Use k-quant model
Browse files
app.py
CHANGED
@@ -19,7 +19,7 @@ def generate(history):
|
|
19 |
return streamer
|
20 |
|
21 |
|
22 |
-
llm = AutoModelForCausalLM.from_pretrained("theodotus/llama-uk", model_file="
|
23 |
end_token = "</s>"
|
24 |
|
25 |
|
|
|
19 |
return streamer
|
20 |
|
21 |
|
22 |
+
llm = AutoModelForCausalLM.from_pretrained("theodotus/llama-uk", model_file="model.bin", model_type='llama')
|
23 |
end_token = "</s>"
|
24 |
|
25 |
|