LLM: fix llama 4.36 forward(#10047)
This commit is contained in:
		
							parent
							
								
									53a5140eff
								
							
						
					
					
						commit
						3685622f29
					
				
					 1 changed files with 1 additions and 1 deletions
				
			
		| 
						 | 
				
			
			@ -638,7 +638,7 @@ def llama_attention_forward_4_36(
 | 
			
		|||
            "Please make sure use `attention_mask` instead.`"
 | 
			
		||||
        )
 | 
			
		||||
 | 
			
		||||
    bsz, q_len, _ = hidden_states.size()
 | 
			
		||||
    bsz, q_len, hidden_size = hidden_states.size()
 | 
			
		||||
    device = hidden_states.device
 | 
			
		||||
    # for flash attention
 | 
			
		||||
    original_dtype = hidden_states.dtype
 | 
			
		||||
| 
						 | 
				
			
			
 | 
			
		|||
		Loading…
	
		Reference in a new issue