Spaces:
Sleeping
Sleeping
Commit
·
d36dc81
1
Parent(s):
ad04391
updated requirement
Browse files
model.py
CHANGED
@@ -10,16 +10,27 @@ HF_TOKEN = os.getenv("HUGGINGFACE_TOKEN")
|
|
10 |
if not HF_TOKEN:
|
11 |
raise ValueError("Missing Hugging Face token. Set HUGGINGFACE_TOKEN as an environment variable.")
|
12 |
|
13 |
-
#
|
|
|
|
|
|
|
14 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, token=HF_TOKEN)
|
15 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
|
17 |
def generate_code(prompt: str, max_tokens: int = 256):
|
18 |
"""Generates code based on the input prompt."""
|
19 |
if not prompt.strip():
|
20 |
return "Error: Empty prompt provided."
|
21 |
|
22 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
23 |
inputs = tokenizer(prompt, return_tensors="pt").to(device)
|
24 |
output = model.generate(**inputs, max_new_tokens=max_tokens)
|
25 |
return tokenizer.decode(output[0], skip_special_tokens=True)
|
|
|
10 |
if not HF_TOKEN:
|
11 |
raise ValueError("Missing Hugging Face token. Set HUGGINGFACE_TOKEN as an environment variable.")
|
12 |
|
13 |
+
# Set device
|
14 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
15 |
+
|
16 |
+
# Load tokenizer with authentication
|
17 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, token=HF_TOKEN)
|
18 |
+
|
19 |
+
# Load model with optimizations
|
20 |
+
model = AutoModelForCausalLM.from_pretrained(
|
21 |
+
MODEL_NAME,
|
22 |
+
token=HF_TOKEN,
|
23 |
+
torch_dtype=torch.float16, # Reduce memory usage
|
24 |
+
low_cpu_mem_usage=True, # Optimize loading
|
25 |
+
device_map="auto", # Automatic device placement
|
26 |
+
offload_folder="offload" # Offload to disk if needed
|
27 |
+
).to(device)
|
28 |
|
29 |
def generate_code(prompt: str, max_tokens: int = 256):
|
30 |
"""Generates code based on the input prompt."""
|
31 |
if not prompt.strip():
|
32 |
return "Error: Empty prompt provided."
|
33 |
|
|
|
34 |
inputs = tokenizer(prompt, return_tensors="pt").to(device)
|
35 |
output = model.generate(**inputs, max_new_tokens=max_tokens)
|
36 |
return tokenizer.decode(output[0], skip_special_tokens=True)
|