add fp8 e5 to use_xmx (#10015)
This commit is contained in:
		
							parent
							
								
									9e18ea187f
								
							
						
					
					
						commit
						a3322e2a6c
					
				
					 1 changed files with 3 additions and 2 deletions
				
			
		| 
						 | 
				
			
			@ -23,7 +23,8 @@ from bigdl.llm.transformers.utils import get_ipex_version, get_xpu_device_type
 | 
			
		|||
 | 
			
		||||
SYM_INT4 = ggml_tensor_qtype["sym_int4"]
 | 
			
		||||
SYM_INT8 = ggml_tensor_qtype["sym_int8"]
 | 
			
		||||
FP8 = ggml_tensor_qtype["fp8"]
 | 
			
		||||
FP8E4 = ggml_tensor_qtype["fp8_e4m3"]
 | 
			
		||||
FP8E5 = ggml_tensor_qtype["fp8_e5m2"]
 | 
			
		||||
 | 
			
		||||
 | 
			
		||||
def init_kv_cache(batch_size, num_heads, head_dim, current_length, max_length, dtype, device):
 | 
			
		||||
| 
						 | 
				
			
			@ -297,7 +298,7 @@ def use_xmx(x: torch.Tensor, qtype: int):
 | 
			
		|||
    return (
 | 
			
		||||
        os.environ.get("BIGDL_LLM_XMX_DISABLED", "0") != "1"
 | 
			
		||||
        and device in ["arc", "flex", "pvc"]
 | 
			
		||||
        and qtype in [SYM_INT4, SYM_INT8, FP8]
 | 
			
		||||
        and qtype in [SYM_INT4, SYM_INT8, FP8E4, FP8E5]
 | 
			
		||||
        and (
 | 
			
		||||
            (device == "pvc" and 1 < x.size(0) <= 16)
 | 
			
		||||
            or
 | 
			
		||||
| 
						 | 
				
			
			
 | 
			
		|||
		Loading…
	
		Reference in a new issue