Spaces:
Runtime error
Runtime error
Commit
·
c2d0dc7
1
Parent(s):
59e3ffd
fixing app.py
Browse files
app.py
CHANGED
@@ -2,27 +2,33 @@ import os
|
|
2 |
import torch
|
3 |
from fastapi import FastAPI, HTTPException
|
4 |
from pydantic import BaseModel
|
5 |
-
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
|
6 |
|
7 |
-
# Set
|
8 |
os.environ["HF_HOME"] = "/tmp/huggingface"
|
9 |
os.environ["TRANSFORMERS_CACHE"] = "/tmp/huggingface"
|
10 |
|
11 |
# Model setup
|
12 |
MODEL_NAME = "deepseek-ai/deepseek-llm-7b-base"
|
13 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
14 |
-
DTYPE = torch.float16 if DEVICE == "cuda" else torch.bfloat16
|
15 |
|
16 |
-
# Load model
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
18 |
model = AutoModelForCausalLM.from_pretrained(
|
19 |
-
MODEL_NAME,
|
|
|
|
|
|
|
20 |
)
|
21 |
|
22 |
-
#
|
23 |
-
|
24 |
-
generation_config.pad_token_id = generation_config.eos_token_id
|
25 |
-
generation_config.use_cache = True # Speed up decoding
|
26 |
|
27 |
# FastAPI app
|
28 |
app = FastAPI()
|
@@ -30,28 +36,26 @@ app = FastAPI()
|
|
30 |
# Request payload
|
31 |
class TextGenerationRequest(BaseModel):
|
32 |
prompt: str
|
33 |
-
max_tokens: int = 512 # Default to 512
|
34 |
|
35 |
@app.post("/generate")
|
36 |
async def generate_text(request: TextGenerationRequest):
|
37 |
try:
|
38 |
-
# Tokenize input and move tensors to the correct device
|
39 |
inputs = tokenizer(request.prompt, return_tensors="pt", padding=True, truncation=True).to(DEVICE)
|
40 |
|
41 |
-
# Use no_grad() for faster inference
|
42 |
with torch.no_grad():
|
43 |
outputs = model.generate(
|
44 |
**inputs,
|
45 |
max_new_tokens=request.max_tokens,
|
46 |
-
do_sample=True,
|
47 |
-
temperature=0.7,
|
48 |
-
top_k=50,
|
49 |
-
top_p=0.9,
|
50 |
-
repetition_penalty=1.
|
|
|
51 |
)
|
52 |
|
53 |
-
|
54 |
-
result = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
|
55 |
return {"generated_text": result}
|
56 |
|
57 |
except Exception as e:
|
|
|
2 |
import torch
|
3 |
from fastapi import FastAPI, HTTPException
|
4 |
from pydantic import BaseModel
|
5 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig, BitsAndBytesConfig
|
6 |
|
7 |
+
# Set cache directory
|
8 |
os.environ["HF_HOME"] = "/tmp/huggingface"
|
9 |
os.environ["TRANSFORMERS_CACHE"] = "/tmp/huggingface"
|
10 |
|
11 |
# Model setup
|
12 |
MODEL_NAME = "deepseek-ai/deepseek-llm-7b-base"
|
13 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
14 |
|
15 |
+
# Load 4-bit quantized model (for speed & efficiency)
|
16 |
+
bnb_config = BitsAndBytesConfig(
|
17 |
+
load_in_4bit=True, # Enable 4-bit inference
|
18 |
+
bnb_4bit_compute_dtype=torch.float16,
|
19 |
+
bnb_4bit_use_double_quant=True,
|
20 |
+
)
|
21 |
+
|
22 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
23 |
model = AutoModelForCausalLM.from_pretrained(
|
24 |
+
MODEL_NAME,
|
25 |
+
quantization_config=bnb_config,
|
26 |
+
device_map="auto",
|
27 |
+
attn_implementation="flash_attention_2" # Enables Flash Attention
|
28 |
)
|
29 |
|
30 |
+
# Compile for even faster inference (PyTorch 2.0+)
|
31 |
+
model = torch.compile(model)
|
|
|
|
|
32 |
|
33 |
# FastAPI app
|
34 |
app = FastAPI()
|
|
|
36 |
# Request payload
|
37 |
class TextGenerationRequest(BaseModel):
|
38 |
prompt: str
|
39 |
+
max_tokens: int = 512 # Default to 512
|
40 |
|
41 |
@app.post("/generate")
|
42 |
async def generate_text(request: TextGenerationRequest):
|
43 |
try:
|
|
|
44 |
inputs = tokenizer(request.prompt, return_tensors="pt", padding=True, truncation=True).to(DEVICE)
|
45 |
|
|
|
46 |
with torch.no_grad():
|
47 |
outputs = model.generate(
|
48 |
**inputs,
|
49 |
max_new_tokens=request.max_tokens,
|
50 |
+
do_sample=True,
|
51 |
+
temperature=0.7,
|
52 |
+
top_k=50,
|
53 |
+
top_p=0.9,
|
54 |
+
repetition_penalty=1.05,
|
55 |
+
use_cache=True,
|
56 |
)
|
57 |
|
58 |
+
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
|
|
59 |
return {"generated_text": result}
|
60 |
|
61 |
except Exception as e:
|