Update rwkv train.

wit/inference.py

@@ -12,6 +12,7 @@ if __name__ == "__main__":
     checkpoint_path = "log/bigger/version_1/checkpoints/epoch=14-step=74040.ckpt"
     checkpoint_path = "log/bigger/version_3/checkpoints/epoch=46-step=231992.ckpt"
     checkpoint_path = "log/bigger/version_8/checkpoints/epoch=49-step=246800.ckpt"
+    # checkpoint_path = "log/bigger/version_11/checkpoints/epoch=25-step=128336.ckpt"
 
     qwen = QwenModule.load_from_checkpoint(checkpoint_path=checkpoint_path)
     qwen.eval()
@@ -41,4 +42,4 @@ if __name__ == "__main__":
         if item[i] != next_token:
             node.set_seq_prop(i, "ERR_" + str(next_token))
             print(str(item[i]) + "  " + str(next_token) + "  ERROR")
-    # node.print()
+    node.print()

wit/model/modeling_rwkv7.py

@@ -6,24 +6,23 @@ import torch.nn.functional as F
 import torch.utils.checkpoint
 from torch.nn import CrossEntropyLoss
 from torch import nn
-
+import torch.nn.init as init
 
 # for 0.1B
-n_layer = 12
+n_layer = 3
-n_embd = 768
+n_embd = 256
 D_DECAY_LORA = 64
 D_AAA_LORA = 64
 D_MV_LORA = 32
 D_GATE_LORA = 128
 
-dim_att = 768
+dim_att = n_embd
-dim_ffn = 3072
+dim_ffn = n_embd * 4
 
-# vocab_size = 65536
+vocab_size = 32
-vocab_size = 65536
 
 # DTYPE = torch.bfloat16
-DTYPE = torch.half  # better
+DTYPE = torch.float32  # better
 
 head_size_a = 64  # don't change
 HS = head_size_a
@@ -172,8 +171,8 @@ class RWKV_CMix_x070(nn.Module):
         super().__init__()
         self.layer_id = layer_id
 
-        with torch.no_grad():
+        # with torch.no_grad():
-            self.x_k = nn.Parameter(torch.empty(1, 1, n_embd))
+        self.x_k = nn.Parameter(torch.empty(1, 1, n_embd))
 
         self.key = nn.Linear(n_embd, dim_ffn, bias=False)
         self.value = nn.Linear(dim_ffn, n_embd, bias=False)
@@ -218,8 +217,6 @@ class Block(nn.Module):
 class RWKV(nn.Module):
     def __init__(self):
         super().__init__()
-        dim_att = n_embd
-        dim_ffn = n_embd * 4
         self.emb = nn.Embedding(vocab_size, n_embd)
 
         self.blocks = nn.ModuleList([Block(i) for i in range(n_layer)])
@@ -245,10 +242,13 @@ class RWKVLMHeadModel(nn.Module):
     def __init__(self, config):
         super().__init__()
         self.config = config
-        self.transformer = RWKV()
+        self.rwkv = RWKV()
         self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
         self.hook_attention = None
 
+        for name, param in self.rwkv.named_parameters():
+            init.normal_(param.data)
+
     def forward(
         self,
         input_ids: Optional[torch.LongTensor] = None,
@@ -256,7 +256,7 @@ class RWKVLMHeadModel(nn.Module):
         token_type_ids: Optional[torch.LongTensor] = None,
         **kwargs,
     ):
-        lm_logits = self.transformer(input_ids)
+        lm_logits = self.rwkv(input_ids)
 
         loss = None
         if labels is not None:

wit/model/qwen_module.py

@@ -13,7 +13,6 @@ import pytorch_lightning as pl
 import torch
 import torchmetrics
 
-from model.modeling_wit import QWenLMHeadModel
 from configuration import ModelConfig, TrainConfig
 
 
@@ -166,7 +165,7 @@ class ModelRunner:
 
 
 class QwenModule(pl.LightningModule):
-    def __init__(self, conf: TrainConfig = None):
+    def __init__(self, conf: TrainConfig, model):
         self.config = conf
         pretrained_model_dir = conf.pretrain_model_name
         learning_rate = conf.learning_rate
@@ -176,7 +175,6 @@ class QwenModule(pl.LightningModule):
         self.save_hyperparameters()
         if mconf == None:
             mconf = ModelConfig()
-        model = QWenLMHeadModel(mconf)
         if pretrained_model_dir != None:
             from modelscope import snapshot_download
 

wit/train.py

@@ -2,8 +2,9 @@ import pytorch_lightning as pl
 import torch
 
 from model.qwen_module import QwenModule
-from model.tokenization_qwen import QWenTokenizer
+from model.modeling_wit import QWenLMHeadModel
-from logger import MLFLogger, TBLogger
+from model.modeling_rwkv7 import RWKVLMHeadModel
+from logger import TBLogger
 
 import configuration
 import dataset.dataset as ds
@@ -42,7 +43,9 @@ if __name__ == "__main__":
 
     torch.manual_seed(conf.seed)
     np.random.seed(conf.seed)
-    qwen = QwenModule(conf)
+    model = QWenLMHeadModel(conf.model_config)
+    # model = RWKVLMHeadModel(conf.model_config)
+    qwen = QwenModule(conf, model)
 
     train_dataloader, val_dataloader = ds.InitDataset(conf)
     # for i in range(len(train_dataloader)):