fix cpu and xpu issue (#2116)

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

server/text_generation_server/models/flash_causal_lm.py

@@ -768,7 +768,10 @@ class FlashCausalLM(Model):
         empty_cache()
 
         element_size = torch.tensor([], dtype=dtype).element_size()
-        x = BLOCK_SIZE // element_size
+        if SYSTEM == "ipex" and device.type == "xpu":
+            x = 1
+        else:
+            x = BLOCK_SIZE // element_size
 
         if SYSTEM == "ipex" and device == torch.device("cpu"):
             self.kv_cache = [

server/text_generation_server/models/flash_gpt2.py

@@ -37,9 +37,10 @@ class FlashGPT2(FlashCausalLM):
         elif SYSTEM == "ipex":
             if hasattr(torch, "xpu") and torch.xpu.is_available():
                 device = torch.device(f"xpu:{rank}")
+                dtype = torch.float16 if dtype is None else dtype
             else:
                 device = torch.device("cpu")
-            dtype = torch.float16 if dtype is None else dtype
+                dtype = torch.bfloat16 if dtype is None else dtype
         else:
             raise NotImplementedError("FlashGPT2 is only available on GPU")
 

server/text_generation_server/models/flash_llama.py

@@ -37,9 +37,10 @@ class FlashLlama(FlashCausalLM):
         elif SYSTEM == "ipex":
             if hasattr(torch, "xpu") and torch.xpu.is_available():
                 device = torch.device(f"xpu:{rank}")
+                dtype = torch.float16 if dtype is None else dtype
             else:
                 device = torch.device("cpu")
-            dtype = torch.float16 if dtype is None else dtype
+                dtype = torch.bfloat16 if dtype is None else dtype
         else:
             raise NotImplementedError("FlashLlama is only available on GPU")
 

server/text_generation_server/models/flash_mistral.py

@@ -41,9 +41,10 @@ class BaseFlashMistral(FlashCausalLM):
         elif SYSTEM == "ipex":
             if hasattr(torch, "xpu") and torch.xpu.is_available():
                 device = torch.device(f"xpu:{rank}")
+                dtype = torch.float16 if dtype is None else dtype
             else:
                 device = torch.device("cpu")
-            dtype = torch.float16 if dtype is None else dtype
+                dtype = torch.bfloat16 if dtype is None else dtype
         else:
             raise NotImplementedError("FlashMistral is only available on GPU")
 

server/text_generation_server/models/flash_neox.py

@@ -36,9 +36,10 @@ class FlashNeoXSharded(FlashCausalLM):
         elif SYSTEM == "ipex":
             if hasattr(torch, "xpu") and torch.xpu.is_available():
                 device = torch.device(f"xpu:{rank}")
+                dtype = torch.float16 if dtype is None else dtype
             else:
                 device = torch.device("cpu")
-            dtype = torch.float16 if dtype is None else dtype
+                dtype = torch.bfloat16 if dtype is None else dtype
         else:
             raise NotImplementedError("FlashNeoX is only available on GPU")
 

server/text_generation_server/models/flash_rw.py

@@ -37,9 +37,10 @@ class FlashRWSharded(FlashCausalLM):
         elif SYSTEM == "ipex":
             if hasattr(torch, "xpu") and torch.xpu.is_available():
                 device = torch.device(f"xpu:{rank}")
+                dtype = torch.float16 if dtype is None else dtype
             else:
                 device = torch.device("cpu")
-            dtype = torch.float16 if dtype is None else dtype
+                dtype = torch.bfloat16 if dtype is None else dtype
         else:
             raise NotImplementedError("FlashRW is only available on GPU")
 

server/text_generation_server/models/flash_santacoder.py

@@ -40,9 +40,10 @@ class FlashSantacoderSharded(FlashCausalLM):
         elif SYSTEM == "ipex":
             if hasattr(torch, "xpu") and torch.xpu.is_available():
                 device = torch.device(f"xpu:{rank}")
+                dtype = torch.float16 if dtype is None else dtype
             else:
                 device = torch.device("cpu")
-            dtype = torch.float16 if dtype is None else dtype
+                dtype = torch.bfloat16 if dtype is None else dtype
         else:
             raise NotImplementedError("FlashSantacoderSharded is only available on GPU")