Spaces:
Paused
Paused
Upload 2 files
Browse files- app.py +120 -28
- requirements.txt +2 -1
app.py
CHANGED
|
@@ -1,11 +1,16 @@
|
|
| 1 |
import os
|
| 2 |
import spaces
|
| 3 |
import torch
|
|
|
|
|
|
|
| 4 |
import gradio as gr
|
| 5 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
|
|
|
|
|
|
| 6 |
|
| 7 |
MODEL_ID = "Qwen/Qwen2.5-Coder-7B-Instruct"
|
| 8 |
API_KEY = os.environ.get("BRAIN_API_KEY", "")
|
|
|
|
| 9 |
|
| 10 |
SYSTEM_DEFAULT = (
|
| 11 |
"Eres FΓ©nix Brain, experto en Python, Gradio, FastAPI y HuggingFace Spaces. "
|
|
@@ -14,17 +19,50 @@ SYSTEM_DEFAULT = (
|
|
| 14 |
|
| 15 |
print("Cargando tokenizer...")
|
| 16 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
|
| 17 |
-
print("Cargando modelo
|
| 18 |
model = AutoModelForCausalLM.from_pretrained(
|
| 19 |
MODEL_ID,
|
| 20 |
torch_dtype=torch.bfloat16,
|
| 21 |
low_cpu_mem_usage=True,
|
| 22 |
trust_remote_code=True,
|
| 23 |
)
|
| 24 |
-
print("Modelo listo")
|
| 25 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
@spaces.GPU
|
| 27 |
def generar(prompt: str, system: str, key: str, max_tokens: int = 1024) -> str:
|
|
|
|
| 28 |
if API_KEY and key != API_KEY:
|
| 29 |
return "β Unauthorized"
|
| 30 |
if not prompt.strip():
|
|
@@ -37,24 +75,38 @@ def generar(prompt: str, system: str, key: str, max_tokens: int = 1024) -> str:
|
|
| 37 |
inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
| 38 |
with torch.no_grad():
|
| 39 |
out = model.generate(
|
| 40 |
-
**inputs,
|
| 41 |
-
|
| 42 |
-
temperature=0.2,
|
| 43 |
-
do_sample=True,
|
| 44 |
pad_token_id=tokenizer.eos_token_id,
|
| 45 |
)
|
| 46 |
gen = out[0][inputs["input_ids"].shape[1]:]
|
| 47 |
return tokenizer.decode(gen, skip_special_tokens=True)
|
| 48 |
|
|
|
|
| 49 |
@spaces.GPU
|
| 50 |
-
def chat_fn(mensaje: str, historial: list, system_prompt: str
|
| 51 |
-
if not mensaje.strip():
|
| 52 |
-
return historial, ""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
messages = [{"role": "system", "content": system_prompt or SYSTEM_DEFAULT}]
|
| 54 |
for h in historial:
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 59 |
inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
| 60 |
with torch.no_grad():
|
|
@@ -64,32 +116,72 @@ def chat_fn(mensaje: str, historial: list, system_prompt: str) -> tuple:
|
|
| 64 |
)
|
| 65 |
gen = out[0][inputs["input_ids"].shape[1]:]
|
| 66 |
respuesta = tokenizer.decode(gen, skip_special_tokens=True)
|
|
|
|
| 67 |
historial = historial + [
|
| 68 |
-
{"role": "user", "content": mensaje},
|
| 69 |
{"role": "assistant", "content": respuesta},
|
| 70 |
]
|
| 71 |
-
return historial, ""
|
| 72 |
|
| 73 |
-
|
| 74 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
|
| 76 |
with gr.Tab("π¬ Chat"):
|
| 77 |
-
system_box = gr.Textbox(label="System prompt", value=SYSTEM_DEFAULT, lines=2)
|
| 78 |
-
chatbot = gr.Chatbot(height=500, type="messages")
|
| 79 |
-
msg_box = gr.Textbox(placeholder="Escribe tu preguntaβ¦", label="Mensaje")
|
| 80 |
with gr.Row():
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 89 |
api_prompt = gr.Textbox(label="Prompt")
|
| 90 |
api_system = gr.Textbox(label="System", value=SYSTEM_DEFAULT)
|
| 91 |
api_key_box = gr.Textbox(label="API Key", type="password")
|
| 92 |
api_out = gr.Textbox(label="Respuesta", lines=10)
|
| 93 |
-
gr.Button("Probar").click(
|
|
|
|
|
|
|
| 94 |
|
| 95 |
demo.launch(server_name="0.0.0.0", server_port=7860)
|
|
|
|
| 1 |
import os
|
| 2 |
import spaces
|
| 3 |
import torch
|
| 4 |
+
import tempfile
|
| 5 |
+
import asyncio
|
| 6 |
import gradio as gr
|
| 7 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 8 |
+
import edge_tts
|
| 9 |
+
import pdfplumber
|
| 10 |
|
| 11 |
MODEL_ID = "Qwen/Qwen2.5-Coder-7B-Instruct"
|
| 12 |
API_KEY = os.environ.get("BRAIN_API_KEY", "")
|
| 13 |
+
TTS_VOZ = "es-ES-AlvaroNeural"
|
| 14 |
|
| 15 |
SYSTEM_DEFAULT = (
|
| 16 |
"Eres FΓ©nix Brain, experto en Python, Gradio, FastAPI y HuggingFace Spaces. "
|
|
|
|
| 19 |
|
| 20 |
print("Cargando tokenizer...")
|
| 21 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
|
| 22 |
+
print("Cargando modelo...")
|
| 23 |
model = AutoModelForCausalLM.from_pretrained(
|
| 24 |
MODEL_ID,
|
| 25 |
torch_dtype=torch.bfloat16,
|
| 26 |
low_cpu_mem_usage=True,
|
| 27 |
trust_remote_code=True,
|
| 28 |
)
|
| 29 |
+
print("Modelo listo β
")
|
| 30 |
|
| 31 |
+
|
| 32 |
+
# ββ TTS βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 33 |
+
async def _tts_async(texto: str, voz: str) -> str:
|
| 34 |
+
tmp = tempfile.NamedTemporaryFile(suffix=".mp3", delete=False)
|
| 35 |
+
await edge_tts.Communicate(texto[:3000], voz).save(tmp.name)
|
| 36 |
+
return tmp.name
|
| 37 |
+
|
| 38 |
+
def texto_a_voz(texto: str, voz: str) -> str:
|
| 39 |
+
try:
|
| 40 |
+
return asyncio.new_event_loop().run_until_complete(_tts_async(texto, voz))
|
| 41 |
+
except Exception as e:
|
| 42 |
+
print(f"TTS error: {e}")
|
| 43 |
+
return None
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
# ββ Leer documentos βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 47 |
+
def _leer_archivo(path: str) -> str:
|
| 48 |
+
if path.lower().endswith(".pdf"):
|
| 49 |
+
try:
|
| 50 |
+
with pdfplumber.open(path) as pdf:
|
| 51 |
+
return "\n".join(p.extract_text() or "" for p in pdf.pages[:10])
|
| 52 |
+
except Exception as e:
|
| 53 |
+
return f"[Error leyendo PDF: {e}]"
|
| 54 |
+
else:
|
| 55 |
+
try:
|
| 56 |
+
with open(path, "r", encoding="utf-8", errors="ignore") as f:
|
| 57 |
+
return f.read()[:8000]
|
| 58 |
+
except Exception as e:
|
| 59 |
+
return f"[Error leyendo archivo: {e}]"
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
# ββ Inferencia ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 63 |
@spaces.GPU
|
| 64 |
def generar(prompt: str, system: str, key: str, max_tokens: int = 1024) -> str:
|
| 65 |
+
"""Endpoint API para La Forja."""
|
| 66 |
if API_KEY and key != API_KEY:
|
| 67 |
return "β Unauthorized"
|
| 68 |
if not prompt.strip():
|
|
|
|
| 75 |
inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
| 76 |
with torch.no_grad():
|
| 77 |
out = model.generate(
|
| 78 |
+
**inputs, max_new_tokens=max_tokens,
|
| 79 |
+
temperature=0.2, do_sample=True,
|
|
|
|
|
|
|
| 80 |
pad_token_id=tokenizer.eos_token_id,
|
| 81 |
)
|
| 82 |
gen = out[0][inputs["input_ids"].shape[1]:]
|
| 83 |
return tokenizer.decode(gen, skip_special_tokens=True)
|
| 84 |
|
| 85 |
+
|
| 86 |
@spaces.GPU
|
| 87 |
+
def chat_fn(mensaje: str, archivo, historial: list, system_prompt: str, voz: str, activar_voz: bool):
|
| 88 |
+
if not mensaje.strip() and archivo is None:
|
| 89 |
+
return historial, "", None
|
| 90 |
+
|
| 91 |
+
# AΓ±adir contenido del archivo al mensaje si existe
|
| 92 |
+
texto_archivo = ""
|
| 93 |
+
if archivo is not None:
|
| 94 |
+
contenido = _leer_archivo(archivo.name)
|
| 95 |
+
nombre = os.path.basename(archivo.name)
|
| 96 |
+
texto_archivo = f"\n\n[Archivo: {nombre}]\n{contenido}"
|
| 97 |
+
|
| 98 |
+
prompt_completo = mensaje + texto_archivo
|
| 99 |
+
|
| 100 |
+
# Construir historial
|
| 101 |
messages = [{"role": "system", "content": system_prompt or SYSTEM_DEFAULT}]
|
| 102 |
for h in historial:
|
| 103 |
+
if isinstance(h, dict):
|
| 104 |
+
messages.append(h)
|
| 105 |
+
else:
|
| 106 |
+
messages.append({"role": "user", "content": h[0]})
|
| 107 |
+
messages.append({"role": "assistant", "content": h[1]})
|
| 108 |
+
messages.append({"role": "user", "content": prompt_completo})
|
| 109 |
+
|
| 110 |
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 111 |
inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
| 112 |
with torch.no_grad():
|
|
|
|
| 116 |
)
|
| 117 |
gen = out[0][inputs["input_ids"].shape[1]:]
|
| 118 |
respuesta = tokenizer.decode(gen, skip_special_tokens=True)
|
| 119 |
+
|
| 120 |
historial = historial + [
|
| 121 |
+
{"role": "user", "content": mensaje + (" π" if archivo else "")},
|
| 122 |
{"role": "assistant", "content": respuesta},
|
| 123 |
]
|
|
|
|
| 124 |
|
| 125 |
+
audio_path = texto_a_voz(respuesta, voz) if activar_voz else None
|
| 126 |
+
return historial, "", audio_path
|
| 127 |
+
|
| 128 |
+
|
| 129 |
+
# ββ UI ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 130 |
+
VOCES = {
|
| 131 |
+
"πͺπΈ Γlvaro (ES)": "es-ES-AlvaroNeural",
|
| 132 |
+
"πͺπΈ Elvira (ES)": "es-ES-ElviraNeural",
|
| 133 |
+
"π²π½ Jorge (MX)": "es-MX-JorgeNeural",
|
| 134 |
+
"π²π½ Dalia (MX)": "es-MX-DaliaNeural",
|
| 135 |
+
"π¦π· TomΓ‘s (AR)": "es-AR-TomasNeural",
|
| 136 |
+
}
|
| 137 |
+
|
| 138 |
+
with gr.Blocks(title="π§ FΓ©nix Brain", theme=gr.themes.Base()) as demo:
|
| 139 |
+
gr.Markdown("# π§ FΓ©nix Brain\nQwen2.5-Coder-7B Β· Documentos Β· Voz")
|
| 140 |
|
| 141 |
with gr.Tab("π¬ Chat"):
|
|
|
|
|
|
|
|
|
|
| 142 |
with gr.Row():
|
| 143 |
+
with gr.Column(scale=3):
|
| 144 |
+
system_box = gr.Textbox(label="System prompt", value=SYSTEM_DEFAULT, lines=2)
|
| 145 |
+
chatbot = gr.Chatbot(height=450, type="messages", label="FΓ©nix Brain")
|
| 146 |
+
msg_box = gr.Textbox(placeholder="Escribe tu mensajeβ¦", label="Mensaje", lines=2)
|
| 147 |
+
with gr.Row():
|
| 148 |
+
send_btn = gr.Button("π¨ Enviar", variant="primary")
|
| 149 |
+
clear_btn = gr.Button("ποΈ Limpiar")
|
| 150 |
+
with gr.Column(scale=1):
|
| 151 |
+
archivo_in = gr.File(label="π Archivo", file_types=[".pdf", ".txt", ".py", ".md", ".json"])
|
| 152 |
+
activar_voz = gr.Checkbox(label="π Voz activada", value=True)
|
| 153 |
+
voz_sel = gr.Dropdown(list(VOCES.keys()), value="πͺπΈ Γlvaro (ES)", label="Voz")
|
| 154 |
+
audio_out = gr.Audio(label="π Respuesta", autoplay=True)
|
| 155 |
+
|
| 156 |
+
def _chat(mensaje, archivo, historial, system, voz_nombre, activar):
|
| 157 |
+
voz = VOCES.get(voz_nombre, TTS_VOZ)
|
| 158 |
+
return chat_fn(mensaje, archivo, historial, system, voz, activar)
|
| 159 |
+
|
| 160 |
+
send_btn.click(_chat, [msg_box, archivo_in, chatbot, system_box, voz_sel, activar_voz], [chatbot, msg_box, audio_out])
|
| 161 |
+
msg_box.submit(_chat, [msg_box, archivo_in, chatbot, system_box, voz_sel, activar_voz], [chatbot, msg_box, audio_out])
|
| 162 |
+
clear_btn.click(lambda: ([], "", None), outputs=[chatbot, msg_box, audio_out])
|
| 163 |
+
|
| 164 |
+
with gr.Tab("π API (La Forja)"):
|
| 165 |
+
gr.Markdown("""
|
| 166 |
+
### Llamada desde La Forja
|
| 167 |
+
```python
|
| 168 |
+
from gradio_client import Client
|
| 169 |
+
client = Client("Cristobal299/Chat")
|
| 170 |
+
result = client.predict(
|
| 171 |
+
prompt="Tu pregunta",
|
| 172 |
+
system="",
|
| 173 |
+
key="TU_BRAIN_API_KEY",
|
| 174 |
+
max_tokens=1024,
|
| 175 |
+
api_name="/generar"
|
| 176 |
+
)
|
| 177 |
+
```
|
| 178 |
+
""")
|
| 179 |
api_prompt = gr.Textbox(label="Prompt")
|
| 180 |
api_system = gr.Textbox(label="System", value=SYSTEM_DEFAULT)
|
| 181 |
api_key_box = gr.Textbox(label="API Key", type="password")
|
| 182 |
api_out = gr.Textbox(label="Respuesta", lines=10)
|
| 183 |
+
gr.Button("Probar API").click(
|
| 184 |
+
generar, [api_prompt, api_system, api_key_box], api_out, api_name="generar"
|
| 185 |
+
)
|
| 186 |
|
| 187 |
demo.launch(server_name="0.0.0.0", server_port=7860)
|
requirements.txt
CHANGED
|
@@ -4,4 +4,5 @@ torch>=2.1.0
|
|
| 4 |
accelerate>=0.26.0
|
| 5 |
spaces
|
| 6 |
gradio_client
|
| 7 |
-
|
|
|
|
|
|
| 4 |
accelerate>=0.26.0
|
| 5 |
spaces
|
| 6 |
gradio_client
|
| 7 |
+
edge-tts
|
| 8 |
+
pdfplumber
|