Spaces:
Runtime error
Runtime error
Create triton_inference/inference_engine.py
Browse files
triton_inference/inference_engine.py
ADDED
|
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import tritonclient.grpc as grpcclient
|
| 2 |
+
import numpy as np
|
| 3 |
+
from typing import Optional
|
| 4 |
+
import logging
|
| 5 |
+
import os
|
| 6 |
+
|
| 7 |
+
class TritonInferenceEngine:
|
| 8 |
+
def __init__(self):
|
| 9 |
+
self.triton_url = os.getenv("TRITON_URL", "localhost:8001")
|
| 10 |
+
self.model_name = os.getenv("MODEL_NAME", "wizardlm-7b")
|
| 11 |
+
self.client = None
|
| 12 |
+
|
| 13 |
+
async def initialize(self):
|
| 14 |
+
try:
|
| 15 |
+
self.client = grpcclient.InferenceServerClient(
|
| 16 |
+
url=self.triton_url,
|
| 17 |
+
verbose=False
|
| 18 |
+
)
|
| 19 |
+
if not self.client.is_model_ready(self.model_name):
|
| 20 |
+
raise RuntimeError(f"Model {self.model_name} is not ready")
|
| 21 |
+
logging.info(f"Connected to Triton Inference Server at {self.triton_url}")
|
| 22 |
+
except Exception as e:
|
| 23 |
+
logging.error(f"Error connecting to Triton: {e}")
|
| 24 |
+
raise
|
| 25 |
+
|
| 26 |
+
async def generate(
|
| 27 |
+
self,
|
| 28 |
+
prompt: str,
|
| 29 |
+
max_tokens: int = 100,
|
| 30 |
+
temperature: float = 0.7,
|
| 31 |
+
top_p: float = 0.9
|
| 32 |
+
) -> str:
|
| 33 |
+
# Preprocesar el prompt (aquí necesitarías tokenizar con el tokenizer adecuado)
|
| 34 |
+
# Para simplificar, asumimos que el modelo espera input_ids y attention_mask
|
| 35 |
+
# En un caso real, usarías un tokenizer como en el ejemplo de HuggingFace
|
| 36 |
+
|
| 37 |
+
# Este es un ejemplo simplificado. En producción, necesitarás adaptarlo a tu modelo.
|
| 38 |
+
inputs = self._prepare_inputs(prompt)
|
| 39 |
+
|
| 40 |
+
# Configurar inputs para Triton
|
| 41 |
+
triton_inputs = [
|
| 42 |
+
grpcclient.InferInput("input_ids", inputs["input_ids"].shape, "INT64"),
|
| 43 |
+
grpcclient.InferInput("attention_mask", inputs["attention_mask"].shape, "INT64"),
|
| 44 |
+
]
|
| 45 |
+
triton_inputs[0].set_data_from_numpy(inputs["input_ids"])
|
| 46 |
+
triton_inputs[1].set_data_from_numpy(inputs["attention_mask"])
|
| 47 |
+
|
| 48 |
+
# Configurar outputs
|
| 49 |
+
outputs = [grpcclient.InferRequestedOutput("output_ids")]
|
| 50 |
+
|
| 51 |
+
# Realizar inferencia
|
| 52 |
+
response = self.client.infer(
|
| 53 |
+
model_name=self.model_name,
|
| 54 |
+
inputs=triton_inputs,
|
| 55 |
+
outputs=outputs
|
| 56 |
+
)
|
| 57 |
+
|
| 58 |
+
# Postprocesar la respuesta
|
| 59 |
+
output_ids = response.as_numpy("output_ids")
|
| 60 |
+
generated_text = self._decode_output(output_ids)
|
| 61 |
+
|
| 62 |
+
return generated_text
|
| 63 |
+
|
| 64 |
+
def _prepare_inputs(self, prompt: str):
|
| 65 |
+
# Aquí deberías tokenizar el prompt usando el tokenizer de WizardLM-7B
|
| 66 |
+
# Por ahora, devolvemos un ejemplo dummy
|
| 67 |
+
# En producción, carga el tokenizer y úsalo para tokenizar
|
| 68 |
+
return {
|
| 69 |
+
"input_ids": np.array([[1, 2, 3, 4, 5]], dtype=np.int64),
|
| 70 |
+
"attention_mask": np.array([[1, 1, 1, 1, 1]], dtype=np.int64)
|
| 71 |
+
}
|
| 72 |
+
|
| 73 |
+
def _decode_output(self, output_ids):
|
| 74 |
+
# Decodificar los output_ids a texto
|
| 75 |
+
# Por ahora, devolvemos un texto dummy
|
| 76 |
+
return "This is a dummy response from the AI model."
|
| 77 |
+
|
| 78 |
+
async def close(self):
|
| 79 |
+
if self.client:
|
| 80 |
+
self.client.close()
|