Update app.py
Browse files
app.py
CHANGED
@@ -1,14 +1,21 @@
|
|
1 |
import os
|
2 |
-
|
3 |
-
from fastapi
|
|
|
4 |
from pydantic import BaseModel, field_validator
|
5 |
-
from transformers import
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
import boto3
|
7 |
import uvicorn
|
8 |
-
import
|
9 |
-
import
|
10 |
-
from
|
11 |
-
import torch # Import torch
|
12 |
|
13 |
AWS_ACCESS_KEY_ID = os.getenv("AWS_ACCESS_KEY_ID")
|
14 |
AWS_SECRET_ACCESS_KEY = os.getenv("AWS_SECRET_ACCESS_KEY")
|
@@ -16,34 +23,24 @@ AWS_REGION = os.getenv("AWS_REGION")
|
|
16 |
S3_BUCKET_NAME = os.getenv("S3_BUCKET_NAME")
|
17 |
HUGGINGFACE_HUB_TOKEN = os.getenv("HUGGINGFACE_HUB_TOKEN")
|
18 |
|
19 |
-
if not all([AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION, S3_BUCKET_NAME]):
|
20 |
-
raise ValueError("Missing one or more AWS environment variables.")
|
21 |
-
|
22 |
s3_client = boto3.client('s3', aws_access_key_id=AWS_ACCESS_KEY_ID, aws_secret_access_key=AWS_SECRET_ACCESS_KEY, region_name=AWS_REGION)
|
23 |
|
24 |
app = FastAPI()
|
25 |
|
26 |
-
SPECIAL_TOKENS = {
|
27 |
-
"bos_token": "<|startoftext|>",
|
28 |
-
"eos_token": "<|endoftext|>",
|
29 |
-
"pad_token": "[PAD]",
|
30 |
-
"unk_token": "[UNK]",
|
31 |
-
}
|
32 |
-
|
33 |
class GenerateRequest(BaseModel):
|
34 |
model_name: str
|
35 |
input_text: str = ""
|
36 |
task_type: str
|
37 |
temperature: float = 1.0
|
38 |
-
max_new_tokens: int =
|
|
|
39 |
top_p: float = 1.0
|
40 |
top_k: int = 50
|
41 |
-
repetition_penalty: float = 1.
|
42 |
num_return_sequences: int = 1
|
43 |
do_sample: bool = True
|
|
|
44 |
stop_sequences: list[str] = []
|
45 |
-
no_repeat_ngram_size: int = 2
|
46 |
-
continuation_id: Optional[str] = None
|
47 |
|
48 |
@field_validator("model_name")
|
49 |
def model_name_cannot_be_empty(cls, v):
|
@@ -58,12 +55,6 @@ class GenerateRequest(BaseModel):
|
|
58 |
raise ValueError(f"task_type must be one of: {valid_types}")
|
59 |
return v
|
60 |
|
61 |
-
@field_validator("max_new_tokens")
|
62 |
-
def max_new_tokens_must_be_within_limit(cls, v):
|
63 |
-
if v > 500:
|
64 |
-
raise ValueError("max_new_tokens cannot be greater than 500.")
|
65 |
-
return v
|
66 |
-
|
67 |
class S3ModelLoader:
|
68 |
def __init__(self, bucket_name, s3_client):
|
69 |
self.bucket_name = bucket_name
|
@@ -75,166 +66,175 @@ class S3ModelLoader:
|
|
75 |
async def load_model_and_tokenizer(self, model_name):
|
76 |
s3_uri = self._get_s3_uri(model_name)
|
77 |
try:
|
78 |
-
config = AutoConfig.from_pretrained(s3_uri, local_files_only=
|
79 |
-
model = AutoModelForCausalLM.from_pretrained(s3_uri, config=config, local_files_only=
|
80 |
-
tokenizer = AutoTokenizer.from_pretrained(s3_uri, config=config, local_files_only=
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
return model, tokenizer
|
86 |
-
except
|
87 |
-
|
|
|
|
|
|
|
88 |
|
89 |
-
|
|
|
90 |
|
91 |
-
|
|
|
|
|
|
|
|
|
92 |
|
93 |
-
|
94 |
-
try:
|
95 |
-
return await model_loader.load_model_and_tokenizer(model_name)
|
96 |
-
except Exception as e:
|
97 |
-
raise HTTPException(status_code=500, detail=f"Error loading model: {e}")
|
98 |
|
99 |
@app.post("/generate")
|
100 |
-
async def generate(request: GenerateRequest
|
101 |
-
model, tokenizer = model_resources
|
102 |
try:
|
103 |
model_name = request.model_name
|
104 |
input_text = request.input_text
|
|
|
105 |
temperature = request.temperature
|
106 |
max_new_tokens = request.max_new_tokens
|
|
|
107 |
top_p = request.top_p
|
108 |
top_k = request.top_k
|
109 |
repetition_penalty = request.repetition_penalty
|
110 |
num_return_sequences = request.num_return_sequences
|
111 |
do_sample = request.do_sample
|
|
|
112 |
stop_sequences = request.stop_sequences
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
generation_config.pad_token_id = tokenizer.pad_token_id
|
134 |
-
|
135 |
-
generated_text = generate_text_internal(model, tokenizer, input_text, generation_config, stop_sequences)
|
136 |
-
|
137 |
-
new_continuation_id = continuation_id if continuation_id else os.urandom(16).hex()
|
138 |
-
active_generations[new_continuation_id] = {"model_name": model_name, "output": generated_text}
|
139 |
-
|
140 |
-
return JSONResponse({"text": generated_text, "continuation_id": new_continuation_id, "model_name": model_name})
|
141 |
-
|
142 |
-
except HTTPException as http_err:
|
143 |
-
raise http_err
|
144 |
except Exception as e:
|
145 |
raise HTTPException(status_code=500, detail=f"Internal server error: {str(e)}")
|
146 |
|
147 |
-
def
|
148 |
-
|
149 |
-
|
|
|
150 |
|
151 |
-
|
|
|
152 |
|
153 |
-
|
154 |
-
def __init__(self, stop_sequences, tokenizer):
|
155 |
-
self.stop_sequences = stop_sequences
|
156 |
-
self.tokenizer = tokenizer
|
157 |
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
if decoded_output.endswith(stop):
|
162 |
-
return True
|
163 |
-
return False
|
164 |
|
165 |
-
stopping_criteria
|
166 |
|
|
|
167 |
outputs = model.generate(
|
168 |
-
encoded_input
|
169 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
170 |
stopping_criteria=stopping_criteria,
|
171 |
-
|
|
|
172 |
)
|
173 |
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
183 |
|
184 |
@app.post("/generate-image")
|
185 |
async def generate_image(request: GenerateRequest):
|
186 |
try:
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
image_generator =
|
191 |
-
image = image_generator(
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
raise http_err
|
200 |
except Exception as e:
|
201 |
raise HTTPException(status_code=500, detail=f"Internal server error: {str(e)}")
|
202 |
|
203 |
@app.post("/generate-text-to-speech")
|
204 |
async def generate_text_to_speech(request: GenerateRequest):
|
205 |
try:
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
raise http_err
|
219 |
except Exception as e:
|
220 |
raise HTTPException(status_code=500, detail=f"Internal server error: {str(e)}")
|
221 |
|
222 |
@app.post("/generate-video")
|
223 |
async def generate_video(request: GenerateRequest):
|
224 |
try:
|
225 |
-
|
226 |
-
|
227 |
-
|
228 |
-
|
229 |
-
|
230 |
-
|
231 |
-
|
232 |
-
|
233 |
-
|
234 |
-
return
|
235 |
-
|
236 |
-
except HTTPException as http_err:
|
237 |
-
raise http_err
|
238 |
except Exception as e:
|
239 |
raise HTTPException(status_code=500, detail=f"Internal server error: {str(e)}")
|
240 |
|
|
|
1 |
import os
|
2 |
+
import torch
|
3 |
+
from fastapi import FastAPI, HTTPException
|
4 |
+
from fastapi.responses import StreamingResponse
|
5 |
from pydantic import BaseModel, field_validator
|
6 |
+
from transformers import (
|
7 |
+
AutoConfig,
|
8 |
+
pipeline,
|
9 |
+
AutoModelForCausalLM,
|
10 |
+
AutoTokenizer,
|
11 |
+
GenerationConfig,
|
12 |
+
StoppingCriteriaList
|
13 |
+
)
|
14 |
import boto3
|
15 |
import uvicorn
|
16 |
+
import asyncio
|
17 |
+
from io import BytesIO
|
18 |
+
from transformers import pipeline
|
|
|
19 |
|
20 |
AWS_ACCESS_KEY_ID = os.getenv("AWS_ACCESS_KEY_ID")
|
21 |
AWS_SECRET_ACCESS_KEY = os.getenv("AWS_SECRET_ACCESS_KEY")
|
|
|
23 |
S3_BUCKET_NAME = os.getenv("S3_BUCKET_NAME")
|
24 |
HUGGINGFACE_HUB_TOKEN = os.getenv("HUGGINGFACE_HUB_TOKEN")
|
25 |
|
|
|
|
|
|
|
26 |
s3_client = boto3.client('s3', aws_access_key_id=AWS_ACCESS_KEY_ID, aws_secret_access_key=AWS_SECRET_ACCESS_KEY, region_name=AWS_REGION)
|
27 |
|
28 |
app = FastAPI()
|
29 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
class GenerateRequest(BaseModel):
|
31 |
model_name: str
|
32 |
input_text: str = ""
|
33 |
task_type: str
|
34 |
temperature: float = 1.0
|
35 |
+
max_new_tokens: int = 200
|
36 |
+
stream: bool = True
|
37 |
top_p: float = 1.0
|
38 |
top_k: int = 50
|
39 |
+
repetition_penalty: float = 1.0
|
40 |
num_return_sequences: int = 1
|
41 |
do_sample: bool = True
|
42 |
+
chunk_delay: float = 0.0
|
43 |
stop_sequences: list[str] = []
|
|
|
|
|
44 |
|
45 |
@field_validator("model_name")
|
46 |
def model_name_cannot_be_empty(cls, v):
|
|
|
55 |
raise ValueError(f"task_type must be one of: {valid_types}")
|
56 |
return v
|
57 |
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
class S3ModelLoader:
|
59 |
def __init__(self, bucket_name, s3_client):
|
60 |
self.bucket_name = bucket_name
|
|
|
66 |
async def load_model_and_tokenizer(self, model_name):
|
67 |
s3_uri = self._get_s3_uri(model_name)
|
68 |
try:
|
69 |
+
config = AutoConfig.from_pretrained(s3_uri, local_files_only=True)
|
70 |
+
model = AutoModelForCausalLM.from_pretrained(s3_uri, config=config, local_files_only=True)
|
71 |
+
tokenizer = AutoTokenizer.from_pretrained(s3_uri, config=config, local_files_only=True)
|
72 |
+
|
73 |
+
if tokenizer.eos_token_id is not None and tokenizer.pad_token_id is None:
|
74 |
+
tokenizer.pad_token_id = config.pad_token_id or tokenizer.eos_token_id
|
75 |
+
|
76 |
return model, tokenizer
|
77 |
+
except EnvironmentError:
|
78 |
+
try:
|
79 |
+
config = AutoConfig.from_pretrained(model_name)
|
80 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name, config=config)
|
81 |
+
model = AutoModelForCausalLM.from_pretrained(model_name, config=config)
|
82 |
|
83 |
+
if tokenizer.eos_token_id is not None and tokenizer.pad_token_id is None:
|
84 |
+
tokenizer.pad_token_id = config.pad_token_id or tokenizer.eos_token_id
|
85 |
|
86 |
+
model.save_pretrained(s3_uri)
|
87 |
+
tokenizer.save_pretrained(s3_uri)
|
88 |
+
return model, tokenizer
|
89 |
+
except Exception as e:
|
90 |
+
raise HTTPException(status_code=500, detail=f"Error loading model: {e}")
|
91 |
|
92 |
+
model_loader = S3ModelLoader(S3_BUCKET_NAME, s3_client)
|
|
|
|
|
|
|
|
|
93 |
|
94 |
@app.post("/generate")
|
95 |
+
async def generate(request: GenerateRequest):
|
|
|
96 |
try:
|
97 |
model_name = request.model_name
|
98 |
input_text = request.input_text
|
99 |
+
task_type = request.task_type
|
100 |
temperature = request.temperature
|
101 |
max_new_tokens = request.max_new_tokens
|
102 |
+
stream = request.stream
|
103 |
top_p = request.top_p
|
104 |
top_k = request.top_k
|
105 |
repetition_penalty = request.repetition_penalty
|
106 |
num_return_sequences = request.num_return_sequences
|
107 |
do_sample = request.do_sample
|
108 |
+
chunk_delay = request.chunk_delay
|
109 |
stop_sequences = request.stop_sequences
|
110 |
+
|
111 |
+
model, tokenizer = await model_loader.load_model_and_tokenizer(model_name)
|
112 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
113 |
+
model.to(device)
|
114 |
+
|
115 |
+
generation_config = GenerationConfig(
|
116 |
+
temperature=temperature,
|
117 |
+
max_new_tokens=max_new_tokens,
|
118 |
+
top_p=top_p,
|
119 |
+
top_k=top_k,
|
120 |
+
repetition_penalty=repetition_penalty,
|
121 |
+
do_sample=do_sample,
|
122 |
+
num_return_sequences=num_return_sequences,
|
123 |
+
)
|
124 |
+
|
125 |
+
return StreamingResponse(
|
126 |
+
stream_text(model, tokenizer, input_text, generation_config, stop_sequences, device, chunk_delay),
|
127 |
+
media_type="text/plain"
|
128 |
+
)
|
129 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
130 |
except Exception as e:
|
131 |
raise HTTPException(status_code=500, detail=f"Internal server error: {str(e)}")
|
132 |
|
133 |
+
async def stream_text(model, tokenizer, input_text, generation_config, stop_sequences, device, chunk_delay, max_length=2048):
|
134 |
+
encoded_input = tokenizer(input_text, return_tensors="pt", truncation=True, max_length=max_length).to(device)
|
135 |
+
input_length = encoded_input["input_ids"].shape[1]
|
136 |
+
remaining_tokens = max_length - input_length
|
137 |
|
138 |
+
if remaining_tokens <= 0:
|
139 |
+
yield ""
|
140 |
|
141 |
+
generation_config.max_new_tokens = min(remaining_tokens, generation_config.max_new_tokens)
|
|
|
|
|
|
|
142 |
|
143 |
+
def stop_criteria(input_ids, scores):
|
144 |
+
decoded_output = tokenizer.decode(int(input_ids[0][-1]), skip_special_tokens=True)
|
145 |
+
return decoded_output in stop_sequences
|
|
|
|
|
|
|
146 |
|
147 |
+
stopping_criteria = StoppingCriteriaList([stop_criteria])
|
148 |
|
149 |
+
output_text = ""
|
150 |
outputs = model.generate(
|
151 |
+
**encoded_input,
|
152 |
+
do_sample=generation_config.do_sample,
|
153 |
+
max_new_tokens=generation_config.max_new_tokens,
|
154 |
+
temperature=generation_config.temperature,
|
155 |
+
top_p=generation_config.top_p,
|
156 |
+
top_k=generation_config.top_k,
|
157 |
+
repetition_penalty=generation_config.repetition_penalty,
|
158 |
+
num_return_sequences=generation_config.num_return_sequences,
|
159 |
stopping_criteria=stopping_criteria,
|
160 |
+
output_scores=True,
|
161 |
+
return_dict_in_generate=True
|
162 |
)
|
163 |
|
164 |
+
for output in outputs.sequences:
|
165 |
+
for token_id in output:
|
166 |
+
token = tokenizer.decode(token_id, skip_special_tokens=True)
|
167 |
+
yield token
|
168 |
+
await asyncio.sleep(chunk_delay) # Simula el delay entre tokens
|
169 |
+
|
170 |
+
if stop_sequences and any(stop in output_text for stop in stop_sequences):
|
171 |
+
yield output_text
|
172 |
+
return
|
173 |
+
|
174 |
+
outputs = model.generate(
|
175 |
+
**encoded_input,
|
176 |
+
do_sample=generation_config.do_sample,
|
177 |
+
max_new_tokens=generation_config.max_new_tokens,
|
178 |
+
temperature=generation_config.temperature,
|
179 |
+
top_p=generation_config.top_p,
|
180 |
+
top_k=generation_config.top_k,
|
181 |
+
repetition_penalty=generation_config.repetition_penalty,
|
182 |
+
num_return_sequences=generation_config.num_return_sequences,
|
183 |
+
stopping_criteria=stopping_criteria,
|
184 |
+
output_scores=True,
|
185 |
+
return_dict_in_generate=True
|
186 |
+
)
|
187 |
|
188 |
@app.post("/generate-image")
|
189 |
async def generate_image(request: GenerateRequest):
|
190 |
try:
|
191 |
+
validated_body = request
|
192 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
193 |
+
|
194 |
+
image_generator = pipeline("text-to-image", model=validated_body.model_name, device=device)
|
195 |
+
image = image_generator(validated_body.input_text)[0]
|
196 |
+
|
197 |
+
img_byte_arr = BytesIO()
|
198 |
+
image.save(img_byte_arr, format="PNG")
|
199 |
+
img_byte_arr.seek(0)
|
200 |
+
|
201 |
+
return StreamingResponse(img_byte_arr, media_type="image/png")
|
202 |
+
|
|
|
203 |
except Exception as e:
|
204 |
raise HTTPException(status_code=500, detail=f"Internal server error: {str(e)}")
|
205 |
|
206 |
@app.post("/generate-text-to-speech")
|
207 |
async def generate_text_to_speech(request: GenerateRequest):
|
208 |
try:
|
209 |
+
validated_body = request
|
210 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
211 |
+
|
212 |
+
audio_generator = pipeline("text-to-speech", model=validated_body.model_name, device=device)
|
213 |
+
audio = audio_generator(validated_body.input_text)[0]
|
214 |
+
|
215 |
+
audio_byte_arr = BytesIO()
|
216 |
+
audio.save(audio_byte_arr)
|
217 |
+
audio_byte_arr.seek(0)
|
218 |
+
|
219 |
+
return StreamingResponse(audio_byte_arr, media_type="audio/wav")
|
220 |
+
|
|
|
221 |
except Exception as e:
|
222 |
raise HTTPException(status_code=500, detail=f"Internal server error: {str(e)}")
|
223 |
|
224 |
@app.post("/generate-video")
|
225 |
async def generate_video(request: GenerateRequest):
|
226 |
try:
|
227 |
+
validated_body = request
|
228 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
229 |
+
video_generator = pipeline("text-to-video", model=validated_body.model_name, device=device)
|
230 |
+
video = video_generator(validated_body.input_text)[0]
|
231 |
+
|
232 |
+
video_byte_arr = BytesIO()
|
233 |
+
video.save(video_byte_arr)
|
234 |
+
video_byte_arr.seek(0)
|
235 |
+
|
236 |
+
return StreamingResponse(video_byte_arr, media_type="video/mp4")
|
237 |
+
|
|
|
|
|
238 |
except Exception as e:
|
239 |
raise HTTPException(status_code=500, detail=f"Internal server error: {str(e)}")
|
240 |
|