model/model_runner.py

# model_runner.py
import os
import sys
from typing import List, Optional
from llama_cpp import Llama

print(f"[BOOT] model_runner from {__file__}", file=sys.stderr)

# ---- Phase 2: flags (no behavior change) ------------------------------------
# Reads LAB_* env toggles; all defaults preserve current behavior.
try:
    from utils import flags  # if your package path is different, adjust import
except Exception:
    # Fallback inline flags if utils.flags isn't available in this lab
    def _as_bool(val: Optional[str], default: bool) -> bool:
        if val is None:
            return default
        return val.strip().lower() in {"1", "true", "yes", "on", "y", "t"}
    class _F:
        SANITIZE_ENABLED = _as_bool(os.getenv("LAB_SANITIZE_ENABLED"), False)   # you don't sanitize today
        STOPSEQ_ENABLED  = _as_bool(os.getenv("LAB_STOPSEQ_ENABLED"),  False)   # extra stops only; defaults off
        CRITIC_ENABLED   = _as_bool(os.getenv("LAB_CRITIC_ENABLED"),   False)
        JSON_MODE        = _as_bool(os.getenv("LAB_JSON_MODE"),        False)
        EVIDENCE_GATE    = _as_bool(os.getenv("LAB_EVIDENCE_GATE"),    False)
        @staticmethod
        def snapshot():
            return {
                "LAB_SANITIZE_ENABLED": _F.SANITIZE_ENABLED,
                "LAB_STOPSEQ_ENABLED":  _F.STOPSEQ_ENABLED,
                "LAB_CRITIC_ENABLED":   _F.CRITIC_ENABLED,
                "LAB_JSON_MODE":        _F.JSON_MODE,
                "LAB_EVIDENCE_GATE":    _F.EVIDENCE_GATE,
            }
    flags = _F()

print("[flags] snapshot:", getattr(flags, "snapshot", lambda: {} )(), file=sys.stderr)

# Optional sanitizer hook (kept no-op unless enabled later)
def _sanitize(text: str) -> str:
    # Phase 2: default False -> no behavior change
    if getattr(flags, "SANITIZE_ENABLED", False):
        # TODO: wire your real sanitizer in Phase 3+
        return text.strip()
    return text

# Stop sequences: keep today's defaults ALWAYS.
# If LAB_STOPSEQ_ENABLED=true, add *extra* stops from STOP_SEQUENCES env (comma-separated).
DEFAULT_STOPS: List[str] = ["\nUser:", "\nAssistant:"]

def _extra_stops_from_env() -> List[str]:
    if not getattr(flags, "STOPSEQ_ENABLED", False):
        return []
    raw = os.getenv("STOP_SEQUENCES", "")
    toks = [t.strip() for t in raw.split(",") if t.strip()]
    return toks

# ---- Model cache / load ------------------------------------------------------
_model = None  # module-level cache

def load_model():
    global _model
    if _model is not None:
        return _model

    model_path = os.getenv("MODEL_PATH")
    if not model_path or not os.path.exists(model_path):
        raise ValueError(f"Model path does not exist or is not set: {model_path}")

    print(f"[INFO] Loading model from {model_path}")

    _model = Llama(
        model_path=model_path,
        n_ctx=1024,        # short context to reduce memory use
        n_threads=4,       # number of CPU threads
        n_gpu_layers=0     # CPU only (Hugging Face free tier)
    )
    return _model

# ---- Inference ---------------------------------------------------------------
def generate(prompt: str, max_tokens: int = 256) -> str:
    model = load_model()

    # Preserve existing default stops; optionally extend with extra ones if flag is on
    stops = DEFAULT_STOPS + _extra_stops_from_env()

    output = model(
        prompt,
        max_tokens=max_tokens,
        stop=stops,            # unchanged defaults; may include extra stops if enabled
        echo=False,
        temperature=0.7,
        top_p=0.95,
    )
    raw_text = output["choices"][0]["text"]

    # Preserve current manual truncation by the same default stops (kept intentionally)
    # Extra stops are also applied here if enabled for consistency.
    for stop_token in stops:
        if stop_token and stop_token in raw_text:
            raw_text = raw_text.split(stop_token)[0]

    final = _sanitize(raw_text)
    return final.strip()