fixing nan in large models

src/model.py

@@ -292,8 +292,9 @@ class Block(nn.Module):
         self.ln2 = nn.LayerNorm(config.n_embd)
 
         if config.model_type == 'RWKV':
-            self.ln1 = nn.Identity() # remove first LayerNorm -> faster convergence for deep models
-            self.ln2 = SimpleRMSNorm(config.n_embd) # SimpleRMSNorm is good enough for RWKV -> less parameters
+            self.ln1 = nn.Identity()
+            # self.ln1 = SimpleRMSNorm(config.n_embd) # turn on this if you see nan in large RWKV models
+            self.ln2 = SimpleRMSNorm(config.n_embd)
             self.attn = RWKV_TimeMix(config)
             self.mlp = RWKV_ChannelMix(config)
         elif config.model_type == 'RotaryMHA':
@@ -319,7 +320,7 @@ class GPT(nn.Module):
         self.blocks = nn.Sequential(*[Block(config) for _ in range(config.n_layer)])
 
         if config.model_type == 'RWKV':
-            self.ln_f = SimpleRMSNorm(config.n_embd) # SimpleRMSNorm is good enough for RWKV -> less parameters
+            self.ln_f = SimpleRMSNorm(config.n_embd)
         else:
             self.ln_f = nn.LayerNorm(config.n_embd)
 
@@ -328,6 +329,16 @@ class GPT(nn.Module):
         self.ctx_size = config.ctx_size
         self.apply(self._init_weights)
 
+        if self.config.model_type == 'RWKV':
+            ww = self.state_dict()
+            for k in ww: # reduce weight to avoid nan
+                if 'receptance.weight' in k:
+                    ww[k] /= math.pow(config.n_embd, 0.5)
+                elif 'key.weight' in k:
+                    ww[k] /= math.pow(config.n_embd, 0.25)
+                elif 'value.weight' in k:
+                    ww[k] /= math.pow(config.n_embd, 0.25)
+
         logger.info("number of parameters: %e", sum(p.numel() for p in self.parameters()))
 
     def get_ctx_size(self):
@@ -335,11 +346,14 @@ class GPT(nn.Module):
 
     def _init_weights(self, module):
         if isinstance(module, (nn.Linear, nn.Embedding)):
-            # if self.config.model_type == 'RWKV' and isinstance(module, nn.Linear):
-            #     gain_layer = min(3, module.weight.shape[0] / module.weight.shape[1])
-            #     depth_factor = min(1, 1 / math.sqrt(self.config.n_layer / 5))
-            #     nn.init.orthogonal_(module.weight, gain = gain_layer * depth_factor) # will nan for large models
-            # else:
+            if self.config.model_type == 'RWKV':
+                if isinstance(module, nn.Linear):
+                    gain_layer = min(3, module.weight.shape[0] / module.weight.shape[1])
+                    depth_factor = 1 # min(1, 1 / math.sqrt(self.config.n_layer / 5))
+                    nn.init.orthogonal_(module.weight, gain = gain_layer * depth_factor)
+                else:
+                    nn.init.orthogonal_(module.weight, gain = 1.0)
+            else:
                 module.weight.data.normal_(mean=0.0, std=0.01)
         if isinstance(module, nn.Linear) and module.bias is not None:
             module.bias.data.zero_()

train.py

@@ -29,8 +29,8 @@ model_level = 'character' # 'character' or 'word'
 
 ctx_size = 256 if model_level == 'character' else 128
 nLayers = 5
-nHead = 8
-nEmb = 512
+nHead = 8                                                       # if you see nan in large RWKV models, turn on 'self.ln1' in model.py
+nEmb = nHead * 64
 
 lr_initial = 6e-4 if model_type == 'RWKV' else 4e-4             # RWKV can use higher lr
 lr_final = 2e-4
@@ -39,7 +39,7 @@ lr_initial /= math.sqrt(nLayers / 5) # lower lr for deep models; higher lr for s
 lr_final /= math.sqrt(nLayers / 5)
 
 betas = (0.9, 0.99)
-weight_decay = 0 if model_type == 'RWKV' else 0.01 # seems wd is not very useful when you have enough data
+weight_decay = 0 if model_type == 'RWKV' else 0.01              # seems wd is not very useful when we have enough data
 
 nepoch = 50                                                     # just a quick test. the 'epoch' here is very short
 nbatchsz = 64