Upload app.py
Browse files
app.py
CHANGED
@@ -11,12 +11,16 @@ from transformers import GPT2Tokenizer
|
|
11 |
import spaces
|
12 |
import os
|
13 |
from pathlib import Path
|
|
|
14 |
|
15 |
# Local imports
|
16 |
from smollmv2 import SmollmV2
|
17 |
from config import SmollmConfig, DataConfig
|
18 |
from smollv2_lightning import LitSmollmv2
|
19 |
|
|
|
|
|
|
|
20 |
|
21 |
def combine_model_parts(model_dir="split_models", output_file="checkpoints/last.ckpt"):
|
22 |
"""
|
@@ -56,7 +60,7 @@ def load_model():
|
|
56 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
57 |
|
58 |
# Load model directly from checkpoint
|
59 |
-
checkpoint_path = "last.ckpt"
|
60 |
|
61 |
if not os.path.exists(checkpoint_path):
|
62 |
raise FileNotFoundError(
|
@@ -64,21 +68,25 @@ def load_model():
|
|
64 |
"Please ensure the model checkpoint file 'last.ckpt' is present in the root directory."
|
65 |
)
|
66 |
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
|
|
|
|
|
|
80 |
|
81 |
-
|
|
|
82 |
|
83 |
|
84 |
@spaces.GPU(enable_queue=True)
|
@@ -86,50 +94,59 @@ def generate_text(prompt, num_tokens, temperature=0.8, top_p=0.9):
|
|
86 |
"""
|
87 |
Generate text using the SmollmV2 model.
|
88 |
"""
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
# Tokenize input prompt
|
93 |
-
input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device)
|
94 |
-
|
95 |
-
# Generate tokens one at a time
|
96 |
-
for _ in range(num_tokens):
|
97 |
-
# Get the model's predictions
|
98 |
-
with torch.no_grad():
|
99 |
-
with torch.autocast(device_type=device, dtype=torch.bfloat16):
|
100 |
-
logits, _ = model.model(input_ids)
|
101 |
-
|
102 |
-
# Get the next token probabilities
|
103 |
-
logits = logits[:, -1, :] / temperature
|
104 |
-
probs = F.softmax(logits, dim=-1)
|
105 |
-
|
106 |
-
# Apply top-p sampling
|
107 |
-
if top_p > 0:
|
108 |
-
sorted_probs, sorted_indices = torch.sort(probs, descending=True)
|
109 |
-
cumsum_probs = torch.cumsum(sorted_probs, dim=-1)
|
110 |
-
sorted_indices_to_keep = cumsum_probs <= top_p
|
111 |
-
sorted_indices_to_keep[..., 1:] = sorted_indices_to_keep[..., :-1].clone()
|
112 |
-
sorted_indices_to_keep[..., 0] = 1
|
113 |
-
indices_to_keep = torch.zeros_like(probs, dtype=torch.bool).scatter_(-1, sorted_indices, sorted_indices_to_keep)
|
114 |
-
probs = torch.where(indices_to_keep, probs, torch.zeros_like(probs))
|
115 |
-
probs = probs / probs.sum(dim=-1, keepdim=True)
|
116 |
|
117 |
-
#
|
118 |
-
|
119 |
|
120 |
-
#
|
121 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
122 |
|
123 |
-
#
|
124 |
-
|
125 |
-
|
126 |
|
127 |
-
|
128 |
-
|
129 |
-
return generated_text
|
130 |
|
131 |
# Load the model globally
|
132 |
-
|
|
|
|
|
|
|
|
|
133 |
|
134 |
# Create the Gradio interface
|
135 |
demo = gr.Interface(
|
|
|
11 |
import spaces
|
12 |
import os
|
13 |
from pathlib import Path
|
14 |
+
import warnings
|
15 |
|
16 |
# Local imports
|
17 |
from smollmv2 import SmollmV2
|
18 |
from config import SmollmConfig, DataConfig
|
19 |
from smollv2_lightning import LitSmollmv2
|
20 |
|
21 |
+
# Configure PyTorch to handle the device properties issue
|
22 |
+
torch._dynamo.config.suppress_errors = True
|
23 |
+
warnings.filterwarnings('ignore', category=UserWarning)
|
24 |
|
25 |
def combine_model_parts(model_dir="split_models", output_file="checkpoints/last.ckpt"):
|
26 |
"""
|
|
|
60 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
61 |
|
62 |
# Load model directly from checkpoint
|
63 |
+
checkpoint_path = "last.ckpt"
|
64 |
|
65 |
if not os.path.exists(checkpoint_path):
|
66 |
raise FileNotFoundError(
|
|
|
68 |
"Please ensure the model checkpoint file 'last.ckpt' is present in the root directory."
|
69 |
)
|
70 |
|
71 |
+
try:
|
72 |
+
# Load the model from checkpoint using Lightning module
|
73 |
+
model = LitSmollmv2.load_from_checkpoint(
|
74 |
+
checkpoint_path,
|
75 |
+
model_config=SmollmConfig,
|
76 |
+
strict=False
|
77 |
+
)
|
78 |
+
|
79 |
+
model.to(device)
|
80 |
+
model.eval()
|
81 |
+
|
82 |
+
# Initialize tokenizer
|
83 |
+
tokenizer = GPT2Tokenizer.from_pretrained(DataConfig.tokenizer_path)
|
84 |
+
tokenizer.pad_token = tokenizer.eos_token
|
85 |
+
|
86 |
+
return model, tokenizer, device
|
87 |
|
88 |
+
except Exception as e:
|
89 |
+
raise RuntimeError(f"Error loading model: {str(e)}")
|
90 |
|
91 |
|
92 |
@spaces.GPU(enable_queue=True)
|
|
|
94 |
"""
|
95 |
Generate text using the SmollmV2 model.
|
96 |
"""
|
97 |
+
try:
|
98 |
+
# Ensure num_tokens doesn't exceed model's block size
|
99 |
+
num_tokens = min(num_tokens, SmollmConfig.block_size)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
100 |
|
101 |
+
# Tokenize input prompt
|
102 |
+
input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device)
|
103 |
|
104 |
+
# Generate tokens one at a time
|
105 |
+
with torch.inference_mode(): # Use inference_mode instead of no_grad
|
106 |
+
for _ in range(num_tokens):
|
107 |
+
# Get the model's predictions
|
108 |
+
with torch.autocast(device_type=device, dtype=torch.float16): # Changed to float16
|
109 |
+
outputs = model(input_ids)
|
110 |
+
logits = outputs[0] if isinstance(outputs, tuple) else outputs
|
111 |
+
|
112 |
+
# Get the next token probabilities
|
113 |
+
logits = logits[:, -1, :] / temperature
|
114 |
+
probs = F.softmax(logits, dim=-1)
|
115 |
+
|
116 |
+
# Apply top-p sampling
|
117 |
+
if top_p > 0:
|
118 |
+
sorted_probs, sorted_indices = torch.sort(probs, descending=True)
|
119 |
+
cumsum_probs = torch.cumsum(sorted_probs, dim=-1)
|
120 |
+
sorted_indices_to_keep = cumsum_probs <= top_p
|
121 |
+
sorted_indices_to_keep[..., 1:] = sorted_indices_to_keep[..., :-1].clone()
|
122 |
+
sorted_indices_to_keep[..., 0] = 1
|
123 |
+
indices_to_keep = torch.zeros_like(probs, dtype=torch.bool).scatter_(-1, sorted_indices, sorted_indices_to_keep)
|
124 |
+
probs = torch.where(indices_to_keep, probs, torch.zeros_like(probs))
|
125 |
+
probs = probs / probs.sum(dim=-1, keepdim=True)
|
126 |
+
|
127 |
+
# Sample next token
|
128 |
+
next_token = torch.multinomial(probs, num_samples=1)
|
129 |
+
|
130 |
+
# Append to input_ids
|
131 |
+
input_ids = torch.cat([input_ids, next_token], dim=-1)
|
132 |
+
|
133 |
+
# Stop if we generate an EOS token
|
134 |
+
if next_token.item() == tokenizer.eos_token_id:
|
135 |
+
break
|
136 |
|
137 |
+
# Decode and return the generated text
|
138 |
+
generated_text = tokenizer.decode(input_ids[0], skip_special_tokens=True)
|
139 |
+
return generated_text
|
140 |
|
141 |
+
except Exception as e:
|
142 |
+
return f"Error during text generation: {str(e)}"
|
|
|
143 |
|
144 |
# Load the model globally
|
145 |
+
try:
|
146 |
+
model, tokenizer, device = load_model()
|
147 |
+
except Exception as e:
|
148 |
+
print(f"Error initializing model: {str(e)}")
|
149 |
+
raise
|
150 |
|
151 |
# Create the Gradio interface
|
152 |
demo = gr.Interface(
|