app.py

# app.py
import os, shutil, subprocess, zipfile, traceback, time, io
from pathlib import Path
from datetime import datetime
import gradio as gr

ROOT = Path(__file__).resolve().parent
DATA = ROOT / "dataset.jsonl"
LOG  = ROOT / "train.log"
RUNS = ROOT / "runs"
RUNS.mkdir(exist_ok=True)

# -------- logging helpers --------
def append_log(msg: str):
    msg = msg.rstrip()
    with open(LOG, "a", encoding="utf-8") as lf:
        lf.write(msg + "\n")

def read_logs():
    return LOG.read_text(encoding="utf-8")[-20000:] if LOG.exists() else "⏳ Waiting…"

# -------- workspace + models --------
def ls_workspace() -> str:
    rows = []
    for p in sorted(ROOT.iterdir(), key=lambda x: (x.is_file(), x.name.lower())):
        try:
            size = p.stat().st_size
        except Exception:
            size = 0
        rows.append(f"{'[DIR]' if p.is_dir() else '     '}\t{size:>10}\t{p.name}")
    return "\n".join(rows) or "(empty)"

def list_models():
    out = []
    for base in [ROOT, RUNS]:
        if not base.exists():
            continue
        for p in base.iterdir():
            if p.is_dir() and (p / "config.json").exists() and (
                (p / "tokenizer.json").exists() or (p / "tokenizer_config.json").exists()
            ):
                out.append(str(p))
    return sorted(set(out))

def dropdown_update_safe(models, prefer=None):
    val = prefer if (prefer and prefer in models) else (models[0] if models else None)
    return gr.update(choices=models, value=val)

# -------- dataset upload --------
def upload_dataset(file):
    if not file:
        return "❌ No file selected.", ls_workspace()
    if hasattr(file, "name") and os.path.isfile(file.name):
        shutil.copy(file.name, DATA)
        return f"✅ Uploaded → {DATA.name}", ls_workspace()
    return "⚠ Unexpected item; please upload a .jsonl file.", ls_workspace()

# -------- training (LIVE LOGS) --------
def start_training_live(run_name):
    """
    Streams training logs to the UI while the subprocess runs.
    Yields tuples for outputs: [status, download_file, workspace, logs, model_dropdown]
    """
    run_id = (run_name or "").strip() or datetime.now().strftime("run_%Y%m%d_%H%M%S")
    out_dir  = RUNS / run_id
    zip_path = RUNS / f"{run_id}.zip"

    # clean only this run
    if out_dir.exists():
        shutil.rmtree(out_dir, ignore_errors=True)
    if zip_path.exists():
        zip_path.unlink()

    # init log
    LOG.write_text(f"🔥 Training started…\nRun: {run_id}\n", encoding="utf-8")
    append_log(f"Workspace:\n{ls_workspace()}")

    cmd = [
        "python", str(ROOT / "train.py"),
        "--dataset", str(DATA),
        "--output",  str(out_dir),
        "--zip_path", str(zip_path),
        "--model_name", "Salesforce/codegen-350M-multi",
        "--epochs", "1",
        "--batch_size", "2",
        "--block_size", "256",
        "--learning_rate", "5e-5",
    ]
    append_log("▶ " + " ".join(cmd))

    # start subprocess with live stdout
    proc = subprocess.Popen(
        cmd,
        stdout=subprocess.PIPE,
        stderr=subprocess.STDOUT,
        bufsize=1,
        universal_newlines=True,
        encoding="utf-8",
        errors="replace",
    )

    live_log = io.StringIO()
    status_msg = f"🚀 Training run '{run_id}' in progress…"
    # stream loop
    while True:
        line = proc.stdout.readline()
        if line == "" and proc.poll() is not None:
            break
        if line:
            append_log(line.rstrip("\n"))
            live_log.write(line)
            # Trim to last ~20k chars for UI
            text = live_log.getvalue()[-20000:]
            # yield with download hidden (until zip exists)
            yield (
                status_msg,
                gr.update(value=None, visible=False),
                ls_workspace(),
                text,
                dropdown_update_safe(list_models(), prefer=None),
            )
            # if zip appears during training (e.g., early save), surface it
            if zip_path.exists():
                yield (
                    "📦 Model zip created during run.",
                    gr.update(value=str(zip_path), visible=True),
                    ls_workspace(),
                    text,
                    dropdown_update_safe(list_models(), prefer=None),
                )

    code = proc.wait()

    models = list_models()
    model_update = dropdown_update_safe(models, prefer=str(out_dir) if out_dir.exists() else None)
    final_logs = read_logs()

    if code == 0 and zip_path.exists():
        info = f"✅ Training complete. Saved: {out_dir.name} | Zip: {zip_path.name}"
        append_log(info)
        yield (info, gr.update(value=str(zip_path), visible=True), ls_workspace(), final_logs, model_update)
    else:
        info = f"❌ Training failed (exit {code}). Check logs below."
        append_log(info)
        yield (info, gr.update(value=None, visible=False), ls_workspace(), final_logs, model_update)

def refresh_download():
    zips = sorted(RUNS.glob("*.zip"), key=lambda p: p.stat().st_mtime, reverse=True)
    latest = zips[0] if zips else None
    models = list_models()
    return (
        gr.update(value=(str(latest) if latest else None), visible=bool(latest)),
        ls_workspace(),
        dropdown_update_safe(models)
    )

# -------- import a zip as a model folder --------
def import_zip(zfile):
    if not zfile:
        return "❌ No zip selected.", list_models()
    dest = ROOT / "imported_model"
    if dest.exists():
        shutil.rmtree(dest, ignore_errors=True)
    dest.mkdir(parents=True, exist_ok=True)
    with zipfile.ZipFile(zfile.name, "r") as z:
        z.extractall(dest)
    return f"✅ Imported to {dest.name}", list_models()

# -------- generation: cached pipeline --------
_GEN_CACHE = {"path": None, "pipe": None}

def get_generation_pipeline(model_path: str):
    from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
    import torch

    if _GEN_CACHE["path"] == model_path and _GEN_CACHE["pipe"] is not None:
        return _GEN_CACHE["pipe"]

    append_log(f"🧩 Loading pipeline from: {model_path}")
    tok = AutoTokenizer.from_pretrained(model_path, use_fast=True)
    if tok.pad_token_id is None:
        if tok.eos_token_id is not None:
            tok.pad_token = tok.eos_token
            append_log("ℹ No pad_token; using eos_token as pad_token.")
        else:
            tok.add_special_tokens({"pad_token": "[PAD]"})
            append_log("ℹ Added [PAD] token to tokenizer.")
    model = AutoModelForCausalLM.from_pretrained(model_path)
    if getattr(model, "config", None) and getattr(model.config, "vocab_size", None) and len(tok) > model.config.vocab_size:
        model.resize_token_embeddings(len(tok))
        append_log(f"ℹ Resized embeddings to {len(tok)}.")

    pipe = pipeline(
        "text-generation",
        model=model,
        tokenizer=tok,
        device_map="auto" if torch.cuda.is_available() else None,
    )
    _GEN_CACHE["path"] = model_path
    _GEN_CACHE["pipe"] = pipe
    append_log("✅ Pipeline loaded.")
    return pipe

def generate(model_path, prompt):
    from pathlib import Path

    # Coerce Dropdown value (can be list)
    if isinstance(model_path, list):
        model_path = model_path[0] if model_path else None

    # validate
    if not model_path:
        return "❌ Select a model from the dropdown first."
    if not isinstance(model_path, str):
        # fix minor bug: _name_ not name
        return f"❌ Invalid model path type: {type(model_path)._name_}"
    if not Path(model_path).exists():
        return f"❌ Model folder not found: {model_path}"
    if not prompt or not prompt.strip():
        return "❌ Enter a prompt."

    try:
        pipe = get_generation_pipeline(model_path)
        append_log(f"📝 Generating for prompt ({len(prompt)} chars)…")
        out = pipe(
            prompt.strip(),
            max_new_tokens=120,
            do_sample=True,
            temperature=0.4,
            top_p=0.9,
            repetition_penalty=1.15,
            no_repeat_ngram_size=4,
            truncation=True
        )[0]["generated_text"]
        append_log("✅ Generation OK.")
        return out
    except Exception as e:
        tb = traceback.format_exc()
        append_log("❌ Generation error:\n" + tb)
        return "❌ Error during generation:\n" + "".join(traceback.format_exception_only(type(e), e))

# -------- UI --------
with gr.Blocks(title="Python AI — Train & Test") as app:
    gr.Markdown("## 🧠 Python AI — Train & Test\n• Unique runs • Safe download • Cached generation • Live logs\n")

    # Test first (so Train can update its dropdown)
    with gr.Tab("Test"):
        gr.Markdown("### Choose a model folder or upload a .zip, then prompt it")
        refresh_btn = gr.Button("↻ Refresh Model List")
        model_list = gr.Dropdown(
            choices=list_models(),
            label="Available AIs",
            interactive=True,
            allow_custom_value=True,  # keeps UI quiet when empty
            multiselect=False         # force single selection
        )
        zip_in = gr.File(label="Or upload a model .zip", file_types=[".zip"])
        import_status = gr.Textbox(label="Import Status", interactive=False)
        prompt = gr.Textbox(label="Prompt", lines=8, placeholder="### Instruction:\nPython: write a function ...\n### Response:\n")
        go = gr.Button("Generate")
        out = gr.Textbox(label="AI Response", lines=20)

    with gr.Tab("Train"):
        with gr.Row():
            ds = gr.File(label="📥 Upload JSONL", file_types=[".jsonl"])
            ws = gr.Textbox(label="Workspace", lines=16, value=ls_workspace())
        run_name = gr.Textbox(label="Run name (optional)", placeholder="e.g., python_small_v1")
        up_status = gr.Textbox(label="Upload Status", interactive=False)
        start = gr.Button("🚀 Start Training (Live Logs)", variant="primary")
        logs = gr.Textbox(label="📜 Training Logs (live)", lines=18, value=read_logs())
        status = gr.Textbox(label="Status", interactive=False)
        download_file = gr.File(label="📦 Latest trained zip", visible=False)
        refresh_dl_btn = gr.Button("Refresh Download")

    # wiring
    ds.change(upload_dataset, inputs=ds, outputs=[up_status, ws])

    # STREAMED training: function yields updates
    start.click(
        start_training_live,
        inputs=[run_name],
        outputs=[status, download_file, ws, logs, model_list]
    )

    refresh_dl_btn.click(
        refresh_download,
        outputs=[download_file, ws, model_list]
    )
    refresh_btn.click(lambda: dropdown_update_safe(list_models()), outputs=model_list)
    zip_in.change(import_zip, inputs=zip_in, outputs=[import_status, model_list])
    go.click(generate, inputs=[model_list, prompt], outputs=out)

app.queue(default_concurrency_limit=1).launch()