Finally it works
Browse filesIts working fine
@Walmart-the-bag
merge this PR.
app.py
CHANGED
@@ -13,9 +13,54 @@ subprocess.run(
|
|
13 |
env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
|
14 |
shell=True,
|
15 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
model_name = "microsoft/Phi-3-medium-128k-instruct"
|
17 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
18 |
-
model = AutoModelForCausalLM.from_pretrained(model_name, device_map='cuda', _attn_implementation="flash_attention_2",
|
19 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
20 |
|
21 |
class StopOnTokens(StoppingCriteria):
|
@@ -25,9 +70,9 @@ class StopOnTokens(StoppingCriteria):
|
|
25 |
if input_ids[0][-1] == stop_id:
|
26 |
return True
|
27 |
return False
|
28 |
-
|
29 |
@spaces.GPU(queue=False)
|
30 |
-
def predict(message, history, temperature, max_tokens,
|
31 |
history_transformer_format = history + [[message, ""]]
|
32 |
stop = StopOnTokens()
|
33 |
messages = "".join(["".join(["\n<|end|>\n<|user|>\n"+item[0], "\n<|end|>\n<|assistant|>\n"+item[1]]) for item in history_transformer_format])
|
@@ -39,7 +84,7 @@ def predict(message, history, temperature, max_tokens, top_p, top_k):
|
|
39 |
max_new_tokens=max_tokens,
|
40 |
do_sample=True,
|
41 |
top_p=top_p,
|
42 |
-
|
43 |
temperature=temperature,
|
44 |
stopping_criteria=StoppingCriteriaList([stop])
|
45 |
)
|
@@ -51,14 +96,120 @@ def predict(message, history, temperature, max_tokens, top_p, top_k):
|
|
51 |
partial_message += new_token
|
52 |
yield partial_message
|
53 |
|
54 |
-
|
55 |
-
|
56 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
additional_inputs=[
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
demo.launch(share=True)
|
|
|
13 |
env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
|
14 |
shell=True,
|
15 |
)
|
16 |
+
|
17 |
+
theme = gr.themes.Base(
|
18 |
+
font=[gr.themes.GoogleFont('Libre Franklin'), gr.themes.GoogleFont('Public Sans'), 'system-ui', 'sans-serif'],
|
19 |
+
)
|
20 |
+
|
21 |
+
model_name = "microsoft/Phi-3-medium-4k-instruct"
|
22 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
23 |
+
model = AutoModelForCausalLM.from_pretrained(model_name, device_map='cuda', torch_dtype=torch.float16, _attn_implementation="flash_attention_2", trust_remote_code=True)
|
24 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
25 |
+
|
26 |
+
class StopOnTokens(StoppingCriteria):
|
27 |
+
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
|
28 |
+
stop_ids = [29, 0]
|
29 |
+
for stop_id in stop_ids:
|
30 |
+
if input_ids[0][-1] == stop_id:
|
31 |
+
return True
|
32 |
+
return False
|
33 |
+
|
34 |
+
@spaces.GPU(queue=False)
|
35 |
+
def predict1(message, history, temperature1, max_tokens1, repetition_penalty1, top_p1):
|
36 |
+
history_transformer_format = history + [[message, ""]]
|
37 |
+
stop = StopOnTokens()
|
38 |
+
messages = "".join(["".join(["\n<|end|>\n<|user|>\n"+item[0], "\n<|end|>\n<|assistant|>\n"+item[1]]) for item in history_transformer_format])
|
39 |
+
model_inputs = tokenizer([messages], return_tensors="pt").to("cuda")
|
40 |
+
streamer = TextIteratorStreamer(tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True)
|
41 |
+
generate_kwargs = dict(
|
42 |
+
model_inputs,
|
43 |
+
streamer=streamer,
|
44 |
+
max_new_tokens=max_tokens1,
|
45 |
+
do_sample=True,
|
46 |
+
top_p=top_p1,
|
47 |
+
repetition_penalty=repetition_penalty1,
|
48 |
+
temperature=temperature1,
|
49 |
+
stopping_criteria=StoppingCriteriaList([stop])
|
50 |
+
)
|
51 |
+
t = Thread(target=model.generate, kwargs=generate_kwargs)
|
52 |
+
t.start()
|
53 |
+
partial_message = ""
|
54 |
+
for new_token in streamer:
|
55 |
+
if new_token != '<':
|
56 |
+
partial_message += new_token
|
57 |
+
yield partial_message
|
58 |
+
|
59 |
+
|
60 |
+
|
61 |
model_name = "microsoft/Phi-3-medium-128k-instruct"
|
62 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
63 |
+
model = AutoModelForCausalLM.from_pretrained(model_name, device_map='cuda', torch_dtype=torch.float16, _attn_implementation="flash_attention_2", trust_remote_code=True)
|
64 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
65 |
|
66 |
class StopOnTokens(StoppingCriteria):
|
|
|
70 |
if input_ids[0][-1] == stop_id:
|
71 |
return True
|
72 |
return False
|
73 |
+
|
74 |
@spaces.GPU(queue=False)
|
75 |
+
def predict(message, history, temperature, max_tokens, repetition_penalty, top_p):
|
76 |
history_transformer_format = history + [[message, ""]]
|
77 |
stop = StopOnTokens()
|
78 |
messages = "".join(["".join(["\n<|end|>\n<|user|>\n"+item[0], "\n<|end|>\n<|assistant|>\n"+item[1]]) for item in history_transformer_format])
|
|
|
84 |
max_new_tokens=max_tokens,
|
85 |
do_sample=True,
|
86 |
top_p=top_p,
|
87 |
+
repetition_penalty=repetition_penalty,
|
88 |
temperature=temperature,
|
89 |
stopping_criteria=StoppingCriteriaList([stop])
|
90 |
)
|
|
|
96 |
partial_message += new_token
|
97 |
yield partial_message
|
98 |
|
99 |
+
max_tokens1 = gr.Slider(
|
100 |
+
minimum=512,
|
101 |
+
maximum=4096,
|
102 |
+
value=4000,
|
103 |
+
step=32,
|
104 |
+
interactive=True,
|
105 |
+
label="Maximum number of new tokens to generate",
|
106 |
+
)
|
107 |
+
repetition_penalty1 = gr.Slider(
|
108 |
+
minimum=0.01,
|
109 |
+
maximum=5.0,
|
110 |
+
value=1,
|
111 |
+
step=0.01,
|
112 |
+
interactive=True,
|
113 |
+
label="Repetition penalty",
|
114 |
+
)
|
115 |
+
temperature1 = gr.Slider(
|
116 |
+
minimum=0.0,
|
117 |
+
maximum=1.0,
|
118 |
+
value=0.7,
|
119 |
+
step=0.05,
|
120 |
+
visible=True,
|
121 |
+
interactive=True,
|
122 |
+
label="Temperature",
|
123 |
+
)
|
124 |
+
top_p1 = gr.Slider(
|
125 |
+
minimum=0.01,
|
126 |
+
maximum=0.99,
|
127 |
+
value=0.9,
|
128 |
+
step=0.01,
|
129 |
+
visible=True,
|
130 |
+
interactive=True,
|
131 |
+
label="Top P",
|
132 |
+
)
|
133 |
+
|
134 |
+
chatbot1 = gr.Chatbot(
|
135 |
+
label="Phi3-medium-4k",
|
136 |
+
show_copy_button=True,
|
137 |
+
likeable=True,
|
138 |
+
layout="panel"
|
139 |
+
)
|
140 |
+
|
141 |
+
output=gr.Textbox(label="Prompt")
|
142 |
+
|
143 |
+
with gr.Blocks() as min:
|
144 |
+
gr.ChatInterface(
|
145 |
+
fn=predict1,
|
146 |
+
chatbot=chatbot1,
|
147 |
additional_inputs=[
|
148 |
+
temperature1,
|
149 |
+
max_tokens1,
|
150 |
+
repetition_penalty1,
|
151 |
+
top_p1,
|
152 |
+
],
|
153 |
+
)
|
154 |
+
|
155 |
+
max_tokens = gr.Slider(
|
156 |
+
minimum=64000,
|
157 |
+
maximum=128000,
|
158 |
+
value=100000,
|
159 |
+
step=1000,
|
160 |
+
interactive=True,
|
161 |
+
label="Maximum number of new tokens to generate",
|
162 |
+
)
|
163 |
+
repetition_penalty = gr.Slider(
|
164 |
+
minimum=0.01,
|
165 |
+
maximum=5.0,
|
166 |
+
value=1,
|
167 |
+
step=0.01,
|
168 |
+
interactive=True,
|
169 |
+
label="Repetition penalty",
|
170 |
+
)
|
171 |
+
temperature = gr.Slider(
|
172 |
+
minimum=0.0,
|
173 |
+
maximum=1.0,
|
174 |
+
value=0.7,
|
175 |
+
step=0.05,
|
176 |
+
visible=True,
|
177 |
+
interactive=True,
|
178 |
+
label="Temperature",
|
179 |
+
)
|
180 |
+
top_p = gr.Slider(
|
181 |
+
minimum=0.01,
|
182 |
+
maximum=0.99,
|
183 |
+
value=0.9,
|
184 |
+
step=0.01,
|
185 |
+
visible=True,
|
186 |
+
interactive=True,
|
187 |
+
label="Top P",
|
188 |
)
|
189 |
+
|
190 |
+
chatbot = gr.Chatbot(
|
191 |
+
label="Phi3-medium-128k",
|
192 |
+
show_copy_button=True,
|
193 |
+
likeable=True,
|
194 |
+
layout="panel"
|
195 |
+
)
|
196 |
+
|
197 |
+
output=gr.Textbox(label="Prompt")
|
198 |
+
|
199 |
+
with gr.Blocks() as max:
|
200 |
+
gr.ChatInterface(
|
201 |
+
fn=predict,
|
202 |
+
chatbot=chatbot,
|
203 |
+
additional_inputs=[
|
204 |
+
temperature,
|
205 |
+
max_tokens,
|
206 |
+
repetition_penalty,
|
207 |
+
top_p,
|
208 |
+
],
|
209 |
+
)
|
210 |
+
|
211 |
+
with gr.Blocks(title="Phi 3 Medium DEMO", theme=theme) as demo:
|
212 |
+
gr.Markdown("# Phi3 Medium all in one")
|
213 |
+
gr.TabbedInterface([max, min], ['Phi3 medium 128k','Phi3 medium 4k'])
|
214 |
+
|
215 |
demo.launch(share=True)
|