Update app.py
Browse files
app.py
CHANGED
|
@@ -5,14 +5,13 @@ import asyncio
|
|
| 5 |
import threading
|
| 6 |
import time
|
| 7 |
import torch
|
| 8 |
-
from fastapi import FastAPI, HTTPException, BackgroundTasks
|
| 9 |
-
from fastapi.responses import StreamingResponse, JSONResponse, PlainTextResponse, HTMLResponse
|
| 10 |
from pydantic import BaseModel, Field
|
| 11 |
from transformers import (
|
| 12 |
AutoConfig,
|
| 13 |
AutoTokenizer,
|
| 14 |
GenerationConfig,
|
| 15 |
-
BitsAndBytesConfig,
|
| 16 |
AutoModelForCausalLM,
|
| 17 |
AutoProcessor,
|
| 18 |
MusicgenForConditionalGeneration
|
|
@@ -28,25 +27,15 @@ from duckduckgo_search import DDGS
|
|
| 28 |
# -----------------------
|
| 29 |
MODEL_NAME = "jnjj/gemma-3-4b-it-1layer-actual"
|
| 30 |
MAX_CONTEXT_LEN = 1024
|
| 31 |
-
# Cuantizaci贸n y offload para texto
|
| 32 |
-
bnb_config = BitsAndBytesConfig(
|
| 33 |
-
load_in_8bit=False,
|
| 34 |
-
llm_int8_threshold=6.0,
|
| 35 |
-
llm_int8_has_fp16_weight=False
|
| 36 |
-
)
|
| 37 |
-
# Contexto m谩ximo para MusicGen
|
| 38 |
MUSICGEN_MAX_TOKENS = 256
|
| 39 |
|
| 40 |
global_model = None
|
| 41 |
global_tokenizer = None
|
| 42 |
global_tokens = {}
|
| 43 |
-
# Diffusers animation pipeline globals
|
| 44 |
motion_adapter = None
|
| 45 |
anim_pipe = None
|
| 46 |
-
# MusicGen globals
|
| 47 |
music_processor = None
|
| 48 |
music_model = None
|
| 49 |
-
|
| 50 |
executor = ThreadPoolExecutor(max_workers=4)
|
| 51 |
|
| 52 |
# -----------------------
|
|
@@ -74,7 +63,6 @@ def load_global_models():
|
|
| 74 |
global_model = AutoModelForCausalLM.from_pretrained(
|
| 75 |
MODEL_NAME,
|
| 76 |
config=text_config,
|
| 77 |
-
quantization_config=bnb_config,
|
| 78 |
device_map="auto",
|
| 79 |
offload_folder="./offload",
|
| 80 |
offload_state_dict=True,
|
|
@@ -85,7 +73,6 @@ def load_global_models():
|
|
| 85 |
)
|
| 86 |
global_model = torch.compile(global_model, backend="inductor")
|
| 87 |
|
| 88 |
-
# Tokens especiales
|
| 89 |
if global_tokenizer.eos_token_id is not None and global_tokenizer.pad_token_id is None:
|
| 90 |
global_tokenizer.pad_token_id = global_tokenizer.eos_token_id
|
| 91 |
global_tokens.update({
|
|
@@ -121,7 +108,7 @@ def load_global_models():
|
|
| 121 |
)
|
| 122 |
music_model.to(torch.device("cuda" if torch.cuda.is_available() else "cpu"))
|
| 123 |
|
| 124 |
-
print("Modelos de texto, animaci贸n y audio cargados
|
| 125 |
|
| 126 |
@app.get("/", response_class=HTMLResponse)
|
| 127 |
def index():
|
|
@@ -131,7 +118,7 @@ def index():
|
|
| 131 |
<body>
|
| 132 |
<h1>Servicio de Generaci贸n Multimedia</h1>
|
| 133 |
<ul>
|
| 134 |
-
<li>Texto:
|
| 135 |
<li>Animaci贸n: AnimateDiffPipeline con LoRA y CPU offload.</li>
|
| 136 |
<li>Audio: MusicGen small, max tokens 256.</li>
|
| 137 |
</ul>
|
|
@@ -144,9 +131,6 @@ def index():
|
|
| 144 |
def health():
|
| 145 |
return {"status": "ok"}
|
| 146 |
|
| 147 |
-
# -----------------------
|
| 148 |
-
# Pydantic Schemas
|
| 149 |
-
# -----------------------
|
| 150 |
class GenerateRequest(BaseModel):
|
| 151 |
input_text: str = ""
|
| 152 |
max_new_tokens: int = 2
|
|
@@ -178,10 +162,7 @@ class MusicRequest(BaseModel):
|
|
| 178 |
texts: list[str]
|
| 179 |
max_new_tokens: int = MUSICGEN_MAX_TOKENS
|
| 180 |
|
| 181 |
-
|
| 182 |
-
# Utility Functions
|
| 183 |
-
# -----------------------
|
| 184 |
-
def cleanup_memory():
|
| 185 |
gc.collect()
|
| 186 |
if torch.cuda.is_available():
|
| 187 |
torch.cuda.empty_cache()
|
|
@@ -199,9 +180,6 @@ async def perform_duckasgo_search(query: str, max_results: int = 3) -> str:
|
|
| 199 |
text += f"{i}. {r.get('title','')}\n URL: {r.get('href','')}\n {r.get('body','')}\n"
|
| 200 |
return text
|
| 201 |
|
| 202 |
-
# -----------------------
|
| 203 |
-
# Text Generation
|
| 204 |
-
# -----------------------
|
| 205 |
async def generate_next_token(input_ids, past_key_values, gen_config, device):
|
| 206 |
with torch.autocast(device_type=device, dtype=torch.float16):
|
| 207 |
outputs = global_model(
|
|
@@ -240,7 +218,7 @@ async def stream_text(request: GenerateRequest, device: str):
|
|
| 240 |
top_p=request.top_p,
|
| 241 |
top_k=request.top_k,
|
| 242 |
repetition_penalty=request.repetition_penalty,
|
| 243 |
-
frequency_penalty=request.
|
| 244 |
presence_penalty=request.presence_penalty,
|
| 245 |
do_sample=request.do_sample
|
| 246 |
)
|
|
@@ -266,11 +244,8 @@ async def stream_text(request: GenerateRequest, device: str):
|
|
| 266 |
yield buffer
|
| 267 |
if request.include_duckasgo:
|
| 268 |
yield "\n" + await perform_duckasgo_search(request.input_text)
|
| 269 |
-
cleanup_memory()
|
| 270 |
|
| 271 |
-
# -----------------------
|
| 272 |
-
# Endpoints
|
| 273 |
-
# -----------------------
|
| 274 |
@app.post("/generate")
|
| 275 |
async def generate_text(request: GenerateRequest, background_tasks: BackgroundTasks):
|
| 276 |
if global_model is None:
|
|
@@ -318,14 +293,10 @@ async def generate_music(request: MusicRequest):
|
|
| 318 |
return_tensors="pt"
|
| 319 |
).to(device)
|
| 320 |
with torch.no_grad():
|
| 321 |
-
audio = music_model.generate(
|
| 322 |
-
**inputs,
|
| 323 |
-
max_new_tokens=request.max_new_tokens
|
| 324 |
-
)
|
| 325 |
-
# audio is a tensor of shape (batch, seq_len)
|
| 326 |
-
# convert to WAV bytes
|
| 327 |
wav_bytes = music_processor.decode(audio[0].cpu()).numpy().tobytes()
|
| 328 |
return Response(wav_bytes, media_type="audio/wav")
|
| 329 |
|
| 330 |
if __name__ == "__main__":
|
| 331 |
uvicorn.run(app, host="0.0.0.0", port=7860)
|
|
|
|
|
|
| 5 |
import threading
|
| 6 |
import time
|
| 7 |
import torch
|
| 8 |
+
from fastapi import FastAPI, HTTPException, BackgroundTasks, Response
|
| 9 |
+
from fastapi.responses import StreamingResponse, JSONResponse, PlainTextResponse, HTMLResponse
|
| 10 |
from pydantic import BaseModel, Field
|
| 11 |
from transformers import (
|
| 12 |
AutoConfig,
|
| 13 |
AutoTokenizer,
|
| 14 |
GenerationConfig,
|
|
|
|
| 15 |
AutoModelForCausalLM,
|
| 16 |
AutoProcessor,
|
| 17 |
MusicgenForConditionalGeneration
|
|
|
|
| 27 |
# -----------------------
|
| 28 |
MODEL_NAME = "jnjj/gemma-3-4b-it-1layer-actual"
|
| 29 |
MAX_CONTEXT_LEN = 1024
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 30 |
MUSICGEN_MAX_TOKENS = 256
|
| 31 |
|
| 32 |
global_model = None
|
| 33 |
global_tokenizer = None
|
| 34 |
global_tokens = {}
|
|
|
|
| 35 |
motion_adapter = None
|
| 36 |
anim_pipe = None
|
|
|
|
| 37 |
music_processor = None
|
| 38 |
music_model = None
|
|
|
|
| 39 |
executor = ThreadPoolExecutor(max_workers=4)
|
| 40 |
|
| 41 |
# -----------------------
|
|
|
|
| 63 |
global_model = AutoModelForCausalLM.from_pretrained(
|
| 64 |
MODEL_NAME,
|
| 65 |
config=text_config,
|
|
|
|
| 66 |
device_map="auto",
|
| 67 |
offload_folder="./offload",
|
| 68 |
offload_state_dict=True,
|
|
|
|
| 73 |
)
|
| 74 |
global_model = torch.compile(global_model, backend="inductor")
|
| 75 |
|
|
|
|
| 76 |
if global_tokenizer.eos_token_id is not None and global_tokenizer.pad_token_id is None:
|
| 77 |
global_tokenizer.pad_token_id = global_tokenizer.eos_token_id
|
| 78 |
global_tokens.update({
|
|
|
|
| 108 |
)
|
| 109 |
music_model.to(torch.device("cuda" if torch.cuda.is_available() else "cpu"))
|
| 110 |
|
| 111 |
+
print("Modelos de texto, animaci贸n y audio cargados sin bitsandbytes.")
|
| 112 |
|
| 113 |
@app.get("/", response_class=HTMLResponse)
|
| 114 |
def index():
|
|
|
|
| 118 |
<body>
|
| 119 |
<h1>Servicio de Generaci贸n Multimedia</h1>
|
| 120 |
<ul>
|
| 121 |
+
<li>Texto: FP16, offload, torch.compile.</li>
|
| 122 |
<li>Animaci贸n: AnimateDiffPipeline con LoRA y CPU offload.</li>
|
| 123 |
<li>Audio: MusicGen small, max tokens 256.</li>
|
| 124 |
</ul>
|
|
|
|
| 131 |
def health():
|
| 132 |
return {"status": "ok"}
|
| 133 |
|
|
|
|
|
|
|
|
|
|
| 134 |
class GenerateRequest(BaseModel):
|
| 135 |
input_text: str = ""
|
| 136 |
max_new_tokens: int = 2
|
|
|
|
| 162 |
texts: list[str]
|
| 163 |
max_new_tokens: int = MUSICGEN_MAX_TOKENS
|
| 164 |
|
| 165 |
+
async def cleanup_memory():
|
|
|
|
|
|
|
|
|
|
| 166 |
gc.collect()
|
| 167 |
if torch.cuda.is_available():
|
| 168 |
torch.cuda.empty_cache()
|
|
|
|
| 180 |
text += f"{i}. {r.get('title','')}\n URL: {r.get('href','')}\n {r.get('body','')}\n"
|
| 181 |
return text
|
| 182 |
|
|
|
|
|
|
|
|
|
|
| 183 |
async def generate_next_token(input_ids, past_key_values, gen_config, device):
|
| 184 |
with torch.autocast(device_type=device, dtype=torch.float16):
|
| 185 |
outputs = global_model(
|
|
|
|
| 218 |
top_p=request.top_p,
|
| 219 |
top_k=request.top_k,
|
| 220 |
repetition_penalty=request.repetition_penalty,
|
| 221 |
+
frequency_penalty=request.frequency.penalty,
|
| 222 |
presence_penalty=request.presence_penalty,
|
| 223 |
do_sample=request.do_sample
|
| 224 |
)
|
|
|
|
| 244 |
yield buffer
|
| 245 |
if request.include_duckasgo:
|
| 246 |
yield "\n" + await perform_duckasgo_search(request.input_text)
|
| 247 |
+
await cleanup_memory()
|
| 248 |
|
|
|
|
|
|
|
|
|
|
| 249 |
@app.post("/generate")
|
| 250 |
async def generate_text(request: GenerateRequest, background_tasks: BackgroundTasks):
|
| 251 |
if global_model is None:
|
|
|
|
| 293 |
return_tensors="pt"
|
| 294 |
).to(device)
|
| 295 |
with torch.no_grad():
|
| 296 |
+
audio = music_model.generate(**inputs, max_new_tokens=request.max_new_tokens)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 297 |
wav_bytes = music_processor.decode(audio[0].cpu()).numpy().tobytes()
|
| 298 |
return Response(wav_bytes, media_type="audio/wav")
|
| 299 |
|
| 300 |
if __name__ == "__main__":
|
| 301 |
uvicorn.run(app, host="0.0.0.0", port=7860)
|
| 302 |
+
|