Spaces:
Paused
Paused
Commit
·
28fa644
1
Parent(s):
eda2ff2
Small refactor, moving endpoints to the routes.py file. Also added streaming endpoint, and from_pretrained
Browse files- main/main.py +4 -166
- main/routes.py +365 -0
main/main.py
CHANGED
@@ -1,13 +1,9 @@
|
|
1 |
-
from fastapi import FastAPI
|
2 |
from fastapi.middleware.cors import CORSMiddleware
|
3 |
-
from pydantic import BaseModel
|
4 |
-
from typing import Optional, Union
|
5 |
-
import torch
|
6 |
import logging
|
7 |
-
from pathlib import Path
|
8 |
-
from litgpt.api import LLM
|
9 |
import os
|
10 |
import uvicorn
|
|
|
11 |
|
12 |
# Set up logging
|
13 |
logging.basicConfig(level=logging.INFO)
|
@@ -30,166 +26,8 @@ app.add_middleware(
|
|
30 |
allow_headers=["*"],
|
31 |
)
|
32 |
|
33 |
-
#
|
34 |
-
|
35 |
-
|
36 |
-
class InitializeRequest(BaseModel):
|
37 |
-
"""
|
38 |
-
Configuration for model initialization including model path
|
39 |
-
"""
|
40 |
-
mode: str = "cpu"
|
41 |
-
precision: Optional[str] = None
|
42 |
-
quantize: Optional[str] = None
|
43 |
-
gpu_count: Union[str, int] = "auto"
|
44 |
-
model_path: str
|
45 |
-
|
46 |
-
class GenerateRequest(BaseModel):
|
47 |
-
prompt: str
|
48 |
-
max_new_tokens: int = 50
|
49 |
-
temperature: float = 1.0
|
50 |
-
top_k: Optional[int] = None
|
51 |
-
top_p: float = 1.0
|
52 |
-
return_as_token_ids: bool = False
|
53 |
-
stream: bool = False
|
54 |
-
|
55 |
-
@app.get("/")
|
56 |
-
async def root():
|
57 |
-
"""Root endpoint to verify service is running"""
|
58 |
-
return {
|
59 |
-
"status": "running",
|
60 |
-
"service": "LLM Engine",
|
61 |
-
"endpoints": {
|
62 |
-
"initialize": "/initialize",
|
63 |
-
"generate": "/generate",
|
64 |
-
"health": "/health"
|
65 |
-
}
|
66 |
-
}
|
67 |
-
|
68 |
-
@app.post("/initialize")
|
69 |
-
async def initialize_model(request: InitializeRequest):
|
70 |
-
"""
|
71 |
-
Initialize the LLM model with specified configuration.
|
72 |
-
"""
|
73 |
-
global llm_instance
|
74 |
-
|
75 |
-
try:
|
76 |
-
# Get the project root directory (where main.py is located)
|
77 |
-
project_root = Path(__file__).parent
|
78 |
-
checkpoints_dir = project_root / "checkpoints"
|
79 |
-
logger.info(f"Checkpoint dir is: {checkpoints_dir}")
|
80 |
-
|
81 |
-
# For LitGPT downloaded models, path includes organization
|
82 |
-
if "/" in request.model_path:
|
83 |
-
# e.g., "mistralai/Mistral-7B-Instruct-v0.3"
|
84 |
-
org, model_name = request.model_path.split("/")
|
85 |
-
model_path = str(checkpoints_dir / org / model_name)
|
86 |
-
else:
|
87 |
-
# Fallback for direct model paths
|
88 |
-
model_path = str(checkpoints_dir / request.model_path)
|
89 |
-
|
90 |
-
logger.info(f"Using model path: {model_path}")
|
91 |
-
|
92 |
-
# Load the model
|
93 |
-
llm_instance = LLM.load(
|
94 |
-
model=model_path,
|
95 |
-
distribute=None if request.precision or request.quantize else "auto"
|
96 |
-
)
|
97 |
-
|
98 |
-
# If manual distribution is needed
|
99 |
-
if request.precision or request.quantize:
|
100 |
-
llm_instance.distribute(
|
101 |
-
accelerator="cuda" if request.mode == "gpu" else "cpu",
|
102 |
-
devices=request.gpu_count,
|
103 |
-
precision=request.precision,
|
104 |
-
quantize=request.quantize
|
105 |
-
)
|
106 |
-
|
107 |
-
logger.info(
|
108 |
-
f"Model initialized successfully with config:\n"
|
109 |
-
f"Mode: {request.mode}\n"
|
110 |
-
f"Precision: {request.precision}\n"
|
111 |
-
f"Quantize: {request.quantize}\n"
|
112 |
-
f"GPU Count: {request.gpu_count}\n"
|
113 |
-
f"Model Path: {model_path}\n"
|
114 |
-
f"Current GPU Memory: {torch.cuda.memory_allocated()/1024**3:.2f}GB allocated, "
|
115 |
-
f"{torch.cuda.memory_reserved()/1024**3:.2f}GB reserved"
|
116 |
-
)
|
117 |
-
|
118 |
-
return {"success": True, "message": "Model initialized successfully"}
|
119 |
-
|
120 |
-
except Exception as e:
|
121 |
-
logger.error(f"Error initializing model: {str(e)}")
|
122 |
-
# Print detailed memory statistics on failure
|
123 |
-
logger.error(f"GPU Memory Stats:\n"
|
124 |
-
f"Allocated: {torch.cuda.memory_allocated()/1024**3:.2f}GB\n"
|
125 |
-
f"Reserved: {torch.cuda.memory_reserved()/1024**3:.2f}GB\n"
|
126 |
-
f"Max Allocated: {torch.cuda.max_memory_allocated()/1024**3:.2f}GB")
|
127 |
-
raise HTTPException(status_code=500, detail=f"Error initializing model: {str(e)}")
|
128 |
-
|
129 |
-
@app.post("/generate")
|
130 |
-
async def generate(request: GenerateRequest):
|
131 |
-
"""
|
132 |
-
Generate text using the initialized model.
|
133 |
-
"""
|
134 |
-
global llm_instance
|
135 |
-
|
136 |
-
if llm_instance is None:
|
137 |
-
raise HTTPException(status_code=400, detail="Model not initialized. Call /initialize first.")
|
138 |
-
|
139 |
-
try:
|
140 |
-
if request.stream:
|
141 |
-
raise HTTPException(
|
142 |
-
status_code=400,
|
143 |
-
detail="Streaming is not currently supported through the API"
|
144 |
-
)
|
145 |
-
|
146 |
-
generated_text = llm_instance.generate(
|
147 |
-
prompt=request.prompt,
|
148 |
-
max_new_tokens=request.max_new_tokens,
|
149 |
-
temperature=request.temperature,
|
150 |
-
top_k=request.top_k,
|
151 |
-
top_p=request.top_p,
|
152 |
-
return_as_token_ids=request.return_as_token_ids,
|
153 |
-
stream=False # Force stream to False for now
|
154 |
-
)
|
155 |
-
|
156 |
-
response = {
|
157 |
-
"generated_text": generated_text if not request.return_as_token_ids else generated_text.tolist(),
|
158 |
-
"metadata": {
|
159 |
-
"prompt": request.prompt,
|
160 |
-
"max_new_tokens": request.max_new_tokens,
|
161 |
-
"temperature": request.temperature,
|
162 |
-
"top_k": request.top_k,
|
163 |
-
"top_p": request.top_p
|
164 |
-
}
|
165 |
-
}
|
166 |
-
|
167 |
-
return response
|
168 |
-
|
169 |
-
except Exception as e:
|
170 |
-
logger.error(f"Error generating text: {str(e)}")
|
171 |
-
raise HTTPException(status_code=500, detail=f"Error generating text: {str(e)}")
|
172 |
-
|
173 |
-
@app.get("/health")
|
174 |
-
async def health_check():
|
175 |
-
"""
|
176 |
-
Check if the service is running and model is loaded.
|
177 |
-
"""
|
178 |
-
global llm_instance
|
179 |
-
|
180 |
-
status = {
|
181 |
-
"status": "healthy",
|
182 |
-
"model_loaded": llm_instance is not None,
|
183 |
-
}
|
184 |
-
|
185 |
-
if llm_instance is not None:
|
186 |
-
logger.info(f"llm_instance is: {llm_instance}")
|
187 |
-
status["model_info"] = {
|
188 |
-
"model_path": llm_instance.config.name,
|
189 |
-
"device": str(next(llm_instance.model.parameters()).device)
|
190 |
-
}
|
191 |
-
|
192 |
-
return status
|
193 |
|
194 |
def main():
|
195 |
# Load environment variables or configuration here
|
|
|
1 |
+
from fastapi import FastAPI
|
2 |
from fastapi.middleware.cors import CORSMiddleware
|
|
|
|
|
|
|
3 |
import logging
|
|
|
|
|
4 |
import os
|
5 |
import uvicorn
|
6 |
+
from routes import router
|
7 |
|
8 |
# Set up logging
|
9 |
logging.basicConfig(level=logging.INFO)
|
|
|
26 |
allow_headers=["*"],
|
27 |
)
|
28 |
|
29 |
+
# Include the router from routes.py
|
30 |
+
app.include_router(router)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
|
32 |
def main():
|
33 |
# Load environment variables or configuration here
|
main/routes.py
ADDED
@@ -0,0 +1,365 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from fastapi import APIRouter, HTTPException
|
2 |
+
from fastapi.responses import StreamingResponse
|
3 |
+
from pydantic import BaseModel
|
4 |
+
from typing import Optional, Union, AsyncGenerator
|
5 |
+
import torch
|
6 |
+
import logging
|
7 |
+
from pathlib import Path
|
8 |
+
from litgpt.api import LLM
|
9 |
+
import json
|
10 |
+
import asyncio
|
11 |
+
|
12 |
+
# Set up logging
|
13 |
+
logger = logging.getLogger(__name__)
|
14 |
+
|
15 |
+
# Create router instance
|
16 |
+
router = APIRouter()
|
17 |
+
|
18 |
+
# Global variable to store the LLM instance
|
19 |
+
llm_instance = None
|
20 |
+
|
21 |
+
class InitializeRequest(BaseModel):
|
22 |
+
"""
|
23 |
+
Configuration for model initialization including model path
|
24 |
+
"""
|
25 |
+
mode: str = "cpu"
|
26 |
+
precision: Optional[str] = None
|
27 |
+
quantize: Optional[str] = None
|
28 |
+
gpu_count: Union[str, int] = "auto"
|
29 |
+
model_path: str
|
30 |
+
|
31 |
+
class GenerateRequest(BaseModel):
|
32 |
+
prompt: str
|
33 |
+
max_new_tokens: int = 50
|
34 |
+
temperature: float = 1.0
|
35 |
+
top_k: Optional[int] = None
|
36 |
+
top_p: float = 1.0
|
37 |
+
return_as_token_ids: bool = False
|
38 |
+
stream: bool = False
|
39 |
+
|
40 |
+
# A Pydantic model for the streaming generation request
|
41 |
+
class StreamGenerateRequest(BaseModel):
|
42 |
+
prompt: str
|
43 |
+
max_new_tokens: int = 50
|
44 |
+
temperature: float = 1.0
|
45 |
+
top_k: Optional[int] = None
|
46 |
+
top_p: float = 1.0
|
47 |
+
|
48 |
+
class InitializeCustomRequest(BaseModel):
|
49 |
+
"""
|
50 |
+
Configuration for custom model initialization using from_pretrained
|
51 |
+
"""
|
52 |
+
mode: str = "cpu"
|
53 |
+
precision: Optional[str] = None
|
54 |
+
quantize: Optional[str] = None
|
55 |
+
gpu_count: Union[str, int] = "auto"
|
56 |
+
folder_path: str # Path to the model folder relative to checkpoints
|
57 |
+
model_filename: str # Name of the model file (e.g., "lit_model.pth")
|
58 |
+
config_filename: str = "config.json" # Default config filename
|
59 |
+
tokenizer_filename: Optional[str] = "tokenizer.json" # Optional tokenizer filename
|
60 |
+
|
61 |
+
|
62 |
+
@router.post("/initialize/custom")
|
63 |
+
async def initialize_custom_model(request: InitializeCustomRequest):
|
64 |
+
"""
|
65 |
+
Initialize a custom model using from_pretrained method.
|
66 |
+
This is for models that are already downloaded and stored in the checkpoints directory.
|
67 |
+
"""
|
68 |
+
global llm_instance
|
69 |
+
|
70 |
+
try:
|
71 |
+
# Get the project root directory and construct paths
|
72 |
+
project_root = Path(__file__).parent
|
73 |
+
checkpoints_dir = project_root / "checkpoints"
|
74 |
+
model_dir = checkpoints_dir / request.folder_path
|
75 |
+
|
76 |
+
logger.info(f"Loading custom model from directory: {model_dir}")
|
77 |
+
|
78 |
+
# Verify that all required files exist
|
79 |
+
model_path = model_dir / request.model_filename
|
80 |
+
config_path = model_dir / request.config_filename
|
81 |
+
|
82 |
+
if not model_path.exists():
|
83 |
+
raise HTTPException(
|
84 |
+
status_code=400,
|
85 |
+
detail=f"Model file not found: {request.model_filename}"
|
86 |
+
)
|
87 |
+
|
88 |
+
if not config_path.exists():
|
89 |
+
raise HTTPException(
|
90 |
+
status_code=400,
|
91 |
+
detail=f"Config file not found: {request.config_filename}"
|
92 |
+
)
|
93 |
+
|
94 |
+
# Check for tokenizer if specified
|
95 |
+
tokenizer_path = None
|
96 |
+
if request.tokenizer_filename:
|
97 |
+
tokenizer_path = model_dir / request.tokenizer_filename
|
98 |
+
if not tokenizer_path.exists():
|
99 |
+
raise HTTPException(
|
100 |
+
status_code=400,
|
101 |
+
detail=f"Tokenizer file not found: {request.tokenizer_filename}"
|
102 |
+
)
|
103 |
+
|
104 |
+
# Load the model using from_pretrained
|
105 |
+
llm_instance = LLM.from_pretrained(
|
106 |
+
path=str(model_dir),
|
107 |
+
model_file=request.model_filename,
|
108 |
+
config_file=request.config_filename,
|
109 |
+
tokenizer_file=request.tokenizer_filename if request.tokenizer_filename else None,
|
110 |
+
distribute=None if request.precision or request.quantize else "auto"
|
111 |
+
)
|
112 |
+
|
113 |
+
# If manual distribution is needed
|
114 |
+
if request.precision or request.quantize:
|
115 |
+
llm_instance.distribute(
|
116 |
+
accelerator="cuda" if request.mode == "gpu" else "cpu",
|
117 |
+
devices=request.gpu_count,
|
118 |
+
precision=request.precision,
|
119 |
+
quantize=request.quantize
|
120 |
+
)
|
121 |
+
|
122 |
+
# Log success and memory stats
|
123 |
+
logger.info(
|
124 |
+
f"Custom model initialized successfully with config:\n"
|
125 |
+
f"Mode: {request.mode}\n"
|
126 |
+
f"Precision: {request.precision}\n"
|
127 |
+
f"Quantize: {request.quantize}\n"
|
128 |
+
f"GPU Count: {request.gpu_count}\n"
|
129 |
+
f"Model Directory: {model_dir}\n"
|
130 |
+
f"Model File: {request.model_filename}\n"
|
131 |
+
f"Config File: {request.config_filename}\n"
|
132 |
+
f"Tokenizer File: {request.tokenizer_filename}\n"
|
133 |
+
f"Current GPU Memory: {torch.cuda.memory_allocated()/1024**3:.2f}GB allocated, "
|
134 |
+
f"{torch.cuda.memory_reserved()/1024**3:.2f}GB reserved"
|
135 |
+
)
|
136 |
+
|
137 |
+
return {
|
138 |
+
"success": True,
|
139 |
+
"message": "Custom model initialized successfully",
|
140 |
+
"model_info": {
|
141 |
+
"folder": str(model_dir),
|
142 |
+
"model_file": request.model_filename,
|
143 |
+
"config_file": request.config_filename,
|
144 |
+
"tokenizer_file": request.tokenizer_filename
|
145 |
+
}
|
146 |
+
}
|
147 |
+
|
148 |
+
except Exception as e:
|
149 |
+
logger.error(f"Error initializing custom model: {str(e)}")
|
150 |
+
# Print detailed memory statistics on failure
|
151 |
+
logger.error(f"GPU Memory Stats:\n"
|
152 |
+
f"Allocated: {torch.cuda.memory_allocated()/1024**3:.2f}GB\n"
|
153 |
+
f"Reserved: {torch.cuda.memory_reserved()/1024**3:.2f}GB\n"
|
154 |
+
f"Max Allocated: {torch.cuda.max_memory_allocated()/1024**3:.2f}GB")
|
155 |
+
raise HTTPException(status_code=500, detail=f"Error initializing custom model: {str(e)}")
|
156 |
+
|
157 |
+
|
158 |
+
# Endpoint for streaming generation
|
159 |
+
@router.post("/generate/stream")
|
160 |
+
async def generate_stream(request: StreamGenerateRequest):
|
161 |
+
"""
|
162 |
+
Generate text using the initialized model with streaming response.
|
163 |
+
Returns a StreamingResponse that yields JSON-formatted chunks of text.
|
164 |
+
"""
|
165 |
+
global llm_instance
|
166 |
+
|
167 |
+
if llm_instance is None:
|
168 |
+
raise HTTPException(
|
169 |
+
status_code=400,
|
170 |
+
detail="Model not initialized. Call /initialize first."
|
171 |
+
)
|
172 |
+
|
173 |
+
async def event_generator() -> AsyncGenerator[str, None]:
|
174 |
+
try:
|
175 |
+
# Start the generation with streaming enabled
|
176 |
+
async for token in llm_instance.generate(
|
177 |
+
prompt=request.prompt,
|
178 |
+
max_new_tokens=request.max_new_tokens,
|
179 |
+
temperature=request.temperature,
|
180 |
+
top_k=request.top_k,
|
181 |
+
top_p=request.top_p,
|
182 |
+
stream=True # Enable streaming
|
183 |
+
):
|
184 |
+
# Create a JSON response for each token
|
185 |
+
chunk = {
|
186 |
+
"token": token,
|
187 |
+
"metadata": {
|
188 |
+
"prompt": request.prompt,
|
189 |
+
"is_finished": False
|
190 |
+
}
|
191 |
+
}
|
192 |
+
# Format as SSE data
|
193 |
+
yield f"data: {json.dumps(chunk)}\n\n"
|
194 |
+
|
195 |
+
# Small delay to prevent overwhelming the client
|
196 |
+
await asyncio.sleep(0.01)
|
197 |
+
|
198 |
+
# Send final message indicating completion
|
199 |
+
final_chunk = {
|
200 |
+
"token": "",
|
201 |
+
"metadata": {
|
202 |
+
"prompt": request.prompt,
|
203 |
+
"is_finished": True
|
204 |
+
}
|
205 |
+
}
|
206 |
+
yield f"data: {json.dumps(final_chunk)}\n\n"
|
207 |
+
|
208 |
+
except Exception as e:
|
209 |
+
logger.error(f"Error in stream generation: {str(e)}")
|
210 |
+
error_chunk = {
|
211 |
+
"error": str(e),
|
212 |
+
"metadata": {
|
213 |
+
"prompt": request.prompt,
|
214 |
+
"is_finished": True
|
215 |
+
}
|
216 |
+
}
|
217 |
+
yield f"data: {json.dumps(error_chunk)}\n\n"
|
218 |
+
|
219 |
+
return StreamingResponse(
|
220 |
+
event_generator(),
|
221 |
+
media_type="text/event-stream",
|
222 |
+
headers={
|
223 |
+
'Cache-Control': 'no-cache',
|
224 |
+
'Connection': 'keep-alive',
|
225 |
+
}
|
226 |
+
)
|
227 |
+
|
228 |
+
@router.get("/")
|
229 |
+
async def root():
|
230 |
+
"""Root endpoint to verify service is running"""
|
231 |
+
return {
|
232 |
+
"status": "running",
|
233 |
+
"service": "LLM Engine",
|
234 |
+
"endpoints": {
|
235 |
+
"initialize": "/initialize",
|
236 |
+
"generate": "/generate",
|
237 |
+
"health": "/health"
|
238 |
+
}
|
239 |
+
}
|
240 |
+
|
241 |
+
@router.post("/initialize")
|
242 |
+
async def initialize_model(request: InitializeRequest):
|
243 |
+
"""
|
244 |
+
Initialize the LLM model with specified configuration.
|
245 |
+
"""
|
246 |
+
global llm_instance
|
247 |
+
|
248 |
+
try:
|
249 |
+
# Get the project root directory (where main.py is located)
|
250 |
+
project_root = Path(__file__).parent
|
251 |
+
checkpoints_dir = project_root / "checkpoints"
|
252 |
+
logger.info(f"Checkpoint dir is: {checkpoints_dir}")
|
253 |
+
|
254 |
+
# For LitGPT downloaded models, path includes organization
|
255 |
+
if "/" in request.model_path:
|
256 |
+
# e.g., "mistralai/Mistral-7B-Instruct-v0.3"
|
257 |
+
org, model_name = request.model_path.split("/")
|
258 |
+
model_path = str(checkpoints_dir / org / model_name)
|
259 |
+
else:
|
260 |
+
# Fallback for direct model paths
|
261 |
+
model_path = str(checkpoints_dir / request.model_path)
|
262 |
+
|
263 |
+
logger.info(f"Using model path: {model_path}")
|
264 |
+
|
265 |
+
# Load the model
|
266 |
+
llm_instance = LLM.load(
|
267 |
+
model=model_path,
|
268 |
+
distribute=None if request.precision or request.quantize else "auto"
|
269 |
+
)
|
270 |
+
|
271 |
+
# If manual distribution is needed
|
272 |
+
if request.precision or request.quantize:
|
273 |
+
llm_instance.distribute(
|
274 |
+
accelerator="cuda" if request.mode == "gpu" else "cpu",
|
275 |
+
devices=request.gpu_count,
|
276 |
+
precision=request.precision,
|
277 |
+
quantize=request.quantize
|
278 |
+
)
|
279 |
+
|
280 |
+
logger.info(
|
281 |
+
f"Model initialized successfully with config:\n"
|
282 |
+
f"Mode: {request.mode}\n"
|
283 |
+
f"Precision: {request.precision}\n"
|
284 |
+
f"Quantize: {request.quantize}\n"
|
285 |
+
f"GPU Count: {request.gpu_count}\n"
|
286 |
+
f"Model Path: {model_path}\n"
|
287 |
+
f"Current GPU Memory: {torch.cuda.memory_allocated()/1024**3:.2f}GB allocated, "
|
288 |
+
f"{torch.cuda.memory_reserved()/1024**3:.2f}GB reserved"
|
289 |
+
)
|
290 |
+
|
291 |
+
return {"success": True, "message": "Model initialized successfully"}
|
292 |
+
|
293 |
+
except Exception as e:
|
294 |
+
logger.error(f"Error initializing model: {str(e)}")
|
295 |
+
# Print detailed memory statistics on failure
|
296 |
+
logger.error(f"GPU Memory Stats:\n"
|
297 |
+
f"Allocated: {torch.cuda.memory_allocated()/1024**3:.2f}GB\n"
|
298 |
+
f"Reserved: {torch.cuda.memory_reserved()/1024**3:.2f}GB\n"
|
299 |
+
f"Max Allocated: {torch.cuda.max_memory_allocated()/1024**3:.2f}GB")
|
300 |
+
raise HTTPException(status_code=500, detail=f"Error initializing model: {str(e)}")
|
301 |
+
|
302 |
+
@router.post("/generate")
|
303 |
+
async def generate(request: GenerateRequest):
|
304 |
+
"""
|
305 |
+
Generate text using the initialized model.
|
306 |
+
"""
|
307 |
+
global llm_instance
|
308 |
+
|
309 |
+
if llm_instance is None:
|
310 |
+
raise HTTPException(status_code=400, detail="Model not initialized. Call /initialize first.")
|
311 |
+
|
312 |
+
try:
|
313 |
+
if request.stream:
|
314 |
+
raise HTTPException(
|
315 |
+
status_code=400,
|
316 |
+
detail="Streaming is not currently supported through the API"
|
317 |
+
)
|
318 |
+
|
319 |
+
generated_text = llm_instance.generate(
|
320 |
+
prompt=request.prompt,
|
321 |
+
max_new_tokens=request.max_new_tokens,
|
322 |
+
temperature=request.temperature,
|
323 |
+
top_k=request.top_k,
|
324 |
+
top_p=request.top_p,
|
325 |
+
return_as_token_ids=request.return_as_token_ids,
|
326 |
+
stream=False # Force stream to False for now
|
327 |
+
)
|
328 |
+
|
329 |
+
response = {
|
330 |
+
"generated_text": generated_text if not request.return_as_token_ids else generated_text.tolist(),
|
331 |
+
"metadata": {
|
332 |
+
"prompt": request.prompt,
|
333 |
+
"max_new_tokens": request.max_new_tokens,
|
334 |
+
"temperature": request.temperature,
|
335 |
+
"top_k": request.top_k,
|
336 |
+
"top_p": request.top_p
|
337 |
+
}
|
338 |
+
}
|
339 |
+
|
340 |
+
return response
|
341 |
+
|
342 |
+
except Exception as e:
|
343 |
+
logger.error(f"Error generating text: {str(e)}")
|
344 |
+
raise HTTPException(status_code=500, detail=f"Error generating text: {str(e)}")
|
345 |
+
|
346 |
+
@router.get("/health")
|
347 |
+
async def health_check():
|
348 |
+
"""
|
349 |
+
Check if the service is running and model is loaded.
|
350 |
+
"""
|
351 |
+
global llm_instance
|
352 |
+
|
353 |
+
status = {
|
354 |
+
"status": "healthy",
|
355 |
+
"model_loaded": llm_instance is not None,
|
356 |
+
}
|
357 |
+
|
358 |
+
if llm_instance is not None:
|
359 |
+
logger.info(f"llm_instance is: {llm_instance}")
|
360 |
+
status["model_info"] = {
|
361 |
+
"model_path": llm_instance.config.name,
|
362 |
+
"device": str(next(llm_instance.model.parameters()).device)
|
363 |
+
}
|
364 |
+
|
365 |
+
return status
|