Update app.py
Browse files
app.py
CHANGED
@@ -81,8 +81,8 @@ model = AutoModelForCausalLM.from_pretrained(
|
|
81 |
torch_dtype="auto",
|
82 |
attn_implementation="flash_attention_2",
|
83 |
quantization_config=quantization_config,
|
84 |
-
low_cpu_mem_usage=True,
|
85 |
-
device_map="auto"
|
86 |
).to(device)
|
87 |
model.eval()
|
88 |
|
|
|
81 |
torch_dtype="auto",
|
82 |
attn_implementation="flash_attention_2",
|
83 |
quantization_config=quantization_config,
|
84 |
+
# low_cpu_mem_usage=True,
|
85 |
+
# device_map="auto"
|
86 |
).to(device)
|
87 |
model.eval()
|
88 |
|