gemma_gguf_backend.py

#!/usr/bin/env python3
"""
Working Gemma 3n GGUF Backend Service
Minimal FastAPI backend using only llama-cpp-python for GGUF models
"""

import os
import logging
import time
from contextlib import asynccontextmanager
from typing import List, Dict, Any, Optional
import uuid
import sys
import subprocess
import threading
from pathlib import Path
import signal  # Use signal.SIGTERM for process termination

from fastapi import FastAPI, HTTPException, Query
from fastapi.responses import JSONResponse
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel, Field, field_validator

# Import llama-cpp-python for GGUF model support
try:
    from llama_cpp import Llama
    llama_cpp_available = True
except ImportError:
    llama_cpp_available = False

import uvicorn
import sqlite3
import json  # For persisting job metadata

# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Pydantic models for OpenAI-compatible API
class ChatMessage(BaseModel):
    role: str = Field(..., description="The role of the message author")
    content: str = Field(..., description="The content of the message")
    
    @field_validator('role')
    @classmethod
    def validate_role(cls, v: str) -> str:
        if v not in ["system", "user", "assistant"]:
            raise ValueError("Role must be one of: system, user, assistant")
        return v

class ChatCompletionRequest(BaseModel):
    model: str = Field(default="gemma-3n-e4b-it", description="The model to use for completion")
    messages: List[ChatMessage] = Field(..., description="List of messages in the conversation")
    max_tokens: Optional[int] = Field(default=256, ge=1, le=1024, description="Maximum tokens to generate (reduced for memory efficiency)")
    temperature: Optional[float] = Field(default=1.0, ge=0.0, le=2.0, description="Sampling temperature")
    top_p: Optional[float] = Field(default=0.95, ge=0.0, le=1.0, description="Top-p sampling")
    top_k: Optional[int] = Field(default=64, ge=1, le=100, description="Top-k sampling")
    stream: Optional[bool] = Field(default=False, description="Whether to stream responses")

class ChatCompletionChoice(BaseModel):
    index: int
    message: ChatMessage
    finish_reason: str

class ChatCompletionResponse(BaseModel):
    id: str
    object: str = "chat.completion"
    created: int
    model: str
    choices: List[ChatCompletionChoice]

class HealthResponse(BaseModel):
    status: str
    model: str
    version: str
    backend: str

from pathlib import Path
# Global variables for model management
current_model = os.environ.get("AI_MODEL", "unsloth/gemma-3n-E4B-it-GGUF")
llm = None

def convert_messages_to_gemma_prompt(messages: List[ChatMessage]) -> str:
    """Convert OpenAI messages format to Gemma 3n chat format."""
    # Gemma 3n uses specific format with <start_of_turn> and <end_of_turn>
    prompt_parts = ["<bos>"]
    
    for message in messages:
        role = message.role
        content = message.content
        
        if role == "system":
            prompt_parts.append(f"<start_of_turn>system\n{content}<end_of_turn>")
        elif role == "user":
            prompt_parts.append(f"<start_of_turn>user\n{content}<end_of_turn>")
        elif role == "assistant":
            prompt_parts.append(f"<start_of_turn>model\n{content}<end_of_turn>")
    
    # Add the start for model response
    prompt_parts.append("<start_of_turn>model\n")
    
    return "\n".join(prompt_parts)

@asynccontextmanager
async def lifespan(app: FastAPI):
    """Application lifespan manager for startup and shutdown events"""
    global llm
    logger.info("🚀 Starting Gemma 3n GGUF Backend Service...")
    if os.environ.get("DEMO_MODE", "").strip() not in ("", "0", "false", "False"):
        logger.info("🧪 DEMO_MODE enabled: skipping model load")
        llm = None
        yield
        logger.info("🔄 Shutting down Gemma 3n Backend Service (demo mode)...")
        return
    
    if not llama_cpp_available:
        logger.error("❌ llama-cpp-python is not available. Please install with: pip install llama-cpp-python")
        raise RuntimeError("llama-cpp-python not available")
    
    try:
        logger.info(f"📥 Loading Gemma 3n GGUF model from {current_model}...")
        
        # Configure model parameters optimized for HF Spaces memory constraints
        llm = Llama.from_pretrained(
            repo_id=current_model,
            filename="*Q4_0.gguf",  # Use Q4_0 instead of Q4_K_M for lower memory usage
            verbose=True,
            # Memory-optimized settings for HF Spaces
            n_ctx=2048,  # Reduced context length to save memory (was 4096)
            n_threads=2,  # Fewer threads for lower memory usage (was 4)
            n_gpu_layers=0,  # Force CPU-only to avoid GPU memory issues
            # Additional memory optimizations
            n_batch=512,  # Smaller batch size to reduce memory peaks
            use_mmap=True,  # Use memory mapping to reduce RAM usage
            use_mlock=False,  # Don't lock memory pages
            low_vram=True,  # Enable low VRAM mode for additional memory savings
            # Chat template for Gemma 3n format
            chat_format="gemma",  # Try built-in gemma format first
        )
        
        logger.info("✅ Successfully loaded Gemma 3n GGUF model with memory optimizations")
        
    except Exception as e:
        logger.error(f"❌ Failed to initialize Gemma 3n model: {e}")
        logger.warning("⚠️ Please download the GGUF model file locally and update the path")
        logger.warning("⚠️ You can download from: https://huggingface.co/unsloth/gemma-3n-E4B-it-GGUF")
        
        # For demo purposes, we'll continue without the model
        logger.info("🔄 Starting service in demo mode (responses will be mocked)")
    
    yield
    logger.info("🔄 Shutting down Gemma 3n Backend Service...")
    if llm:
        # Clean up model resources
        llm = None

# Initialize FastAPI app
app = FastAPI(
    title="Gemma 3n GGUF Backend Service",
    description="OpenAI-compatible chat completion API powered by Gemma-3n-E4B-it-GGUF",
    version="1.0.0",
    lifespan=lifespan
)

# Add CORS middleware
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],  # Configure appropriately for production
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

def ensure_model_ready():
    """Check if model is loaded and ready"""
    # For demo mode, we'll allow the service to run even without a model
    pass

def generate_response_gguf(messages: List[ChatMessage], max_tokens: int = 256, temperature: float = 1.0, top_p: float = 0.95, top_k: int = 64) -> str:
    """Generate response using GGUF model via llama-cpp-python (memory-optimized)."""
    if llm is None:
        # Demo mode response
        return "🤖 Demo mode: Gemma 3n model not loaded. This would be a real response from the Gemma 3n model. Please download the GGUF model from https://huggingface.co/unsloth/gemma-3n-E4B-it-GGUF"
    
    # Limit max_tokens for memory efficiency on HF Spaces
    max_tokens = min(max_tokens, 512)  # Cap at 512 tokens max
    
    try:
        # Use the chat completion method if available
        if hasattr(llm, 'create_chat_completion'):
            # Convert to dict format for llama-cpp-python
            messages_dict = [{"role": msg.role, "content": msg.content} for msg in messages]
            
            response = llm.create_chat_completion(
                messages=messages_dict,
                max_tokens=max_tokens,
                temperature=temperature,
                top_p=top_p,
                top_k=top_k,
                stop=["<end_of_turn>", "<eos>", "</s>"]  # Gemma 3n stop tokens
            )
            
            return response['choices'][0]['message']['content'].strip()
        
        else:
            # Fallback to direct prompt completion
            prompt = convert_messages_to_gemma_prompt(messages)
            
            response = llm(
                prompt,
                max_tokens=max_tokens,
                temperature=temperature,
                top_p=top_p,
                top_k=top_k,
                stop=["<end_of_turn>", "<eos>", "</s>"],
                echo=False
            )
            
            return response['choices'][0]['text'].strip()
            
    except Exception as e:
        logger.error(f"GGUF generation failed: {e}")
        return "I apologize, but I'm having trouble generating a response right now. Please try again."

@app.get("/", response_class=JSONResponse)
async def root() -> Dict[str, Any]:
    """Root endpoint with service information"""
    return {
        "message": "Gemma 3n GGUF Backend Service is running!",
        "model": current_model,
        "version": "1.0.0",
        "backend": "llama-cpp-python",
        "model_loaded": llm is not None,
        "endpoints": {
            "health": "/health",
            "chat_completions": "/v1/chat/completions"
        }
    }

@app.get("/health", response_model=HealthResponse)
async def health_check():
    """Health check endpoint"""
    return HealthResponse(
        status="healthy" if (llm is not None) else "demo_mode",
        model=current_model,
        version="1.0.0",
        backend="llama-cpp-python"
    )

@app.post("/v1/chat/completions", response_model=ChatCompletionResponse)
async def create_chat_completion(
    request: ChatCompletionRequest
) -> ChatCompletionResponse:
    """Create a chat completion (OpenAI-compatible) using Gemma 3n GGUF"""
    try:
        ensure_model_ready()
        
        if not request.messages:
            raise HTTPException(status_code=400, detail="Messages cannot be empty")
        
        logger.info(f"Generating Gemma 3n response for {len(request.messages)} messages")
        
        response_text = generate_response_gguf(
            request.messages,
            request.max_tokens or 512,
            request.temperature or 1.0,
            request.top_p or 0.95,
            request.top_k or 64
        )
        
        response_text = response_text.strip() if response_text else "No response generated."
        
        return ChatCompletionResponse(
            id=f"chatcmpl-{int(time.time())}",
            created=int(time.time()),
            model=request.model,
            choices=[ChatCompletionChoice(
                index=0,
                message=ChatMessage(role="assistant", content=response_text),
                finish_reason="stop"
            )]
        )
        
    except Exception as e:
        logger.error(f"Error in chat completion: {e}")
        raise HTTPException(status_code=500, detail=f"Internal server error: {str(e)}")

# -----------------------------
# Training Job Management (Unsloth)
# -----------------------------

# Persistent job store: in-memory dict backed by SQLite
TRAIN_JOBS: Dict[str, Dict[str, Any]] = {}
# Initialize SQLite DB for job persistence
DB_PATH = Path(os.environ.get("JOB_DB_PATH", "./jobs.db"))
conn = sqlite3.connect(str(DB_PATH), check_same_thread=False)
cursor = conn.cursor()
cursor.execute(
    """
    CREATE TABLE IF NOT EXISTS jobs (
        job_id TEXT PRIMARY KEY,
        data   TEXT NOT NULL
    )
    """
)
conn.commit()

def load_jobs() -> None:
    cursor.execute("SELECT job_id, data FROM jobs")
    for job_id, data in cursor.fetchall():
        TRAIN_JOBS[job_id] = json.loads(data)

def save_job(job_id: str) -> None:
    cursor.execute(
        "INSERT OR REPLACE INTO jobs (job_id, data) VALUES (?, ?)",
        (job_id, json.dumps(TRAIN_JOBS[job_id]))
    )
    conn.commit()

# Load existing jobs on startup
load_jobs()

TRAIN_DIR = Path(os.environ.get("TRAIN_DIR", "./training_runs")).resolve()
TRAIN_DIR.mkdir(parents=True, exist_ok=True)
# Maximum concurrent training jobs
MAX_CONCURRENT_JOBS = int(os.environ.get("MAX_CONCURRENT_JOBS", "5"))

def _start_training_subprocess(job_id: str, args: Dict[str, Any]) -> subprocess.Popen[Any]:
    """Spawn a subprocess to run the Unsloth fine-tuning script."""
    logs_dir = TRAIN_DIR / job_id
    logs_dir.mkdir(parents=True, exist_ok=True)
    log_file = open(logs_dir / "train.log", "w", encoding="utf-8")
    # Store log file handle to close later
    TRAIN_JOBS.setdefault(job_id, {})["log_file"] = log_file
    save_job(job_id)

    # Build absolute script path to avoid module/package resolution issues
    script_path = (Path(__file__).parent / "training" / "train_gemma_unsloth.py").resolve()
    # Verify training script exists
    if not script_path.exists():
        logger.error(f"Training script not found at {script_path}")
        raise HTTPException(status_code=500, detail=f"Training script not found at {script_path}")
    python_exec = sys.executable

    cmd = [
        python_exec,
        str(script_path),
        "--job-id", job_id,
        "--output-dir", str(logs_dir),
    ]

    # Optional user-specified args
    def _extend(k: str, v: Any):
        if v is None:
            return
        if isinstance(v, bool):
            cmd.extend([f"--{k}"] if v else [])
        else:
            cmd.extend([f"--{k}", str(v)])

    _extend("dataset", args.get("dataset"))
    _extend("text-field", args.get("text_field"))
    _extend("prompt-field", args.get("prompt_field"))
    _extend("response-field", args.get("response_field"))
    _extend("max-steps", args.get("max_steps"))
    _extend("epochs", args.get("epochs"))
    _extend("lr", args.get("lr"))
    _extend("batch-size", args.get("batch_size"))
    _extend("gradient-accumulation", args.get("gradient_accumulation"))
    _extend("lora-r", args.get("lora_r"))
    _extend("lora-alpha", args.get("lora_alpha"))
    _extend("cutoff-len", args.get("cutoff_len"))
    _extend("model-id", args.get("model_id"))
    _extend("use-bf16", args.get("use_bf16"))
    _extend("use-fp16", args.get("use_fp16"))
    _extend("seed", args.get("seed"))
    _extend("dry-run", args.get("dry_run"))

    logger.info(f"🧵 Starting training subprocess for job {job_id}: {' '.join(cmd)}")
    logger.info(f"🐍 Using interpreter: {python_exec}")
    proc = subprocess.Popen(cmd, stdout=log_file, stderr=subprocess.STDOUT, cwd=str(Path(__file__).parent))
    return proc

def _watch_process(job_id: str, proc: subprocess.Popen[Any]):
    """Monitor a training process and update job state on exit."""
    return_code = proc.wait()
    status = "completed" if return_code == 0 else "failed"
    TRAIN_JOBS[job_id]["status"] = status
    TRAIN_JOBS[job_id]["return_code"] = return_code
    TRAIN_JOBS[job_id]["ended_at"] = int(time.time())
    # Persist updated job status
    save_job(job_id)
    # Close the log file handle to prevent resource leaks
    log_file = TRAIN_JOBS[job_id].get("log_file")
    if log_file:
        try:
            log_file.close()
        except Exception as close_err:
            logger.warning(f"Failed to close log file for job {job_id}: {close_err}")
    logger.info(f"🏁 Training job {job_id} finished with status={status}, code={return_code}")

class StartTrainingRequest(BaseModel):
    dataset: str = Field(..., description="HF dataset name or path to local JSONL/JSON file")
    model_id: Optional[str] = Field(default="unsloth/gemma-3n-E4B-it", description="Base model for training (HF Transformers format)")
    text_field: Optional[str] = Field(default=None, description="Single text field name (SFT)")
    prompt_field: Optional[str] = Field(default=None, description="Prompt/instruction field (chat data)")
    response_field: Optional[str] = Field(default=None, description="Response/output field (chat data)")
    max_steps: Optional[int] = Field(default=None)
    epochs: Optional[int] = Field(default=1)
    lr: Optional[float] = Field(default=2e-4)
    batch_size: Optional[int] = Field(default=1)
    gradient_accumulation: Optional[int] = Field(default=8)
    lora_r: Optional[int] = Field(default=16)
    lora_alpha: Optional[int] = Field(default=32)
    cutoff_len: Optional[int] = Field(default=4096)
    use_bf16: Optional[bool] = Field(default=True)
    use_fp16: Optional[bool] = Field(default=False)
    seed: Optional[int] = Field(default=42)
    dry_run: Optional[bool] = Field(default=False, description="Write DONE and exit without running (for CI/macOS)")

class StartTrainingResponse(BaseModel):
    job_id: str
    status: str
    output_dir: str

class TrainStatusResponse(BaseModel):
    job_id: str
    status: str
    created_at: int
    started_at: Optional[int] = None
    ended_at: Optional[int] = None
    output_dir: Optional[str] = None
    return_code: Optional[int] = None

@app.post("/train/start", response_model=StartTrainingResponse)
def start_training(req: StartTrainingRequest):
    """Start a background Unsloth fine-tuning job. Returns a job_id to poll."""
    # Enforce maximum concurrent training jobs
    running_jobs = sum(1 for job in TRAIN_JOBS.values() if job.get("status") == "running")
    if running_jobs >= MAX_CONCURRENT_JOBS:
        raise HTTPException(
            status_code=429,
            detail=f"Maximum concurrent training jobs reached ({MAX_CONCURRENT_JOBS}). Try again later."
        )
    job_id = uuid.uuid4().hex[:12]
    now = int(time.time())
    output_dir = str((TRAIN_DIR / job_id).resolve())
    TRAIN_JOBS[job_id] = {
        "status": "starting",
        "created_at": now,
        "started_at": now,
        "args": req.model_dump(),
        "output_dir": output_dir,
    }
    save_job(job_id)

    try:
        proc = _start_training_subprocess(job_id, req.model_dump())
        TRAIN_JOBS[job_id]["status"] = "running"
        TRAIN_JOBS[job_id]["pid"] = proc.pid
        save_job(job_id)
        watcher = threading.Thread(target=_watch_process, args=(job_id, proc), daemon=True)
        watcher.start()
        return StartTrainingResponse(job_id=job_id, status="running", output_dir=output_dir)
    except Exception as e:
        logger.exception("Failed to start training job")
        TRAIN_JOBS[job_id]["status"] = "failed_to_start"
        save_job(job_id)
        raise HTTPException(status_code=500, detail=f"Failed to start training: {e}")

@app.get("/train/status/{job_id}", response_model=TrainStatusResponse)
def train_status(job_id: str):
    job = TRAIN_JOBS.get(job_id)
    if not job:
        raise HTTPException(status_code=404, detail="Job not found")
    return TrainStatusResponse(
        job_id=job_id,
        status=job.get("status", "unknown"),
        created_at=job.get("created_at", 0),
        started_at=job.get("started_at"),
        ended_at=job.get("ended_at"),
        output_dir=job.get("output_dir"),
        return_code=job.get("return_code"),
    )

@app.get("/train/logs/{job_id}")
def train_logs(
    job_id: str,
    tail: int = Query(200, ge=0, le=1000, description="Number of lines to tail, between 0 and 1000"),
):
    job = TRAIN_JOBS.get(job_id)
    if not job:
        raise HTTPException(status_code=404, detail="Job not found")
    log_path = Path(job["output_dir"]) / "train.log"
    if not log_path.exists():
        return {"job_id": job_id, "logs": "(no logs yet)"}
    try:
        with open(log_path, "r", encoding="utf-8", errors="ignore") as f:
            lines = f.readlines()[-tail:]
        return {"job_id": job_id, "logs": "".join(lines)}
    except Exception as e:
        raise HTTPException(status_code=500, detail=f"Failed to read logs: {e}")

@app.post("/train/stop/{job_id}")
def train_stop(job_id: str):
    job = TRAIN_JOBS.get(job_id)
    if not job:
        raise HTTPException(status_code=404, detail="Job not found")
    pid = job.get("pid")
    if not pid:
        raise HTTPException(status_code=400, detail="Job does not have an active PID")
    try:
        os.kill(pid, signal.SIGTERM)
    except ProcessLookupError:
        logger.warning(
            f"Process {pid} for job {job_id} not found; may have exited already"
        )
        job["status"] = "stopping_failed"
        save_job(job_id)
        return {"job_id": job_id, "status": job["status"]}
    except Exception as e:
        raise HTTPException(status_code=500, detail=f"Failed to stop job: {e}")
    else:
        job["status"] = "stopping"
        save_job(job_id)
        return {"job_id": job_id, "status": "stopping"}

# Main entry point
if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8000)