debug

server/text_generation_server/models/custom_modeling/flash_gpt2_modeling.py

@@ -175,6 +175,7 @@ class FlashGPT2Attention(torch.nn.Module):
         prefix: str,
         config,
         weights,
+        layer_idx
     ):
         super().__init__()
         self.num_heads = config.num_attention_heads
@@ -189,6 +190,7 @@ class FlashGPT2Attention(torch.nn.Module):
                 f"and `num_shards`: {weights.process_group.size()}"
             )
         self.num_heads = self.num_heads // weights.process_group.size()
+        self.layer_idx = layer_idx
 
         self.query_key_value = load_qkv(
             config,
@@ -218,6 +220,7 @@ class FlashGPT2Attention(torch.nn.Module):
         slots,
         input_lengths,
         max_s,
+        step
     ):
         query, key, value = self.query_key_value(hidden_states).split(
             self.head_size * self.num_heads, dim=1
@@ -225,6 +228,11 @@ class FlashGPT2Attention(torch.nn.Module):
         query = query.view(-1, self.num_heads, self.head_size)
         key = key.view(-1, self.num_heads, self.head_size)
         value = value.view(-1, self.num_heads, self.head_size)
+        
+        if self.layer_idx < 5:
+            torch.save(query, f"query_step{step}_layer{self.layer_idx}.pt")
+            torch.save(key, f"key_step{step}_layer{self.layer_idx}.pt")
+            torch.save(value, f"value_step{step}_layer{self.layer_idx}.pt")
 
         reshape_and_cache(key, value, kv_cache[0], kv_cache[1], slots)
 
@@ -256,6 +264,9 @@ class FlashGPT2Attention(torch.nn.Module):
                 input_lengths,
                 max_s,
             )
+        
+        if self.layer_idx < 5:
+            torch.save(attn_output, f"flash_attn_out_step{step}_layer{self.layer_idx}.pt")
 
         return self.o_proj(attn_output.view(-1, self.num_heads * self.head_size))
 
@@ -301,7 +312,7 @@ class FlashGPT2Layer(nn.Module):
     def __init__(self, prefix, config, weights, layer_idx):
         super().__init__()
         self.self_attn = FlashGPT2Attention(
-            prefix=f"{prefix}.attn", config=config, weights=weights
+            prefix=f"{prefix}.attn", config=config, weights=weights, layer_idx=layer_idx
         )
         self.mlp = GPT2MLP(prefix=f"{prefix}.mlp", config=config, weights=weights)
 
@@ -342,6 +353,7 @@ class FlashGPT2Layer(nn.Module):
             slots,
             input_lengths,
             max_s,
+            step
         )
 
         if self.layer_idx < 5: