Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -51,11 +51,11 @@ start_messages = [
|
|
51 |
|
52 |
|
53 |
# functions
|
54 |
-
@dataclass
|
55 |
-
class PastKV:
|
56 |
-
|
57 |
|
58 |
-
past_key_values = PastKV()
|
59 |
|
60 |
|
61 |
def past_kv_to_device(past_kv, device):
|
@@ -63,7 +63,7 @@ def past_kv_to_device(past_kv, device):
|
|
63 |
|
64 |
|
65 |
@spaces.GPU
|
66 |
-
def set_past_key_values(
|
67 |
model, tokenizer = pipe.model, pipe.tokenizer
|
68 |
tokenized = tokenizer.apply_chat_template(start_messages, return_tensors='pt')
|
69 |
|
@@ -71,26 +71,25 @@ def set_past_key_values(past_key_values):
|
|
71 |
test_messages = [*start_messages, {'role': 'user', 'content': 'Hello World!'}]
|
72 |
tokenized_test = tokenizer.apply_chat_template(test_messages, return_tensors='pt')
|
73 |
assert (tokenized_test[:, :tokenized.shape[1]] == tokenized).all().cpu().item()
|
74 |
-
|
75 |
-
return True
|
76 |
|
77 |
|
78 |
@spaces.GPU
|
79 |
-
def generate(text):
|
80 |
messages = [
|
81 |
*start_messages,
|
82 |
{'role': 'user', 'content': text}
|
83 |
]
|
84 |
response = pipe(messages,
|
85 |
-
past_key_values=past_key_values
|
86 |
**generate_kwargs)[0]['generated_text']
|
87 |
return response[-1]['content']
|
88 |
|
89 |
|
90 |
if __name__ == "__main__":
|
91 |
with torch.no_grad():
|
92 |
-
set_past_key_values(
|
93 |
print(f'{past_key_values.past_key_values=}')
|
94 |
demo = gr.Interface(generate,
|
95 |
-
inputs="textbox", outputs="textbox")
|
96 |
demo.launch()
|
|
|
51 |
|
52 |
|
53 |
# functions
|
54 |
+
# @dataclass
|
55 |
+
# class PastKV:
|
56 |
+
# past_key_values: Any = None
|
57 |
|
58 |
+
# past_key_values = PastKV()
|
59 |
|
60 |
|
61 |
def past_kv_to_device(past_kv, device):
|
|
|
63 |
|
64 |
|
65 |
@spaces.GPU
|
66 |
+
def set_past_key_values():
|
67 |
model, tokenizer = pipe.model, pipe.tokenizer
|
68 |
tokenized = tokenizer.apply_chat_template(start_messages, return_tensors='pt')
|
69 |
|
|
|
71 |
test_messages = [*start_messages, {'role': 'user', 'content': 'Hello World!'}]
|
72 |
tokenized_test = tokenizer.apply_chat_template(test_messages, return_tensors='pt')
|
73 |
assert (tokenized_test[:, :tokenized.shape[1]] == tokenized).all().cpu().item()
|
74 |
+
return gr.State(model(tokenized.to(model.device)).past_key_values)
|
|
|
75 |
|
76 |
|
77 |
@spaces.GPU
|
78 |
+
def generate(text, past_key_values):
|
79 |
messages = [
|
80 |
*start_messages,
|
81 |
{'role': 'user', 'content': text}
|
82 |
]
|
83 |
response = pipe(messages,
|
84 |
+
past_key_values=past_key_values, #past_kv_to_device(past_key_values, pipe.model.device),
|
85 |
**generate_kwargs)[0]['generated_text']
|
86 |
return response[-1]['content']
|
87 |
|
88 |
|
89 |
if __name__ == "__main__":
|
90 |
with torch.no_grad():
|
91 |
+
past_key_values = set_past_key_values()
|
92 |
print(f'{past_key_values.past_key_values=}')
|
93 |
demo = gr.Interface(generate,
|
94 |
+
inputs=["textbox", gr.State(past_key_values)], outputs="textbox")
|
95 |
demo.launch()
|