File size: 10,741 Bytes
7d96c7a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6c293ab
7d96c7a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6c293ab
 
 
7d96c7a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6c293ab
 
7d96c7a
6c293ab
 
7d96c7a
 
6c293ab
7d96c7a
 
 
 
6c293ab
7d96c7a
 
 
 
6c293ab
 
7d96c7a
 
 
 
6c293ab
7d96c7a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6c293ab
 
7d96c7a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
from fastapi import FastAPI, HTTPException, status
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import JSONResponse
from pydantic import BaseModel
from transformers import T5Tokenizer, T5ForConditionalGeneration, AutoConfig
import torch
import os
import sys
import traceback
from typing import Optional, Dict, Any
from accelerate import Accelerator
import time
import psutil
from loguru import logger

# Configure production logging to stderr
logger.remove()  # Remove default handler
logger.add(
    sys.stderr,
    level="INFO",
    format="<green>{time:YYYY-MM-DD HH:mm:ss.SSS}</green> | <level>{level: <8}</level> | <cyan>{name}</cyan>:<cyan>{function}</cyan>:<cyan>{line}</cyan> - <level>{message}</level>"
)

# Initialize FastAPI app with metadata
app = FastAPI(
    title="Clinical Report Generator API",
    description="Production API for generating clinical report summaries using T5",
    version="1.0.0",
    docs_url="/documentation",  # Swagger UI
    redoc_url="/redoc"  # ReDoc
)

# Configure CORS for production
app.add_middleware(
    CORSMiddleware,
    allow_origins=["https://pdarleyjr.github.io"],  # GitHub Pages domain
    allow_credentials=True,
    allow_methods=["POST", "GET"],  # Restrict to needed methods
    allow_headers=["*"],
    max_age=3600,  # Cache preflight requests
)

# Model configuration
MODEL_ID = "pdarleyjr/iplc-t5-clinical"

class ModelManager:
    def __init__(self):
        self.model = None
        self.tokenizer = None
        self.accelerator = Accelerator()
        self.last_load_time = None
        self.load_lock = False

    async def load_model(self) -> bool:
        """Load model and tokenizer with proper error handling and logging"""
        if self.load_lock:
            logger.warning("Model load already in progress")
            return False

        try:
            self.load_lock = True
            logger.info("Starting model and tokenizer loading process...")
            
            # Log system resources
            memory = psutil.virtual_memory()
            logger.info(f"System memory: {memory.percent}% used, {memory.available / (1024*1024*1024):.2f}GB available")
            if torch.cuda.is_available():
                logger.info(f"CUDA memory: {torch.cuda.memory_allocated() / (1024*1024*1024):.2f}GB allocated")

            # Load tokenizer
            logger.info("Initializing tokenizer...")
            self.tokenizer = T5Tokenizer.from_pretrained(
                MODEL_ID,
                use_fast=True,
                model_max_length=512
            )
            logger.success("Tokenizer loaded successfully")

            # Load model configuration
            logger.info("Fetching model configuration...")
            config = AutoConfig.from_pretrained(
                MODEL_ID,
                trust_remote_code=False
            )
            logger.success("Model configuration loaded successfully")

            # Load the model
            logger.info("Loading model (this may take a few minutes)...")
            device = "cuda" if torch.cuda.is_available() else "cpu"
            logger.info(f"Using device: {device}")
            
            self.model = T5ForConditionalGeneration.from_pretrained(
                MODEL_ID,
                config=config,
                torch_dtype=torch.float16 if device == "cuda" else torch.float32,
                low_cpu_mem_usage=True
            ).to(device)
            logger.success("Model loaded successfully")

            # Prepare model with accelerator
            self.model = self.accelerator.prepare_model(self.model)
            logger.success("Model prepared with accelerator")

            # Log final memory usage
            memory = psutil.virtual_memory()
            logger.info(f"Final memory usage: {memory.percent}% used, {memory.available / (1024*1024*1024):.2f}GB available")
            if torch.cuda.is_available():
                logger.info(f"Final CUDA memory: {torch.cuda.memory_allocated() / (1024*1024*1024):.2f}GB allocated")

            self.last_load_time = time.time()
            return True

        except Exception as e:
            logger.exception("Error loading model")
            self.model = None
            self.tokenizer = None
            return False

        finally:
            self.load_lock = False

    def is_loaded(self) -> bool:
        """Check if model and tokenizer are loaded"""
        return self.model is not None and self.tokenizer is not None

    def get_load_time(self) -> Optional[float]:
        """Get the last successful load time"""
        return self.last_load_time

# Initialize model manager
model_manager = ModelManager()

class PredictRequest(BaseModel):
    """Request model for prediction endpoint"""
    text: str

    class Config:
        schema_extra = {
            "example": {
                "text": "evaluation type: initial. primary diagnosis: F84.0. severity: mild. primary language: english"
            }
        }

@app.post("/predict", 
         response_model=Dict[str, Any],
         status_code=status.HTTP_200_OK,
         responses={
             500: {"description": "Internal server error"},
             503: {"description": "Service unavailable - model loading"}
         })
async def predict(request: PredictRequest) -> JSONResponse:
    """Generate a clinical report summary"""
    start_time = time.time()
    
    try:
        # Check if model needs to be loaded
        if not model_manager.is_loaded():
            logger.warning("Model not loaded, attempting to load...")
            success = await model_manager.load_model()
            if not success:
                return JSONResponse(
                    status_code=status.HTTP_503_SERVICE_UNAVAILABLE,
                    content={
                        "success": False,
                        "error": "Model is initializing. Please try again in a few moments."
                    }
                )

        # Prepare input text
        input_text = "summarize: " + request.text
        input_ids = model_manager.tokenizer.encode(
            input_text,
            return_tensors="pt",
            max_length=512,
            truncation=True,
            padding=True
        )

        # Generate summary with error handling
        try:
            device = next(model_manager.model.parameters()).device
            input_ids = input_ids.to(device)
            
            with torch.no_grad(), model_manager.accelerator.autocast():
                outputs = model_manager.model.generate(
                    input_ids,
                    max_length=512,  # Increased for longer summaries
                    num_beams=5,     # Increased for better coherence
                    no_repeat_ngram_size=3,
                    length_penalty=2.0,
                    early_stopping=True,
                    pad_token_id=model_manager.tokenizer.pad_token_id,
                    eos_token_id=model_manager.tokenizer.eos_token_id,
                    temperature=0.7   # Added for more natural generation
                )

            summary = model_manager.tokenizer.decode(outputs[0], skip_special_tokens=True)
            
            # Log performance metrics
            process_time = time.time() - start_time
            logger.info(f"Summary generated in {process_time:.2f} seconds")

            return JSONResponse(
                content={
                    "success": True,
                    "data": summary,
                    "error": None,
                    "metrics": {
                        "process_time": process_time
                    }
                }
            )

        except torch.cuda.OutOfMemoryError:
            logger.error("CUDA out of memory error - clearing cache and reducing batch size")
            if torch.cuda.is_available():
                torch.cuda.empty_cache()
                logger.info(f"CUDA memory after cleanup: {torch.cuda.memory_allocated() / (1024*1024*1024):.2f}GB allocated")
            return JSONResponse(
                status_code=status.HTTP_503_SERVICE_UNAVAILABLE,
                content={
                    "success": False,
                    "error": "Server is currently overloaded. Please try again later."
                }
            )

    except Exception as e:
        logger.exception("Error in predict endpoint")
        return JSONResponse(
            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
            content={
                "success": False,
                "error": "An unexpected error occurred. Please try again later."
            }
        )

@app.get("/health",
         response_model=Dict[str, Any],
         status_code=status.HTTP_200_OK)
async def health_check() -> JSONResponse:
    """Check API and model health status"""
    try:
        is_loaded = model_manager.is_loaded()
        load_time = model_manager.get_load_time()
        
        return JSONResponse(
            content={
                "status": "healthy",
                "model_loaded": is_loaded,
                "last_load_time": load_time,
                "version": "1.0.0",
                "gpu_available": torch.cuda.is_available(),
                "gpu_name": torch.cuda.get_device_name(0) if torch.cuda.is_available() else None
            }
        )
    except Exception as e:
        logger.error(f"Error in health check: {str(e)}")
        return JSONResponse(
            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
            content={
                "status": "unhealthy",
                "error": str(e)
            }
        )

@app.on_event("startup")
async def startup_event() -> None:
    """Initialize model on startup"""
    logger.info("Starting application in production mode...")
    logger.info(f"System resources - CPU: {psutil.cpu_percent()}%, Memory: {psutil.virtual_memory().percent}%")
    if torch.cuda.is_available():
        logger.info(f"CUDA device: {torch.cuda.get_device_name(0)}")
    await model_manager.load_model()

@app.on_event("shutdown")
async def shutdown_event() -> None:
    """Clean up resources on shutdown"""
    logger.info("Initiating graceful shutdown...")
    # Clear CUDA cache and log final stats
    if torch.cuda.is_available():
        logger.info(f"Final CUDA memory before cleanup: {torch.cuda.memory_allocated() / (1024*1024*1024):.2f}GB")
        torch.cuda.empty_cache()
        logger.info("CUDA cache cleared")
    logger.info(f"Final system stats - CPU: {psutil.cpu_percent()}%, Memory: {psutil.virtual_memory().percent}%")
    logger.success("Application shutdown complete")

# Run the server
if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=7860)