Update app.py
Browse files
app.py
CHANGED
@@ -4,17 +4,34 @@ import torch
|
|
4 |
import time
|
5 |
import spaces
|
6 |
|
7 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
@spaces.GPU
|
9 |
-
def
|
10 |
-
"""
|
11 |
-
|
|
|
|
|
12 |
start_time = time.time()
|
13 |
|
14 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
|
|
|
|
|
|
15 |
model = AutoModelForCausalLM.from_pretrained(
|
16 |
model_id,
|
17 |
-
torch_dtype=torch.float16,
|
18 |
device_map="auto",
|
19 |
trust_remote_code=True
|
20 |
)
|
@@ -22,15 +39,25 @@ def load_model_gpu(model_id):
|
|
22 |
load_time = time.time() - start_time
|
23 |
print(f"✅ Model loaded in {load_time:.2f}s")
|
24 |
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
|
33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
with torch.no_grad():
|
35 |
outputs = model.generate(
|
36 |
**inputs,
|
@@ -42,26 +69,15 @@ def generate_response(model, tokenizer, prompt, max_length=512, temperature=0.7)
|
|
42 |
eos_token_id=tokenizer.eos_token_id
|
43 |
)
|
44 |
|
45 |
-
generation_time = time.time() -
|
46 |
-
|
|
|
|
|
47 |
outputs[0][inputs['input_ids'].shape[-1]:],
|
48 |
skip_special_tokens=True
|
49 |
).strip()
|
50 |
|
51 |
-
return
|
52 |
-
|
53 |
-
# Model configurations
|
54 |
-
MODELS = {
|
55 |
-
"Athena-R3X 8B": "Spestly/Athena-R3X-8B",
|
56 |
-
"Athena-R3X 4B": "Spestly/Athena-R3X-4B",
|
57 |
-
"Athena-R3 7B": "Spestly/Athena-R3-7B",
|
58 |
-
"Athena-3 3B": "Spestly/Athena-3-3B",
|
59 |
-
"Athena-3 7B": "Spestly/Athena-3-7B",
|
60 |
-
"Athena-3 14B": "Spestly/Athena-3-14B",
|
61 |
-
"Athena-2 1.5B": "Spestly/Athena-2-1.5B",
|
62 |
-
"Athena-1 3B": "Spestly/Athena-1-3B",
|
63 |
-
"Athena-1 7B": "Spestly/Athena-1-7B"
|
64 |
-
}
|
65 |
|
66 |
def chatbot(conversation, user_message, model_name, max_length=512, temperature=0.7):
|
67 |
if not user_message.strip():
|
@@ -74,27 +90,18 @@ def chatbot(conversation, user_message, model_name, max_length=512, temperature=
|
|
74 |
model_id = MODELS.get(model_name, MODELS["Athena-R3X 8B"])
|
75 |
|
76 |
try:
|
77 |
-
#
|
78 |
-
model, tokenizer = load_model_gpu(model_id)
|
79 |
-
|
80 |
-
# Append user message to conversation
|
81 |
conversation.append([user_message, ""])
|
82 |
|
83 |
-
# Build prompt from conversation history
|
84 |
-
prompt = ""
|
85 |
-
for user_msg, assistant_msg in conversation[:-1]: # Exclude the current message
|
86 |
-
prompt += f"User: {user_msg}\nAthena: {assistant_msg}\n"
|
87 |
-
prompt += f"User: {user_message}\nAthena:"
|
88 |
-
|
89 |
# Generate response using ZeroGPU
|
90 |
-
|
91 |
-
|
92 |
)
|
93 |
|
94 |
-
# Update the
|
95 |
-
conversation[-1][1] =
|
96 |
|
97 |
-
stats = f"
|
98 |
|
99 |
return conversation, "", stats
|
100 |
|
@@ -104,7 +111,7 @@ def chatbot(conversation, user_message, model_name, max_length=512, temperature=
|
|
104 |
conversation[-1][1] = error_msg
|
105 |
else:
|
106 |
conversation = [[user_message, error_msg]]
|
107 |
-
return conversation, "", f"❌ Error
|
108 |
|
109 |
def clear_chat():
|
110 |
return [], "", ""
|
|
|
4 |
import time
|
5 |
import spaces
|
6 |
|
7 |
+
# Model configurations
|
8 |
+
MODELS = {
|
9 |
+
"Athena-R3X 8B": "Spestly/Athena-R3X-8B",
|
10 |
+
"Athena-R3X 4B": "Spestly/Athena-R3X-4B",
|
11 |
+
"Athena-R3 7B": "Spestly/Athena-R3-7B",
|
12 |
+
"Athena-3 3B": "Spestly/Athena-3-3B",
|
13 |
+
"Athena-3 7B": "Spestly/Athena-3-7B",
|
14 |
+
"Athena-3 14B": "Spestly/Athena-3-14B",
|
15 |
+
"Athena-2 1.5B": "Spestly/Athena-2-1.5B",
|
16 |
+
"Athena-1 3B": "Spestly/Athena-1-3B",
|
17 |
+
"Athena-1 7B": "Spestly/Athena-1-7B"
|
18 |
+
}
|
19 |
+
|
20 |
@spaces.GPU
|
21 |
+
def generate_response(model_id, conversation, user_message, max_length=512, temperature=0.7):
|
22 |
+
"""Generate response using ZeroGPU - all CUDA operations happen here"""
|
23 |
+
|
24 |
+
# Load model and tokenizer inside the GPU function
|
25 |
+
print(f"🚀 Loading {model_id}...")
|
26 |
start_time = time.time()
|
27 |
|
28 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
29 |
+
if tokenizer.pad_token is None:
|
30 |
+
tokenizer.pad_token = tokenizer.eos_token
|
31 |
+
|
32 |
model = AutoModelForCausalLM.from_pretrained(
|
33 |
model_id,
|
34 |
+
torch_dtype=torch.float16,
|
35 |
device_map="auto",
|
36 |
trust_remote_code=True
|
37 |
)
|
|
|
39 |
load_time = time.time() - start_time
|
40 |
print(f"✅ Model loaded in {load_time:.2f}s")
|
41 |
|
42 |
+
# Build conversation history
|
43 |
+
conversation_history = []
|
44 |
+
for user_msg, assistant_msg in conversation:
|
45 |
+
if user_msg:
|
46 |
+
conversation_history.append(f"User: {user_msg}")
|
47 |
+
if assistant_msg:
|
48 |
+
conversation_history.append(f"Athena: {assistant_msg}")
|
49 |
|
50 |
+
# Add current user message
|
51 |
+
conversation_history.append(f"User: {user_message}")
|
52 |
+
conversation_history.append("Athena:")
|
53 |
+
|
54 |
+
# Create prompt
|
55 |
+
prompt = "\n".join(conversation_history)
|
56 |
+
|
57 |
+
# Tokenize and generate
|
58 |
+
inputs = tokenizer(prompt, return_tensors="pt")
|
59 |
+
|
60 |
+
generation_start = time.time()
|
61 |
with torch.no_grad():
|
62 |
outputs = model.generate(
|
63 |
**inputs,
|
|
|
69 |
eos_token_id=tokenizer.eos_token_id
|
70 |
)
|
71 |
|
72 |
+
generation_time = time.time() - generation_start
|
73 |
+
|
74 |
+
# Decode response
|
75 |
+
response = tokenizer.decode(
|
76 |
outputs[0][inputs['input_ids'].shape[-1]:],
|
77 |
skip_special_tokens=True
|
78 |
).strip()
|
79 |
|
80 |
+
return response, load_time, generation_time
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
def chatbot(conversation, user_message, model_name, max_length=512, temperature=0.7):
|
83 |
if not user_message.strip():
|
|
|
90 |
model_id = MODELS.get(model_name, MODELS["Athena-R3X 8B"])
|
91 |
|
92 |
try:
|
93 |
+
# Add user message to conversation
|
|
|
|
|
|
|
94 |
conversation.append([user_message, ""])
|
95 |
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
# Generate response using ZeroGPU
|
97 |
+
response, load_time, generation_time = generate_response(
|
98 |
+
model_id, conversation[:-1], user_message, max_length, temperature
|
99 |
)
|
100 |
|
101 |
+
# Update the conversation with the response
|
102 |
+
conversation[-1][1] = response
|
103 |
|
104 |
+
stats = f"��� Load: {load_time:.1f}s | Gen: {generation_time:.1f}s | Model: {model_name}"
|
105 |
|
106 |
return conversation, "", stats
|
107 |
|
|
|
111 |
conversation[-1][1] = error_msg
|
112 |
else:
|
113 |
conversation = [[user_message, error_msg]]
|
114 |
+
return conversation, "", f"❌ Error: {str(e)}"
|
115 |
|
116 |
def clear_chat():
|
117 |
return [], "", ""
|