-02, -1.0650e-02, -2.8079e-02,
+         -1.5375e-02,  2.2630e-02, -4.3439e-03,  1.3493e-02, -1.8223e-02,
+          9.9750e-03, -2.4560e-02,  1.0904e-03, -3.1198e-02,  4.7331e-03,
+          1.6713e-02, -1.7653e-02, -3.8674e-02,  1.5458e-02,  4.0555e-02,
+          6.9451e-03,  1.1988e-03,  8.0718e-04,  3.9985e-03, -2.2781e-02,
+          8.1173e-04,  2.0106e-02, -1.2800e-02, -1.2961e-02, -2.1273e-02,
+         -4.4104e-05, -3.6080e-02, -1.9392e-02,  3.2862e-02, -5.6041e-03,
+          2.3288e-02, -4.6795e-02,  1.7282e-02,  5.7052e-03,  2.2405e-02,
+          1.9871e-03, -1.4333e-02,  5.3773e-03,  4.3568e-02,  9.8980e-03,
+         -1.9403e-03,  1.8981e-02, -2.5712e-02, -3.3621e-03,  2.9886e-02,
+          1.3326e-03,  1.1318e-02, -3.3238e-03, -1.5494e-02, -3.0565e-02,
+          1.7137e-02, -2.7874e-02, -1.1257e-02,  3.2250e-02, -2.5293e-02,
+         -3.0693e-03, -2.7787e-02,  1.4931e-02,  2.4202e-03, -4.0572e-03,
+          5.0273e-03,  9.7496e-03,  2.2601e-02,  3.2389e-02, -1.1910e-02,
+          9.1037e-03,  5.6000e-02, -1.9640e-02,  1.5469e-02, -3.3027e-02,
+          1.4839e-02,  2.5071e-02, -1.2687e-02, -1.3466e-02,  1.9031e-02,
+         -7.3403e-03, -1.5207e-02, -1.4486e-02,  2.0678e-02, -4.1996e-02,
+          1.0585e-02,  3.6276e-02,  6.1149e-03,  1.6405e-02,  1.5643e-02,
+          1.5060e-02, -5.1235e-03, -2.2824e-02, -1.3752e-02, -1.5742e-02,
+          2.4032e-02, -2.1782e-03, -1.3158e-02,  3.9482e-03,  3.2267e-02,
+         -2.2632e-03,  1.2055e-02,  4.4731e-02,  1.8271e-02, -1.1486e-02,
+          1.7836e-02,  1.7886e-03, -2.4020e-02,  2.6064e-02, -2.2122e-04,
+          1.8643e-02, -2.9808e-02, -6.1845e-03, -4.4464e-03,  8.8374e-04,
+          1.5268e-02,  1.7205e-03,  5.7832e-02, -1.7486e-02,  1.1897e-02,
+          5.8081e-02,  1.7667e-02, -7.7282e-03,  1.4036e-02, -1.4936e-03,
+          6.0635e-04,  1.6124e-03, -1.6916e-02, -1.1239e-02,  1.8497e-02,
+          1.2334e-03, -2.0706e-02,  3.2959e-03,  2.9186e-02,  3.7506e-02,
+          1.2037e-02, -1.4903e-02,  8.5606e-03,  3.4136e-03,  1.1850e-02,
+         -7.4782e-03,  5.3924e-03, -2.4772e-02,  2.6840e-02, -2.7656e-02,
+         -3.2637e-02, -1.2779e-02,  1.0730e-02,  1.4096e-03,  3.1572e-02,
+          7.8976e-04,  3.1674e-02,  8.5333e-03, -1.2679e-02,  1.1176e-02,
+         -2.0446e-02,  1.8628e-02, -4.0158e-02, -2.3358e-02, -2.2504e-02,
+         -2.8759e-02, -1.4597e-02, -8.5879e-03,  1.0550e-02, -3.5556e-02,
+         -1.9046e-02, -1.9159e-02, -2.2703e-02, -7.2056e-03,  4.2380e-02,
+         -9.7475e-03, -2.4754e-02,  1.3992e-03, -1.0411e-02,  1.5708e-02,
+         -8.2899e-03, -6.4856e-03,  1.6359e-02, -5.1969e-04, -5.0958e-03,
+         -4.1232e-02,  2.7349e-03, -1.7723e-02,  1.3388e-02,  2.2776e-03,
+         -2.0786e-02, -1.8082e-02, -2.4866e-03,  2.2141e-02,  6.9998e-03,
+         -5.5714e-03,  2.1088e-02,  5.8745e-03,  1.2788e-02,  4.2977e-03,
+          5.8631e-03, -1.8121e-02,  1.9242e-03,  2.3622e-02,  1.4917e-02,
+         -5.3198e-03, -3.9222e-02, -2.4697e-02,  9.1218e-03, -1.0711e-02,
+          1.0268e-02,  1.5148e-02, -4.4508e-02,  4.6783e-03,  2.8093e-03,
+          9.1253e-03, -7.3281e-03,  1.0114e-03, -9.2369e-04,  1.4841e-02,
+          2.2642e-02,  2.3675e-02,  1.3902e-02, -5.6343e-03,  1.4851e-02,
+         -9.5169e-03, -3.1721e-02,  1.6696e-02,  2.9285e-02, -1.4090e-02,
+          2.1128e-02,  4.8656e-02,  3.8431e-02, -3.5470e-02, -4.8230e-03,
+         -1.6513e-02,  4.1917e-02,  8.9090e-03, -1.4022e-04,  4.0182e-03,
+          7.1723e-03,  3.1419e-02, -4.8508e-03,  1.7768e-03, -7.3688e-03,
+          3.4637e-03, -2.3227e-02,  3.9606e-05, -2.4731e-02, -1.3640e-02,
+         -5.1718e-03,  2.6662e-02, -1.2871e-02, -1.6009e-02, -5.3720e-03,
+          2.7397e-04, -3.4016e-03,  2.6429e-02,  3.8069e-02,  1.0929e-02,
+         -1.0620e-02,  1.2165e-02, -2.6018e-02,  1.6021e-02,  4.0644e-02,
+         -8.0898e-03, -3.5198e-02, -1.9602e-02,  2.4986e-02, -5.8400e-03,
+          3.2070e-02, -1.8265e-02, -5.4518e-03,  2.8195e-02,  5.5598e-02,
+         -3.9959e-02,  1.5521e-02, -2.8416e-02,  3.1130e-02, -1.0038e-02,
+          2.1522e-02, -1.1654e-02,  2.2382e-02, -5.4467e-03, -2.2840e-02,
+          2.7036e-03, -4.4607e-02, -4.1953e-02,  2.0079e-02, -5.0121e-03,
+         -1.7495e-02,  4.4070e-03,  3.7400e-04,  1.0899e-02,  1.7008e-02,
+         -1.6307e-02, -1.9986e-02, -2.3865e-02, -2.5618e-02, -2.9981e-02,
+         -2.7230e-03,  2.7079e-02,  5.2920e-03,  2.1069e-02, -2.5896e-02,
+         -1.6256e-02, -1.4182e-03,  1.1829e-02,  1.0360e-02,  2.8883e-02,
+         -6.8762e-03,  1.4032e-02, -4.3389e-03]], requires_grad=True)

test1.txt

@@ -0,0 +1,109 @@
+from torchviz import make_dot
+dot = make_dot(query_output.last_hidden_state, params=dict(self.Qformer.bert.named_parameters()))
+log_dir = '/mnt/pfs-guan-ssai/nlu/wanghanzi/multimodal/PromptMoE/'
+dot.render(filename="Pre_PromptMoE_RawProb_backward_graph", directory=log_dir, format="pdf")
+
+
+# Pre-Prompt-MoE
+model.Qformer.bert.encoder.layer[6].experts.gate.weight.grad
+model.Qformer.bert.encoder.layer[8].experts.gate.weight.grad
+model.Qformer.bert.encoder.layer[10].experts.gate.weight.grad
+model.Qformer.bert.encoder.layer[6].experts.experts[0].dense1.weight.grad
+model.Qformer.bert.encoder.layer[10].experts.experts[0].dense1.weight.grad
+model.Qformer.bert.encoder.layer[8].experts.experts[0].dense1.weight.grad
+model.Qformer.bert.encoder.layer[8].experts.experts[1].dense1.weight.grad
+model.Qformer.bert.encoder.layer[8].experts.experts[2].dense1.weight.grad
+
+
+model.Qformer.bert.encoder.layer[8].experts.gate.weight.grad
+model.Qformer.bert.encoder.layer[9].intermediate_query.dense.weight
+model.Qformer.bert.encoder.layer[9].intermediate_query.dense.weight.grad
+model.Qformer.bert.encoder.layer[10].intermediate.dense.weight.grad
+model.Qformer.bert.encoder.layer[11].intermediate.dense.weight.grad
+
+model.Qformer.bert.encoder.layer[10].intermediate_query.dense.weight
+model.Qformer.bert.encoder.layer[10].experts.experts[2].dense1.weight
+model.Qformer.bert.encoder.layer[10].experts.experts[1].dense1.weight
+model.Qformer.bert.encoder.layer[10].experts.experts[0].dense1.weight
+model.Qformer.bert.encoder.layer[10].intermediate_query.dense.weight == model.Qformer.bert.encoder.layer[10].experts.experts[0].dense1.weight
+
+# Pre-MoE gate-sentence
+# model.Qformer.bert.encoder.layer[8].experts.gate.weight.grad 不更新
+
+# Pre-MoE gate-token
+# 正常更新
+
+# Post-MoE gate-sentence
+model.Qformer.bert.encoder.layer[8].experts.gate.weight.grad
+# model.Qformer.bert.encoder.layer[8].experts.gate.weight.grad 正常更新
+# model.Qformer.bert.encoder.layer[6].experts.gate.weight.grad 全是0/-0
+# model.Qformer.bert.encoder.layer[10].experts.gate.weight.grad 全是0/-0
+
+# Route-MoE 
+# Pre-MoE 算的beam_scores有问题
+
+# Post-Route 会更新多个expert的参数；会更新gate的参数
+# Layer 6 更新了两个expert的参数 (layer 6 layer 8)
+# model.Qformer.bert.encoder.layer[11].intermediate.dense.weight.grad 是0？都是0
+# model.Qformer.bert.encoder.layer[11].output.dense.weight.grad
+
+model.Qformer.bert.encoder.layer[6].experts.gate.weight.grad
+model.Qformer.bert.encoder.layer[6].experts.experts[0].intermediate_query.dense.weight.grad
+model.Qformer.bert.encoder.layer[6].experts.experts[1].intermediate_query.dense.weight.grad
+
+model.Qformer.bert.encoder.layer[7].experts.gate.weight.grad
+model.Qformer.bert.encoder.layer[7].experts.experts[0].intermediate_query.dense.weight.grad
+model.Qformer.bert.encoder.layer[7].experts.experts[1].intermediate_query.dense.weight.grad
+
+model.Qformer.bert.encoder.layer[8].experts.gate.weight.grad
+model.Qformer.bert.encoder.layer[8].experts.experts[0].intermediate_query.dense.weight.grad
+model.Qformer.bert.encoder.layer[8].experts.experts[1].intermediate_query.dense.weight.grad
+
+model.Qformer.bert.encoder.layer[9].experts.gate.weight.grad
+model.Qformer.bert.encoder.layer[9].experts.experts[0].intermediate_query.dense.weight.grad
+model.Qformer.bert.encoder.layer[9].experts.experts[1].intermediate_query.dense.weight.grad
+
+model.Qformer.bert.encoder.layer[10].experts.gate.weight.grad
+model.Qformer.bert.encoder.layer[10].experts.experts[0].intermediate_query.dense.weight.grad
+model.Qformer.bert.encoder.layer[10].experts.experts[1].intermediate_query.dense.weight.grad
+
+model.Qformer.bert.encoder.layer[11].experts.gate.weight.grad
+model.Qformer.bert.encoder.layer[11].experts.experts[0].intermediate_query.dense.weight.grad
+model.Qformer.bert.encoder.layer[11].experts.experts[1].intermediate_query.dense.weight.grad
+
+
+(Pdb) [p for n, p in self.model.named_parameters() if n == 'Qformer.bert.encoder.layer.10.experts.experts.0.dense1.weight']
+[Parameter containing:
+tensor([[-0.0328,  0.0414,  0.0010,  ..., -0.0068,  0.0244,  0.0587],
+        [ 0.0120,  0.0458,  0.0171,  ..., -0.0439, -0.0107, -0.0397],
+        [ 0.0239,  0.0191, -0.0145,  ...,  0.0008, -0.0067,  0.0090],
+        ...,
+        [ 0.0174, -0.0465, -0.0106,  ..., -0.0095,  0.0153, -0.0195],
+        [-0.0151, -0.0082, -0.0320,  ..., -0.0016, -0.0232, -0.0147],
+        [ 0.0142, -0.0286,  0.0161,  ..., -0.0160, -0.0306, -0.0272]],
+       device='cuda:0', requires_grad=True)]
+(Pdb) [p for n, p in self.model.named_parameters() if n == 'Qformer.bert.encoder.layer.8.experts.experts.0.dense1.weight']
+[Parameter containing:
+tensor([[ 0.0024,  0.0218, -0.0186,  ..., -0.0178, -0.0067,  0.0820],
+        [-0.0759, -0.0002, -0.0548,  ...,  0.0292,  0.0531,  0.0779],
+        [-0.0220, -0.0037, -0.0520,  ..., -0.0426, -0.0261, -0.0357],
+        ...,
+        [-0.0448,  0.0471,  0.0133,  ..., -0.0062, -0.0217, -0.0203],
+        [ 0.0532,  0.0197,  0.0320,  ..., -0.0010, -0.0838,  0.0682],
+        [ 0.0284,  0.0038, -0.0007,  ..., -0.0305,  0.0296,  0.0056]],
+       device='cuda:0', requires_grad=True)]
+(Pdb) [p for n, p in self.model.named_parameters() if n == 'Qformer.bert.encoder.layer.6.experts.experts.0.dense1.weight']
+[Parameter containing:
+tensor([[ 6.5176e-02, -4.6473e-02, -2.7396e-02,  ...,  2.1774e-03,
+          6.1457e-02,  1.9180e-03],
+        [ 7.3707e-03,  6.1392e-02, -2.7108e-02,  ...,  4.0778e-02,
+         -1.9791e-02, -1.1612e-02],
+        [ 2.1193e-02, -3.8323e-02, -6.0238e-02,  ..., -1.4539e-02,
+          9.2965e-02,  3.9153e-02],
+        ...,
+        [ 5.3203e-03, -1.7276e-02, -3.2191e-02,  ..., -1.6435e-02,
+         -1.8553e-02, -2.8158e-02],
+        [-6.9853e-02,  9.2719e-03, -1.8895e-03,  ..., -2.6425e-02,
+          1.4880e-03,  3.4505e-02],
+        [-1.2168e-03,  3.7038e-02,  4.8047e-02,  ..., -3.4523e-03,
+         -1.3030e-05, -1.4778e-02]], device='cuda:0', requires_grad=True)]