magenta-retry

Runtime error

fastrtc issues solved mostly in gradio app. html web tester still better ux for the moment
6cd5ac5 9 months ago
67.4 kB
import os

# very top of app.py, before importing tensorflow/jax/magenta stuff
try:
    import pylibsrtp
    # This is the key: call srtp_init via cffi binding
    from pylibsrtp import _binding
    _binding.lib.srtp_init()
    print("SRTP init OK (pre-TF)", flush=True)
except Exception as e:
    print("SRTP init failed early:", e, flush=True)

# ---- Space mode gating (place above any JAX import!) ----
SPACE_MODE = os.getenv("SPACE_MODE")
if SPACE_MODE is None:
    try:
        import jax
        SPACE_MODE = "serve" if any(getattr(d, "platform", "") in ("gpu","cuda","rocm") for d in jax.devices()) else "template"
    except Exception:
        SPACE_MODE = "template"



if SPACE_MODE != "serve":
    # In template mode, force JAX to CPU so it won't try to load CUDA plugins
    os.environ.setdefault("JAX_PLATFORMS", "cpu")
else:
    # Only set GPU-friendly XLA flags when we actually intend to serve on GPU
    os.environ.setdefault(
        "XLA_FLAGS",
        " ".join([
            "--xla_gpu_enable_triton_gemm=true",
            "--xla_gpu_enable_latency_hiding_scheduler=true",
            "--xla_gpu_autotune_level=2",
        ])
    )

# Optional: persist JAX compile cache across restarts (reduces warmup time)
os.environ.setdefault("JAX_CACHE_DIR", "/home/appuser/.cache/jax")

import jax
# ✅ Valid choices include: "default", "high", "highest", "tensorfloat32", "float32", etc.
# TF32 is the sweet spot on Ampere/Ada GPUs for ~1.1–1.3× matmul speedups.
try:
    jax.config.update("jax_default_matmul_precision", "tensorfloat32")
except Exception:
    jax.config.update("jax_default_matmul_precision", "high")  # older alias

# Initialize the on-disk compilation cache (best-effort)
try:
    from jax.experimental.compilation_cache import compilation_cache as cc
    cc.initialize_cache(os.environ["JAX_CACHE_DIR"])
except Exception:
    pass
# --------------------------------------------------------------------



from magenta_rt import system, audio as au
import numpy as np
from fastapi import FastAPI, UploadFile, File, Form, Body, HTTPException, Response, Request, WebSocket, WebSocketDisconnect, Query
from fastapi.responses import JSONResponse, FileResponse, HTMLResponse
import tempfile, io, base64, math, threading
from fastapi.middleware.cors import CORSMiddleware
from contextlib import contextmanager
import soundfile as sf
from math import gcd
from scipy.signal import resample_poly
from utils import (
    match_loudness_to_reference, stitch_generated, hard_trim_seconds,
    apply_micro_fades, make_bar_aligned_context, take_bar_aligned_tail,
    resample_and_snap, wav_bytes_base64
)

from jam_worker import JamWorker, JamParams, JamChunk
from one_shot_generation import generate_loop_continuation_with_mrt, generate_style_only_with_mrt

import uuid, threading

import logging

import gradio as gr
from typing import Optional, Union, Literal


import json, asyncio, base64
import time



from starlette.websockets import WebSocketState
try:
    from uvicorn.protocols.utils import ClientDisconnected  # uvicorn >= 0.20
except Exception:
    class ClientDisconnected(Exception):  # fallback
        pass

import re, tarfile
from pathlib import Path
from huggingface_hub import snapshot_download, HfApi

from pydantic import BaseModel

from model_management import CheckpointManager, AssetManager, ModelSelector, ModelSelect

# from fastrtc_magenta import create_magenta_stream, FASTRTC_AVAILABLE

def _gpu_probe() -> dict:
    """
    Returns:
      {
        "ok": bool,
        "backend": str | None,        # "gpu" | "cpu" | "tpu" | None
        "has_gpu": bool,
        "devices": list[str],         # e.g. ["gpu:0", "gpu:1"]
        "error": str | None,
      }
    """
    try:
        import jax
        try:
            backend = jax.default_backend()  # "gpu", "cpu", "tpu"
        except Exception:
            from jax.lib import xla_bridge
            backend = getattr(xla_bridge.get_backend(), "platform", None)

        try:
            devices = jax.devices()
            has_gpu = any(getattr(d, "platform", "") in ("gpu", "cuda", "rocm") for d in devices)
            dev_list = [f"{getattr(d, 'platform', '?')}:{getattr(d, 'id', '?')}" for d in devices]
            return {"ok": True, "backend": backend, "has_gpu": has_gpu, "devices": dev_list, "error": None}
        except Exception as e:
            return {"ok": False, "backend": backend, "has_gpu": False, "devices": [], "error": f"jax.devices failed: {e}"}
    except Exception as e:
        return {"ok": False, "backend": None, "has_gpu": False, "devices": [], "error": f"jax import failed: {e}"}

# ---- Finetune assets (mean & centroids) --------------------------------------
# _FINETUNE_REPO_DEFAULT = os.getenv("MRT_ASSETS_REPO", "thepatch/magenta-ft")
_ASSETS_REPO_ID: str | None = None
_MEAN_EMBED: np.ndarray | None = None           # shape (D,) dtype float32
_CENTROIDS: np.ndarray | None = None            # shape (K, D) dtype float32

# _STEP_RE = re.compile(r"(?:^|/)checkpoint_(\d+)(?:/|\.tar\.gz|\.tgz)?$")

# Create instances (these don't modify globals)
asset_manager = AssetManager()
model_selector = ModelSelector(CheckpointManager(), asset_manager)

def _sync_assets_globals_from_manager():
    # Keeps /model/config in sync with what the asset manager has
    global _MEAN_EMBED, _CENTROIDS, _ASSETS_REPO_ID
    _MEAN_EMBED = asset_manager.mean_embed
    _CENTROIDS = asset_manager.centroids
    _ASSETS_REPO_ID = asset_manager.assets_repo_id

def _any_jam_running() -> bool:
    with jam_lock:
        return any(w.is_alive() for w in jam_registry.values())

def _stop_all_jams(timeout: float = 5.0):
    with jam_lock:
        for sid, w in list(jam_registry.items()):
            if w.is_alive():
                w.stop()
                w.join(timeout=timeout)
                jam_registry.pop(sid, None)


async def send_json_safe(ws: WebSocket, obj) -> bool:
    """Try to send. Returns False if the socket is (or becomes) closed."""
    if ws.client_state == WebSocketState.DISCONNECTED or ws.application_state == WebSocketState.DISCONNECTED:
        return False
    try:
        await ws.send_text(json.dumps(obj))
        return True
    except (WebSocketDisconnect, ClientDisconnected, RuntimeError):
        return False
    except Exception:
        return False

# --- Patch T5X mesh helpers for GPUs on JAX >= 0.7 (coords present, no core_on_chip) ---
def _patch_t5x_for_gpu_coords():
    try:
        import jax
        from t5x import partitioning as _t5x_part

        old_bounds = getattr(_t5x_part, "bounds_from_last_device", None)
        old_getcoords = getattr(_t5x_part, "get_coords", None)

        def _bounds_from_last_device_gpu_safe(last_device):
            # TPU: coords + core_on_chip
            core = getattr(last_device, "core_on_chip", None)
            coords = getattr(last_device, "coords", None)
            if coords is not None and core is not None:
                x, y, z = coords
                return x + 1, y + 1, z + 1, core + 1
            # Non-TPU (or GPU lacking core_on_chip): hosts x local_devices
            return jax.host_count(), jax.local_device_count()

        def _get_coords_gpu_safe(device):
            core = getattr(device, "core_on_chip", None)
            coords = getattr(device, "coords", None)
            if coords is not None and core is not None:
                return (*coords, core)
            # Fallback that works on CPU/GPU
            return (device.process_index, device.id % jax.local_device_count())

        _t5x_part.bounds_from_last_device = _bounds_from_last_device_gpu_safe
        _t5x_part.get_coords = _get_coords_gpu_safe
        import logging; logging.info("Patched t5x.partitioning for GPU coords without core_on_chip.")
    except Exception as e:
        import logging; logging.exception("t5x GPU-coords patch failed: %s", e)

def _patch_magenta_rt_asset_fetch():
    """
    Patch magenta_rt.asset._fetch_single_hf to handle None response gracefully.
    Prevents AttributeError when network timeouts occur.
    """
    try:
        from magenta_rt import asset
        import logging
        
        # Save original function
        _original_fetch = asset._fetch_single_hf
        
        def _fetch_single_hf_safe(*args, **kwargs):
            """Wrapper that fixes the None response.status_code bug"""
            try:
                return _original_fetch(*args, **kwargs)
            except Exception as e:
                # This is the bug fix: check if response exists before accessing it
                response = getattr(e, 'response', None)
                if response is not None and hasattr(response, 'status_code'):
                    if response.status_code == 429:
                        # Original code's rate-limit handling would go here
                        logging.warning("Rate limited by HuggingFace Hub")
                        raise
                # For all other cases (including timeout with no response), re-raise
                logging.error(f"HuggingFace download failed: {e}")
                raise
        
        # Apply the patch
        asset._fetch_single_hf = _fetch_single_hf_safe
        logging.info("Patched magenta_rt.asset._fetch_single_hf for safer error handling.")
        
    except Exception as e:
        logging.exception("magenta_rt asset fetch patch failed: %s", e)

# Call the patch immediately at import time (before MagentaRT init)
_patch_t5x_for_gpu_coords()
_patch_magenta_rt_asset_fetch()

jam_registry: dict[str, JamWorker] = {}
jam_lock = threading.Lock()



@contextmanager
def mrt_overrides(mrt, **kwargs):
    """Temporarily set attributes on MRT if they exist; restore after."""
    old = {}
    try:
        for k, v in kwargs.items():
            if hasattr(mrt, k):
                old[k] = getattr(mrt, k)
                setattr(mrt, k, v)
        yield
    finally:
        for k, v in old.items():
            setattr(mrt, k, v)

# loudness utils
try:
    import pyloudnorm as pyln
    _HAS_LOUDNORM = True
except Exception:
    _HAS_LOUDNORM = False



def build_style_vector(
    mrt,
    *,
    text_styles: list[str] | None = None,
    text_weights: list[float] | None = None,
    loop_embed: np.ndarray | None = None,
    loop_weight: float | None = None,
    mean_weight: float | None = None,
    centroid_weights: list[float] | None = None,
) -> np.ndarray:
    """
    Returns a single style embedding combining:
      - loop embedding (optional)
      - one or more text style embeddings (optional)
      - mean finetune embedding (optional)
      - centroid embeddings (optional)
    All weights are normalized so they sum to 1 if > 0.
    """
    comps: list[np.ndarray] = []
    weights: list[float] = []

    # loop component
    if loop_embed is not None and (loop_weight or 0) > 0:
        comps.append(loop_embed.astype(np.float32, copy=False))
        weights.append(float(loop_weight))

    # text components
    if text_styles:
        for i, s in enumerate(text_styles):
            s = s.strip()
            if not s:
                continue
            w = 1.0
            if text_weights and i < len(text_weights):
                try: w = float(text_weights[i])
                except: w = 1.0
            if w <= 0: 
                continue
            e = mrt.embed_style(s)
            comps.append(e.astype(np.float32, copy=False))
            weights.append(w)

    # mean finetune
    if mean_weight and (_MEAN_EMBED is not None) and mean_weight > 0:
        comps.append(_MEAN_EMBED)
        weights.append(float(mean_weight))

    # centroid components
    if centroid_weights and _CENTROIDS is not None:
        K = _CENTROIDS.shape[0]
        for k, w in enumerate(centroid_weights[:K]):
            try: w = float(w)
            except: w = 0.0
            if w <= 0: 
                continue
            comps.append(_CENTROIDS[k])
            weights.append(w)

    if not comps:
        # fallback: neutral style if nothing provided
        return mrt.embed_style("")

    wsum = sum(weights)
    if wsum <= 0:
        return mrt.embed_style("")
    weights = [w/wsum for w in weights]

    # weighted sum
    out = np.zeros_like(comps[0], dtype=np.float32)
    for w, e in zip(weights, comps):
        out += w * e.astype(np.float32, copy=False)
    return out



# ----------------------------
# FastAPI app with lazy, thread-safe model init
# ----------------------------
app = FastAPI()

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],   # or lock to your domain(s)
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

_MRT = None
_MRT_LOCK = threading.Lock()

_PROGRESS = {}
_PROGRESS_LOCK = threading.Lock()

_GENERATE_COUNTER = 0
_GENERATE_COUNTER_LOCK = threading.Lock()

def _light_reset_mrt(mrt):
    """
    Lightweight reset that clears accumulated state without full recompilation.
    Should take <1 second instead of 30 seconds.
    """
    import logging
    logging.info("Performing light MRT reset after prolonged use...")
    
    try:
        # 1. Clear JAX device arrays (but not compiled functions)
        import jax
        for device in jax.devices():
            # Force garbage collection on device
            try:
                device.clear_memory()  # If available in your JAX version
            except AttributeError:
                pass
        
        # 2. Clear any MRT-level caches
        attrs_to_clear = ['_last_state', '_generation_cache', '_style_cache']
        for attr in attrs_to_clear:
            if hasattr(mrt, attr):
                setattr(mrt, attr, None)
        
        # 3. Clear codec internal state
        codec_attrs = [
            '_encode_state', '_decode_state', 
            '_encoder_cache', '_decoder_cache',
            '_buffer', '_frame_buffer'
        ]
        for attr in codec_attrs:
            if hasattr(mrt.codec, attr):
                setattr(mrt.codec, attr, None)
        
        # 4. Force Python garbage collection
        import gc
        gc.collect()
        
        # 5. If style model has cache, clear it
        if hasattr(mrt, 'style_model'):
            if hasattr(mrt.style_model, 'clear_cache'):
                mrt.style_model.clear_cache()
            # Clear any embedding caches
            for attr in ['_embed_cache', '_text_cache']:
                if hasattr(mrt.style_model, attr):
                    setattr(mrt.style_model, attr, None)
        
        logging.info("Light reset complete")
        return True
        
    except Exception as e:
        logging.warning(f"Light reset partially failed (non-fatal): {e}")
        return False

def _progress_update(req_id: str, n: int, total: int, stage: str = "generating"):
    if not req_id: 
        return
    with _PROGRESS_LOCK:
        _PROGRESS[req_id] = {
            "n": int(n),
            "total": int(total),
            "percent": int(round(100.0 * max(0, min(n, total)) / max(1, total))),
            "stage": stage,
            "ts": time.time(),
        }

def _progress_done(req_id: str):
    if not req_id: 
        return
    with _PROGRESS_LOCK:
        st = _PROGRESS.get(req_id, {})
        total = st.get("total") or st.get("n") or 1
        _PROGRESS[req_id] = {"n": total, "total": total, "percent": 100, "stage": "done", "ts": time.time()}

def _progress_get(req_id: str):
    with _PROGRESS_LOCK:
        return _PROGRESS.get(req_id, {"percent": 0, "stage": "pending"})

def get_mrt():
    global _MRT
    if _MRT is None:
        with _MRT_LOCK:
            if _MRT is None:
                ckpt_dir = CheckpointManager.resolve_checkpoint_dir()  # uses MRT_CKPT_REPO/STEP if present
                _MRT = system.MagentaRT(
                    tag=os.getenv("MRT_SIZE", "large"),
                    guidance_weight=1.0,
                    device="gpu",
                    checkpoint_dir=ckpt_dir,
                    lazy=False
                )
                # If no assets loaded yet, and a repo is configured, load them now.
                if asset_manager.mean_embed is None and asset_manager.centroids is None:
                    repo = os.getenv("MRT_ASSETS_REPO") or os.getenv("MRT_CKPT_REPO")
                    if repo:
                        asset_manager.load_finetune_assets_from_hf(repo, None)
                        _sync_assets_globals_from_manager()
    return _MRT

_WARMED = False
_WARMUP_LOCK = threading.Lock()

def _generate_pink_noise(n_samples: int, channels: int = 2) -> np.ndarray:
    """
    Generate pink (1/f) noise using the Voss-McCartney algorithm approximation.
    Pink noise is more representative of real audio spectrally than silence,
    which helps avoid tainting JAX compilation cache with silence patterns.
    """
    # Simple approximation: filter white noise
    white = np.random.randn(n_samples, channels).astype(np.float32)
    # Apply simple 1/f rolloff via cumulative sum + normalization
    pink = np.cumsum(white, axis=0)
    # High-pass to remove DC drift
    pink = pink - np.mean(pink, axis=0, keepdims=True)
    # Normalize to reasonable level (-20 dBFS ish)
    peak = np.max(np.abs(pink))
    if peak > 0:
        pink = pink * (0.1 / peak)
    return pink.astype(np.float32)


def _mrt_warmup():
    """
    Build a minimal, bar-aligned PINK NOISE context and run one 2s generate_chunk
    to trigger XLA JIT & autotune so first real request is fast.
    
    KEY FIX: Uses pink noise instead of silence for warmup!
    Pure silence can taint JAX compilation cache with silence patterns,
    leading to intermittent silence gaps in generated audio.
    """
    global _WARMED
    with _WARMUP_LOCK:
        if _WARMED:
            return
        try:
            mrt = get_mrt()

            # --- derive timing from model config ---
            codec_fps = float(mrt.codec.frame_rate)
            ctx_seconds = float(mrt.config.context_length_frames) / codec_fps
            sr = int(mrt.sample_rate)

            # We'll align to 120 BPM, 4/4, and generate one ~2s chunk
            bpm = 120.0
            beats_per_bar = 4

            # --- build a PINK NOISE stereo context of ctx_seconds ---
            samples = int(max(1, round(ctx_seconds * sr)))
            warmup_audio = _generate_pink_noise(samples, channels=2)
            logging.info("Warmup using pink noise instead of silence (%d samples)", samples)

            with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
                sf.write(tmp.name, warmup_audio, sr, subtype="PCM_16")
                tmp_path = tmp.name

            try:
                # Load as Waveform and take a tail of exactly ctx_seconds
                loop = au.Waveform.from_file(tmp_path).resample(sr).as_stereo()
                ctx_tail = take_bar_aligned_tail(loop, bpm, beats_per_bar, ctx_seconds)

                # Tokens for context window
                tokens_full = mrt.codec.encode(ctx_tail).astype(np.int32)
                tokens = tokens_full[:, :mrt.config.decoder_codec_rvq_depth]
                context_tokens = make_bar_aligned_context(
                    tokens,
                    bpm=bpm,
                    fps=float(mrt.codec.frame_rate),
                    ctx_frames=mrt.config.context_length_frames,
                    beats_per_bar=beats_per_bar,
                )

                # Init state and a basic style vector (text token is fine)
                state = mrt.init_state()
                state.context_tokens = context_tokens
                style_vec = mrt.embed_style("warmup")

                # --- one throwaway chunk (~2s) ---
                _wav, _state = mrt.generate_chunk(state=state, style=style_vec)

                logging.info("MagentaRT warmup complete (pink noise context).")
            finally:
                try:
                    os.unlink(tmp_path)
                except Exception:
                    pass

            _WARMED = True
        except Exception as e:
            # Never crash on warmup errors; log and continue serving
            logging.exception("MagentaRT warmup failed (continuing without warmup): %s", e)


# ----------------------------
# startup and model selection
# ----------------------------

@app.on_event("startup")
def _boot():
    # 1) Load finetune assets up front (only if envs are present)
    repo = os.getenv("MRT_ASSETS_REPO") or os.getenv("MRT_CKPT_REPO")
    if repo:
        ok, msg = asset_manager.load_finetune_assets_from_hf(repo, None)
        _sync_assets_globals_from_manager()  # keep /model/config in sync
        logging.info("Startup asset load from %s: %s", repo, "ok" if ok else msg)
    else:
        logging.info("Startup asset load: no repo env set; skipping.")

    # 2) Start warmup in the background (unchanged behavior)
    if os.getenv("MRT_WARMUP", "1") != "0":
        threading.Thread(target=_mrt_warmup, name="mrt-warmup", daemon=True).start()



@app.get("/model/status")
def model_status():
    mrt = get_mrt()
    return {
        "tag": getattr(mrt, "_tag", "unknown"),
        "using_checkpoint_dir": True,
        "codec_frame_rate": float(mrt.codec.frame_rate),
        "decoder_rvq_depth": int(mrt.config.decoder_codec_rvq_depth),
        "context_seconds": float(mrt.config.context_length),
        "chunk_seconds": float(mrt.config.chunk_length),
        "crossfade_seconds": float(mrt.config.crossfade_length),
        "selected_step": os.getenv("MRT_CKPT_STEP"),
        "repo": os.getenv("MRT_CKPT_REPO"),
    }

@app.post("/model/swap")
def model_swap(step: int = Form(...)):
    # stop any active jam if you want to be strict (not shown)
    os.environ["MRT_CKPT_STEP"] = str(step)
    global _MRT
    with _MRT_LOCK:
        _MRT = None  # force re-create on next get_mrt()
    # optionally pre-warm here by calling get_mrt()
    return {"reloaded": True, "step": step}

@app.post("/model/assets/load")
def model_assets_load(repo_id: str = Form(None)):
    global _MEAN_EMBED, _CENTROIDS, _ASSETS_REPO_ID
    ok, msg = asset_manager.load_finetune_assets_from_hf(repo_id, get_mrt())
    # Sync globals after loading
    _MEAN_EMBED = asset_manager.mean_embed
    _CENTROIDS = asset_manager.centroids
    _ASSETS_REPO_ID = asset_manager.assets_repo_id
    return {"ok": ok, "message": msg, "repo_id": _ASSETS_REPO_ID,
            "mean": _MEAN_EMBED is not None,
            "centroids": None if _CENTROIDS is None else int(_CENTROIDS.shape[0])}

@app.get("/model/assets/status")
def model_assets_status():
    d = None
    try:
        d = int(get_mrt().style_model.config.embedding_dim)
    except Exception:
        pass
    return {
        "repo_id": _ASSETS_REPO_ID,
        "mean_loaded": _MEAN_EMBED is not None,
        "centroids_loaded": False if _CENTROIDS is None else True,
        "centroid_count": None if _CENTROIDS is None else int(_CENTROIDS.shape[0]),
        "embedding_dim": d,
    }

@app.get("/model/config")
def model_config():
    """
    Lightweight config snapshot:
      - never calls get_mrt() (no model build / no downloads)
      - never calls snapshot_download()
      - reports whether a model instance is currently loaded in memory
      - best-effort local checkpoint presence (no network)
    """
    # Read-only snapshot of in-memory model presence
    with _MRT_LOCK:
        loaded = (_MRT is not None)

    size   = os.getenv("MRT_SIZE", "large")
    repo   = os.getenv("MRT_CKPT_REPO")
    rev    = os.getenv("MRT_CKPT_REV", "main")
    step   = os.getenv("MRT_CKPT_STEP")
    assets = os.getenv("MRT_ASSETS_REPO")

    # Use CheckpointManager for local cache probe (no network)
    local_ckpt = None
    if step:
        try:
            from pathlib import Path
            import re
            step_escaped = re.escape(str(step))
            candidates = []
            for root in ("/home/appuser/.cache/mrt_ckpt/extracted",
                         "/home/appuser/.cache/mrt_ckpt/repo"):
                p = Path(root)
                if not p.exists():
                    continue
                # Look for exact "checkpoint_<step>" directories anywhere under these roots
                for d in p.rglob(f"checkpoint_{step}"):
                    if d.is_dir():
                        candidates.append(str(d))
            local_ckpt = candidates[0] if candidates else None
        except Exception:
            local_ckpt = None

    return {
        "size": size,
        "repo": repo,
        "revision": rev,
        "selected_step": step,
        "assets_repo": assets,

        # in-memory + local cache hints (no network, no model build)
        "loaded": loaded,
        "active_jam": _any_jam_running(),
        "local_checkpoint_dir": local_ckpt,   # None if not found locally

        # steering assets currently resident in memory
        "mean_loaded": (_MEAN_EMBED is not None),
        "centroids_loaded": (_CENTROIDS is not None),
        "centroid_count": (None if _CENTROIDS is None else int(_CENTROIDS.shape[0])),
        "warmup_done": bool(_WARMED),
    }

@app.get("/model/checkpoints")
def model_checkpoints(repo_id: str, revision: str = "main"):
    steps = CheckpointManager.list_ckpt_steps(repo_id, revision)
    return {"repo": repo_id, "revision": revision, "steps": steps, "latest": (steps[-1] if steps else None)}

@app.post("/model/select")
def model_select(req: ModelSelect):
    """
    Swap model/checkpoint/assets. If req.prewarm is True, run the full bar-aligned warmup
    (_mrt_warmup) synchronously so we only report warmed once the new model is actually ready.
    """
    global _MRT, _MEAN_EMBED, _CENTROIDS, _ASSETS_REPO_ID, _WARMED

    # 1) Validate the request (no side-effects)
    success, validation_result = model_selector.validate_selection(req)
    if not success:
        if "error" in validation_result:
            raise HTTPException(status_code=400, detail=validation_result["error"])
        return {"ok": False, **validation_result}
    


    # Augment response surface
    validation_result["active_jam"] = _any_jam_running()

    # Dry-run path
    if req.dry_run:
        return {"ok": True, "dry_run": True, **validation_result}
    
    if isinstance(req.step, str) and req.step.lower() == "none":  # user asked for stock base
        # Clear any resident finetune assets so /model/config reflects "no assets"
        asset_manager.mean_embed = None
        asset_manager.centroids = None
        asset_manager.assets_repo_id = None
        # keep module-level mirrors in sync
        _MEAN_EMBED = None
        _CENTROIDS = None
        _ASSETS_REPO_ID = None

    # 2) Handle jam policy
    if _any_jam_running():
        if req.stop_active:
            _stop_all_jams()
        else:
            raise HTTPException(status_code=409, detail="A jam is running; retry with stop_active=true")

    # 3) Compute environment changes (no mutation yet)
    env_changes = model_selector.prepare_env_changes(req, validation_result)

    # Keep current env for rollback
    old_env = {
        "MRT_SIZE": os.getenv("MRT_SIZE"),
        "MRT_CKPT_REPO": os.getenv("MRT_CKPT_REPO"),
        "MRT_CKPT_REV": os.getenv("MRT_CKPT_REV"),
        "MRT_CKPT_STEP": os.getenv("MRT_CKPT_STEP"),
        "MRT_ASSETS_REPO": os.getenv("MRT_ASSETS_REPO"),
    }

    try:
        # 4) Apply env atomically
        for key, value in env_changes.items():
            if value is None:
                os.environ.pop(key, None)
            else:
                os.environ[key] = str(value)

        # 5) Force rebuild of the model and reset warmup state
        with _MRT_LOCK:
            _MRT = None
        with _WARMUP_LOCK:
            _WARMED = False  # ← critical: don't leak previous model's warmed state

        # 6) Load finetune assets if requested (mean/centroids)
        if req.sync_assets and validation_result.get("assets_repo"):
            ok, msg = asset_manager.load_finetune_assets_from_hf(
                validation_result["assets_repo"],
                None  # don't implicitly instantiate model here; we'll do it below
            )
            if ok:
                _MEAN_EMBED = asset_manager.mean_embed
                _CENTROIDS = asset_manager.centroids
                _ASSETS_REPO_ID = asset_manager.assets_repo_id
            else:
                logging.warning("Asset sync skipped/failed: %s", msg)

        # 7) Prewarm behavior:
        #    - If prewarm=True, run the *real* bar-aligned warmup synchronously.
        #    - This will instantiate the new MRT and set _WARMED=True on success.
        if req.prewarm:
            _mrt_warmup()  # builds MRT internally via get_mrt(), runs generate_chunk, sets _WARMED

        # Optional: if you want to always ensure MRT exists (even without prewarm), uncomment:
        # else:
        #     _ = get_mrt()

        return {
            "ok": True,
            **validation_result,
            "warmup_done": bool(_WARMED),
        }

    except Exception as e:
        # 8) Roll back env on failure
        for k, v in old_env.items():
            if v is None:
                os.environ.pop(k, None)
            else:
                os.environ[k] = v
        # Also reset model pointer & warmed flag to a safe state
        with _MRT_LOCK:
            _MRT = None
        with _WARMUP_LOCK:
            _WARMED = False
        logging.exception("Model select failed: %s", e)
        raise HTTPException(status_code=500, detail=f"Model select failed: {e}")
    


# ----------------------------
# one-shot generation
# ----------------------------

@app.post("/generate/reset")
def generate_reset():
    """
    Manually trigger a light reset of the generation system.
    Useful if user notices quality degradation without full restart.
    """
    global _GENERATE_COUNTER
    
    with _GENERATE_COUNTER_LOCK:
        _GENERATE_COUNTER = 0  # Reset counter
    
    mrt = get_mrt()
    success = _light_reset_mrt(mrt)
    
    return {
        "reset": success,
        "message": "Light reset complete" if success else "Reset partially completed",
        "counter_reset": True
    }

@app.get("/progress")
def progress(request_id: str):
    return _progress_get(request_id)

@app.post("/generate")
def generate(
    loop_audio: UploadFile = File(...),
    bpm: float = Form(...),
    bars: int = Form(8),
    beats_per_bar: int = Form(4),
    styles: str = Form("acid house"),
    style_weights: str = Form(""),
    loop_weight: float = Form(1.0),
    loudness_mode: str = Form("auto"),
    loudness_headroom_db: float = Form(1.0),
    guidance_weight: float = Form(5.0),
    temperature: float = Form(1.1),
    topk: int = Form(40),
    target_sample_rate: int | None = Form(None),
    intro_bars_to_drop: int = Form(0),
    request_id: str = Form(None),
    force_reset: bool = Form(False),  # NEW: Manual reset trigger
):
    global _GENERATE_COUNTER

    req_id = request_id or str(uuid.uuid4())
    tmp_path = None

    try:
        # Check if we need a periodic reset
        with _GENERATE_COUNTER_LOCK:
            _GENERATE_COUNTER += 1
            gen_count = _GENERATE_COUNTER
            
            # Every 1 generations, do a light reset
            # (Or if user explicitly requests it)
            if gen_count % 1 == 0 or force_reset:
                logging.info(f"[Generate {req_id}] Triggering light reset (generation #{gen_count})")
                mrt = get_mrt()
                _light_reset_mrt(mrt)

        # 0) Read file -> tmp wav
        data = loop_audio.file.read()
        if not data:
            # finalize progress as error and return
            with _PROGRESS_LOCK:
                _PROGRESS[req_id] = {
                    "percent": 100,
                    "stage": "error",
                    "error": "Empty file",
                    "ts": time.time(),
                }
            return {"error": "Empty file", "request_id": req_id}

        with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
            tmp.write(data)
            tmp_path = tmp.name

        # 1) Parse styles + weights
        extra_styles = [s.strip() for s in (styles.split(",") if styles else []) if s.strip()]
        weights = [float(x) for x in style_weights.split(",")] if style_weights else None

        # 2) Get model and apply per-request overrides
        mrt = get_mrt()
        with mrt_overrides(
            mrt,
            guidance_weight=guidance_weight,
            temperature=temperature,
            topk=topk,
        ):
            # progress callback (called from the generator loop)
            def on_chunk(i, total):
                _progress_update(req_id, i, total, stage="generating")

            # 2a) (optional) emit initial 0% once steps are known:
            #     We'll do this inside the generator right after steps is computed.
            wav, loud_stats = generate_loop_continuation_with_mrt(
                mrt,
                input_wav_path=tmp_path,
                bpm=bpm,
                extra_styles=extra_styles,
                style_weights=weights,
                bars=bars,
                beats_per_bar=beats_per_bar,
                loop_weight=loop_weight,
                loudness_mode=loudness_mode,
                loudness_headroom_db=loudness_headroom_db,
                intro_bars_to_drop=intro_bars_to_drop,
                progress_cb=on_chunk,
            )

        # 3) Post-process stages (optional: expose sub-stages for nicer UI)
        #    Mark "postprocess" before we resample/snap/encode.
        st = _PROGRESS_GET(req_id) if False else None  # (placeholder so lints don't complain)
        _progress_update(
            req_id,
            _progress_get(req_id).get("total", 1),
            _progress_get(req_id).get("total", 1),
            "postprocess",
        )

        # 3a) Determine SR
        inp_info = sf.info(tmp_path)
        input_sr = int(inp_info.samplerate)
        target_sr_val = int(target_sample_rate or input_sr)

        # 3b) Convert SR + snap to exact bars
        cur_sr = int(mrt.sample_rate)
        x = wav.samples if wav.samples.ndim == 2 else wav.samples[:, None]
        seconds_per_bar = (60.0 / float(bpm)) * int(beats_per_bar)
        expected_secs = float(bars) * seconds_per_bar

        # (optional) sub-stage
        _progress_update(req_id, _progress_get(req_id).get("total", 1), _progress_get(req_id).get("total", 1), "resample_and_snap")
        x = resample_and_snap(x, cur_sr=cur_sr, target_sr=target_sr_val, seconds=expected_secs)

        # 3c) Encode WAV -> base64
        _progress_update(req_id, _progress_get(req_id).get("total", 1), _progress_get(req_id).get("total", 1), "encode")
        audio_b64, total_samples, channels = wav_bytes_base64(x, target_sr_val)
        loop_duration_seconds = total_samples / float(target_sr_val)

        # 4) Metadata
        metadata = {
            "bpm": int(round(bpm)),
            "bars": int(bars),
            "beats_per_bar": int(beats_per_bar),
            "styles": extra_styles,
            "style_weights": weights,
            "loop_weight": loop_weight,
            "loudness": loud_stats,
            "sample_rate": int(target_sr_val),
            "channels": int(channels),
            "crossfade_seconds": mrt.config.crossfade_length,
            "total_samples": int(total_samples),
            "seconds_per_bar": seconds_per_bar,
            "loop_duration_seconds": loop_duration_seconds,
            "guidance_weight": guidance_weight,
            "temperature": temperature,
            "topk": topk,
        }

        _progress_done(req_id)
        return {"audio_base64": audio_b64, "metadata": metadata, "request_id": req_id}

    except Exception as e:
        # Flip to error state so the UI stops polling and can show a message
        with _PROGRESS_LOCK:
            _PROGRESS[req_id] = {
                "percent": 100,
                "stage": "error",
                "error": str(e),
                "ts": time.time(),
            }
        # Re-raise so FastAPI returns a 500 (or your exception handler formats it)
        raise

    finally:
        # Clean up temp file
        if tmp_path:
            try:
                os.unlink(tmp_path)
            except Exception:
                pass


# new endpoint to return a bar-aligned chunk without the need for combined audio

@app.post("/generate_style")
def generate_style(
    bpm: float = Form(...),
    bars: int = Form(8),
    beats_per_bar: int = Form(4),
    styles: str = Form("warmup"),
    style_weights: str = Form(""),
    guidance_weight: float = Form(1.1),
    temperature: float = Form(1.1),
    topk: int = Form(40),
    target_sample_rate: int | None = Form(None),
    intro_bars_to_drop: int = Form(0),
):
    """
    Style-only, bar-aligned generation (no input audio).
    Seeds with 10s of silent context; outputs exactly `bars` at the requested BPM.
    """
    mrt = get_mrt()

    # Override sampling knobs just for this request
    with mrt_overrides(mrt,
                       guidance_weight=guidance_weight,
                       temperature=temperature,
                       topk=topk):
        wav, _ = generate_style_only_with_mrt(
            mrt,
            bpm=bpm,
            bars=bars,
            beats_per_bar=beats_per_bar,
            styles=styles,
            style_weights=style_weights,
            intro_bars_to_drop=intro_bars_to_drop,
        )

    # Determine target SR (defaults to model SR = 48k)
    cur_sr = int(mrt.sample_rate)
    target_sr = int(target_sample_rate or cur_sr)
    x = wav.samples if wav.samples.ndim == 2 else wav.samples[:, None]

    seconds_per_bar = (60.0 / float(bpm)) * int(beats_per_bar)
    expected_secs   = float(bars) * seconds_per_bar

    # Snap exactly to musical length at the requested sample rate
    x = resample_and_snap(x, cur_sr=cur_sr, target_sr=target_sr, seconds=expected_secs)

    audio_b64, total_samples, channels = wav_bytes_base64(x, target_sr)

    metadata = {
        "bpm": int(round(bpm)),
        "bars": int(bars),
        "beats_per_bar": int(beats_per_bar),
        "styles": [s.strip() for s in (styles.split(",") if styles else []) if s.strip()],
        "style_weights": [float(y) for y in style_weights.split(",")] if style_weights else None,
        "sample_rate": int(target_sr),
        "channels": int(channels),
        "crossfade_seconds": mrt.config.crossfade_length,
        "seconds_per_bar": seconds_per_bar,
        "loop_duration_seconds": total_samples / float(target_sr),
        "guidance_weight": guidance_weight,
        "temperature": temperature,
        "topk": topk,
    }
    return {"audio_base64": audio_b64, "metadata": metadata}


# ----------------------------
# the 'keep jamming' button
# ----------------------------

@app.post("/jam/start")
def jam_start(
    loop_audio: UploadFile = File(...),
    bpm: float = Form(...),
    bars_per_chunk: int = Form(4),
    beats_per_bar: int = Form(4),
    styles: str = Form(""),
    style_weights: str = Form(""),
    loop_weight: float = Form(1.0),
    # NEW steering params:
    mean: float = Form(0.0),
    centroid_weights: str = Form(""),
    loudness_mode: str = Form("auto"),
    loudness_headroom_db: float = Form(1.0),
    guidance_weight: float = Form(1.1),
    temperature: float = Form(1.1),
    topk: int = Form(40),
    target_sample_rate: int | None = Form(None),
):
    asset_manager.ensure_assets_loaded(get_mrt())

    # enforce single active jam per GPU
    with jam_lock:
        for sid, w in list(jam_registry.items()):
            if w.is_alive():
                raise HTTPException(status_code=429, detail="A jam is already running. Try again later.")

    # read input + prep context/style (reuse your existing code)
    data = loop_audio.file.read()
    if not data: raise HTTPException(status_code=400, detail="Empty file")
    with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
        tmp.write(data); tmp_path = tmp.name

    mrt = get_mrt()
    loop = au.Waveform.from_file(tmp_path).resample(mrt.sample_rate).as_stereo()

    # build tail context + style vec (tail-biased)
    codec_fps = float(mrt.codec.frame_rate)
    ctx_seconds = float(mrt.config.context_length_frames) / codec_fps
    loop_tail = take_bar_aligned_tail(loop, bpm, beats_per_bar, ctx_seconds)

    # Parse client style fields (preserves your semantics)
    text_list = [s.strip() for s in (styles.split(",") if styles else []) if s.strip()]
    try:
        tw = [float(x) for x in style_weights.split(",")] if style_weights else []
    except ValueError:
        tw = []
    try:
        cw = [float(x) for x in centroid_weights.split(",")] if centroid_weights else []
    except ValueError:
        cw = []

    # Compute loop-tail embed once (same as before)
    loop_tail_embed = mrt.embed_style(loop_tail)

    # Build final style vector:
    # - identical to your previous mix when mean==0 and cw is empty
    # - otherwise includes mean and centroid components (weights auto-normalized)
    style_vec = build_style_vector(
        mrt,
        text_styles=text_list,
        text_weights=tw,
        loop_embed=loop_tail_embed,
        loop_weight=float(loop_weight),
        mean_weight=float(mean),
        centroid_weights=cw,
    ).astype(np.float32, copy=False)

    # target SR (default input SR)
    inp_info = sf.info(tmp_path)
    input_sr = int(inp_info.samplerate)
    target_sr = int(target_sample_rate or input_sr)

    params = JamParams(
        bpm=bpm, 
        beats_per_bar=beats_per_bar, 
        bars_per_chunk=bars_per_chunk,
        target_sr=target_sr, 
        loudness_mode=loudness_mode, 
        headroom_db=loudness_headroom_db,
        style_vec=style_vec, 
        ref_loop=loop_tail,                    # For loudness matching
        combined_loop=loop,                    # NEW: Full loop for context setup
        guidance_weight=guidance_weight, 
        temperature=temperature, 
        topk=topk
    )

    worker = JamWorker(mrt, params)
    sid = str(uuid.uuid4())
    with jam_lock:
        jam_registry[sid] = worker
    worker.start()

    return {"session_id": sid}

@app.get("/jam/next")
def jam_next(session_id: str):
    """
    Get the next sequential chunk in the jam session.
    This ensures chunks are delivered in order without gaps.
    
    The new JamWorker returns raw WAV bytes in chunk.audio_bytes.
    We convert to base64 here for backward compatibility with existing clients.
    """
    with jam_lock:
        worker = jam_registry.get(session_id)
    if worker is None or not worker.is_alive():
        raise HTTPException(status_code=404, detail="Session not found")

    # Get the next sequential chunk (this blocks until ready)
    chunk = worker.get_next_chunk()
    
    if chunk is None:
        raise HTTPException(status_code=408, detail="Chunk not ready within timeout")

    # Convert raw WAV bytes to base64 for backward compatibility
    audio_base64 = base64.b64encode(chunk.audio_bytes).decode("utf-8")

    return {
        "chunk": {
            "index": chunk.index,
            "audio_base64": audio_base64,
            "metadata": chunk.metadata
        }
    }

@app.post("/jam/consume")
def jam_consume(session_id: str = Form(...), chunk_index: int = Form(...)):
    """
    Mark a chunk as consumed by the frontend.
    This helps the worker manage its buffer and generation flow.
    """
    with jam_lock:
        worker = jam_registry.get(session_id)
    if worker is None or not worker.is_alive():
        raise HTTPException(status_code=404, detail="Session not found")

    worker.mark_chunk_consumed(chunk_index)
    
    return {"consumed": chunk_index}



@app.post("/jam/stop")
def jam_stop(session_id: str = Body(..., embed=True)):
    with jam_lock:
        worker = jam_registry.get(session_id)
    if worker is None:
        raise HTTPException(status_code=404, detail="Session not found")

    worker.stop()
    worker.join(timeout=5.0)
    if worker.is_alive():
        # It’s daemon=True, so it won’t block process exit, but report it
        print(f"⚠️ JamWorker {session_id} did not stop within timeout")

    with jam_lock:
        jam_registry.pop(session_id, None)
    return {"stopped": True}

@app.post("/jam/stop_all")
def jam_stop_all():
    """Force stop all active jam sessions (nuclear option for cleanup)"""
    stopped_sessions = []
    
    with jam_lock:
        for session_id, worker in list(jam_registry.items()):
            if worker.is_alive():
                worker.stop()
                worker.join(timeout=2.0)
                stopped_sessions.append(session_id)
            jam_registry.pop(session_id, None)
    
    return {"stopped_sessions": stopped_sessions, "count": len(stopped_sessions)}

@app.post("/jam/update")
def jam_update(
    session_id: str = Form(...),
    # knobs
    guidance_weight: Optional[float] = Form(None),
    temperature: Optional[float]     = Form(None),
    topk: Optional[int]              = Form(None),
    # styles
    styles: str                      = Form(""),
    style_weights: str               = Form(""),
    loop_weight: Optional[float]     = Form(None),
    use_current_mix_as_style: bool   = Form(False),
    # NEW steering
    mean: Optional[float]            = Form(None),
    centroid_weights: str            = Form(""),
):
    asset_manager.ensure_assets_loaded(get_mrt())

    with jam_lock:
        worker = jam_registry.get(session_id)
    if worker is None or not worker.is_alive():
        raise HTTPException(status_code=404, detail="Session not found")

    # 1) fast knob updates
    if any(v is not None for v in (guidance_weight, temperature, topk)):
        worker.update_knobs(
            guidance_weight=guidance_weight,
            temperature=temperature,
            topk=topk
        )

    # 2) rebuild style only if asked
    wants_style_update = (
        use_current_mix_as_style
        or (styles.strip() != "")
        or (mean is not None)
        or (centroid_weights.strip() != "")
    )
    if not wants_style_update:
        return {"ok": True}

    # --- parse inputs (robust) ---
    text_list = [s.strip() for s in (styles.split(",") if styles else []) if s.strip()]
    try:
        tw = [float(x) for x in style_weights.split(",")] if style_weights else []
    except ValueError:
        tw = []
    try:
        cw = [float(x) for x in centroid_weights.split(",")] if centroid_weights else []
    except ValueError:
        cw = []

    # Clamp centroid weights to available centroids (if loaded)
    max_c = 0 if _CENTROIDS is None else int(_CENTROIDS.shape[0])
    if max_c and len(cw) > max_c:
        cw = cw[:max_c]

    # Snapshot minimal state under lock
    with worker._lock:
        combined_loop = worker.params.combined_loop if use_current_mix_as_style else None
        lw = None
        if use_current_mix_as_style:
            lw = 1.0 if (loop_weight is None) else float(loop_weight)
        mrt = worker.mrt

    # Heavy work OUTSIDE the lock
    loop_embed = None
    if combined_loop is not None:
        loop_embed = mrt.embed_style(combined_loop)

    style_vec = build_style_vector(
        mrt,
        text_styles=text_list,
        text_weights=tw,
        loop_embed=loop_embed,             # None => ignored by builder
        loop_weight=lw,                    # None => ignored by builder
        mean_weight=(None if mean is None else float(mean)),
        centroid_weights=cw,               # [] => ignored by builder
    ).astype(np.float32, copy=False)

    # Swap atomically
    with worker._lock:
        worker.params.style_vec = style_vec

    return {"ok": True}


@app.post("/jam/reseed")
def jam_reseed(session_id: str = Form(...), loop_audio: UploadFile = File(None)):
    with jam_lock:
        worker = jam_registry.get(session_id)
    if worker is None or not worker.is_alive():
        raise HTTPException(status_code=404, detail="Session not found")

    # Option 1: use uploaded new combined bounce from the app (preferred)
    if loop_audio is not None:
        data = loop_audio.file.read()
        if not data:
            raise HTTPException(status_code=400, detail="Empty file")

        with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
            tmp.write(data); path = tmp.name
        wav = au.Waveform.from_file(path).resample(worker.mrt.sample_rate).as_stereo()
    else:
        # Option 2: reseed from the original combined_loop
        # (Usually better to reseed from the Swift-side combined mix you upload.)
        combined_loop = worker.params.combined_loop
        if combined_loop is None:
            raise HTTPException(status_code=400, detail="No combined_loop available to reseed from. Please upload loop_audio.")
        wav = combined_loop.as_stereo().resample(worker.mrt.sample_rate)

    worker.reseed_from_waveform(wav)
    return {"ok": True}


@app.post("/jam/reseed_splice")
def jam_reseed_splice(
    session_id: str = Form(...),
    anchor_bars: float = Form(2.0),              # how much of the original to re-inject
    combined_audio: UploadFile = File(None),     # optional: Swift supplies the current combined mix
):
    """
    Seamless reseed by splicing tokens at the next chunk boundary.
    
    The new jam_worker implementation:
    - Keeps the *first* `anchor_bars` worth of context from the original loop
    - Keeps the *remainder* from the CURRENT state.context_tokens (no re-encoding!)
    - This ensures the tail of the spliced context is token-identical to what 
      MagentaRT was already working with, avoiding codec round-trip artifacts.
    
    If combined_audio is provided, it will be parsed but the splice itself 
    uses current tokens to avoid re-encoding artifacts.
    """
    worker = jam_registry.get(session_id)
    if worker is None or not worker.is_alive():
        raise HTTPException(status_code=404, detail="Session not found")

    # If combined_audio is provided, parse it (future use)
    # but the reseed_splice itself doesn't need it anymore
    recent_wav = None
    if combined_audio is not None:
        data = combined_audio.file.read()
        if data:
            with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
                tmp.write(data)
                path = tmp.name
            recent_wav = au.Waveform.from_file(path).resample(worker.mrt.sample_rate).as_stereo()

    # Perform the splice reseed
    # New signature: reseed_splice(recent_wav: Optional, anchor_bars: float)
    # recent_wav is now optional - if None, uses current state tokens
    worker.reseed_splice(recent_wav, anchor_bars=float(anchor_bars))
    return {"ok": True, "anchor_bars": float(anchor_bars)}


@app.get("/jam/status")
def jam_status(session_id: str):
    with jam_lock:
        worker = jam_registry.get(session_id)

    if worker is None:
        raise HTTPException(status_code=404, detail="Session not found")

    running = worker.is_alive()

    # Snapshot safely - use new attribute names from updated jam_worker.py
    with worker._lock:
        last_generated = int(worker.idx)
        # _next_to_deliver is the index we're ABOUT to deliver, so last delivered is one less
        last_delivered = int(worker._next_to_deliver) - 1
        queued = len(worker._outbox)  # Changed from outbox to _outbox
        buffer_ahead = last_generated - max(0, last_delivered)
        p = worker.params
        spb = p.beats_per_bar * (60.0 / p.bpm)
        chunk_secs = p.bars_per_chunk * spb

    return {
        "running": running,
        "last_generated_index": last_generated,       # Last chunk that finished generating
        "last_delivered_index": last_delivered,       # Last chunk sent to frontend
        "buffer_ahead": buffer_ahead,                  # How many chunks ahead we are
        "queued_chunks": queued,                       # Total chunks in outbox
        "bpm": p.bpm,
        "beats_per_bar": p.beats_per_bar,
        "bars_per_chunk": p.bars_per_chunk,
        "seconds_per_bar": spb,
        "chunk_duration_seconds": chunk_secs,
        "target_sample_rate": p.target_sr,
        "last_chunk_started_at": worker.last_chunk_started_at,
        "last_chunk_completed_at": worker.last_chunk_completed_at,
    }


@app.get("/health")
def health():
    # 1) Template mode → not ready (encourage duplication on GPU)
    if SPACE_MODE != "serve":
        return JSONResponse(
            status_code=503,
            content={
                "ok": False,
                "status": "template_mode",
                "message": "This Space is a GPU template. Duplicate it and select an L40s/A100-class runtime to use the API.",
                "mode": SPACE_MODE,
            },
        )

    # 2) Runtime hardware probe
    probe = _gpu_probe()
    if not probe["ok"] or not probe["has_gpu"] or probe.get("backend") != "gpu":
        return JSONResponse(
            status_code=503,
            content={
                "ok": False,
                "status": "gpu_unavailable",
                "message": "GPU is not visible to JAX. Select a GPU runtime (e.g., L40s) to serve.",
                "probe": probe,
                "mode": SPACE_MODE,
            },
        )

    # 3) Ready; include operational hints
    warmed = bool(_WARMED)
    with jam_lock:
        active_jams = sum(1 for w in jam_registry.values() if w.is_alive())
    return {
        "ok": True,
        "status": "ready" if warmed else "initializing",
        "mode": SPACE_MODE,
        "warmed": warmed,
        "active_jams": active_jams,
        "probe": probe,
    }

@app.middleware("http")
async def log_requests(request: Request, call_next):
    rid = request.headers.get("X-Request-ID", "-")
    print(f"📥 {request.method} {request.url.path}?{request.url.query} [rid={rid}]")
    try:
        response = await call_next(request)
    except Exception as e:
        print(f"💥 exception for {request.url.path} [rid={rid}]: {e}")
        raise
    print(f"📤 {response.status_code} {request.url.path} [rid={rid}]")
    return response





# ----------------------------
# websockets route (rt-mode only)
# ----------------------------


@app.websocket("/ws/jam")
async def ws_jam(websocket: WebSocket):
    """
    Real-time streaming WebSocket endpoint for MagentaRT.
    
    This route operates in 'rt' mode only - for bar-aligned jam sessions,
    use the HTTP endpoints (/jam/start, /jam/chunk, etc.) instead.
    
    The server handles crossfading internally via MagentaRTState, so clients
    can simply play back chunks sequentially without additional crossfade logic.
    """
    await websocket.accept()
    binary_audio = False

    async def send_json(obj):
        return await send_json_safe(websocket, obj)

    try:
        while True:
            raw = await websocket.receive_text()
            msg = json.loads(raw)
            mtype = msg.get("type")

            # --- START (rt-mode only) ---
            if mtype == "start":
                binary_audio = bool(msg.get("binary_audio", False))
                params = msg.get("params", {}) or {}

                # Initialize MagentaRT state
                mrt = get_mrt()
                state = mrt.init_state()

                # Build silent context (10s) tokens
                codec_fps   = float(mrt.codec.frame_rate)
                ctx_seconds = float(mrt.config.context_length_frames) / codec_fps
                sr = int(mrt.sample_rate)
                samples = int(max(1, round(ctx_seconds * sr)))
                silent = au.Waveform(np.zeros((samples, 2), np.float32), sr)
                tokens = mrt.codec.encode(silent).astype(np.int32)[:, :mrt.config.decoder_codec_rvq_depth]
                state.context_tokens = tokens

                # Parse params (including steering)
                asset_manager.ensure_assets_loaded(get_mrt())
                styles_str        = params.get("styles", "warmup") or ""
                style_weights_str = params.get("style_weights", "") or ""
                mean_w            = float(params.get("mean", 0.0) or 0.0)
                cw_str            = str(params.get("centroid_weights", "") or "")

                text_list = [s.strip() for s in styles_str.split(",") if s.strip()]
                try:
                    text_w = [float(x) for x in style_weights_str.split(",")] if style_weights_str else []
                except ValueError:
                    text_w = []
                try:
                    cw = [float(x) for x in cw_str.split(",") if x.strip() != ""]
                except ValueError:
                    cw = []

                # Clamp centroid weights to available centroids
                if _CENTROIDS is not None and len(cw) > int(_CENTROIDS.shape[0]):
                    cw = cw[: int(_CENTROIDS.shape[0])]

                # Build initial style vector (no loop_embed in rt mode)
                style_vec = build_style_vector(
                    mrt,
                    text_styles=text_list,
                    text_weights=text_w,
                    loop_embed=None,
                    loop_weight=None,
                    mean_weight=mean_w,
                    centroid_weights=cw,
                )

                # Stash rt session fields on the websocket object
                websocket._mrt   = mrt
                websocket._state = state
                websocket._style_cur = style_vec
                websocket._style_tgt = style_vec
                websocket._style_ramp_s = float(params.get("style_ramp_seconds", 0.0))

                websocket._rt_mean              = mean_w
                websocket._rt_centroid_weights  = cw
                websocket._rt_running           = True
                websocket._rt_sr                = sr
                websocket._rt_topk              = int(params.get("topk", 40))
                websocket._rt_temp              = float(params.get("temperature", 1.1))
                websocket._rt_guid              = float(params.get("guidance_weight", 1.1))
                websocket._pace                 = params.get("pace", "asap")  # "realtime" | "asap"

                # Report whether steering assets were loaded
                assets_ok = (_MEAN_EMBED is not None) or (_CENTROIDS is not None)
                await send_json({"type": "started", "mode": "rt", "steering_assets": "loaded" if assets_ok else "none"})

                # Kick off the ~2s streaming loop
                async def _rt_loop():
                    try:
                        mrt = websocket._mrt
                        chunk_secs = (mrt.config.chunk_length_frames * mrt.config.frame_length_samples) / float(mrt.sample_rate)
                        target_next = time.perf_counter()
                        while websocket._rt_running:
                            mrt.guidance_weight = websocket._rt_guid
                            mrt.temperature     = websocket._rt_temp
                            mrt.topk            = websocket._rt_topk

                            # Ramp style toward target
                            ramp = float(getattr(websocket, "_style_ramp_s", 0.0) or 0.0)
                            if ramp <= 0.0:
                                websocket._style_cur = websocket._style_tgt
                            else:
                                step = min(1.0, chunk_secs / ramp)
                                websocket._style_cur = websocket._style_cur + step * (websocket._style_tgt - websocket._style_cur)

                            # Generate chunk (crossfading handled internally by MagentaRT)
                            wav, new_state = mrt.generate_chunk(state=websocket._state, style=websocket._style_cur)
                            websocket._state = new_state

                            x = wav.samples.astype(np.float32, copy=False)
                            buf = io.BytesIO()
                            sf.write(buf, x, mrt.sample_rate, subtype="FLOAT", format="WAV")

                            ok = True
                            if binary_audio:
                                try:
                                    await websocket.send_bytes(buf.getvalue())
                                    ok = await send_json({"type": "chunk_meta", "metadata": {"sample_rate": mrt.sample_rate}})
                                except Exception:
                                    ok = False
                            else:
                                b64 = base64.b64encode(buf.getvalue()).decode("utf-8")
                                ok = await send_json({"type": "chunk", "audio_base64": b64,
                                                    "metadata": {"sample_rate": mrt.sample_rate}})

                            if not ok:
                                break

                            if getattr(websocket, "_pace", "asap") == "realtime":
                                t1 = time.perf_counter()
                                target_next += chunk_secs
                                sleep_s = max(0.0, target_next - t1 - 0.02)
                                if sleep_s > 0:
                                    await asyncio.sleep(sleep_s)
                    except asyncio.CancelledError:
                        pass
                    except Exception:
                        pass

                websocket._rt_task = asyncio.create_task(_rt_loop())

            # --- UPDATE (rt-mode knobs & style) ---
            elif mtype == "update":
                if not hasattr(websocket, "_rt_running"):
                    await send_json({"type":"error","error":"Session not started"}); continue
                    
                websocket._rt_temp = float(msg.get("temperature", websocket._rt_temp))
                websocket._rt_topk = int(msg.get("topk", websocket._rt_topk))
                websocket._rt_guid = float(msg.get("guidance_weight", websocket._rt_guid))

                # Steering fields
                if "mean" in msg and msg["mean"] is not None:
                    try: websocket._rt_mean = float(msg["mean"])
                    except: websocket._rt_mean = 0.0

                if "centroid_weights" in msg:
                    cw = [w.strip() for w in str(msg["centroid_weights"]).split(",") if w.strip() != ""]
                    try:
                        websocket._rt_centroid_weights = [float(x) for x in cw]
                    except:
                        websocket._rt_centroid_weights = []

                # Styles / text weights (optional, comma-separated)
                styles_str = msg.get("styles", None)
                style_weights_str = msg.get("style_weights", "")

                text_list = [s for s in (styles_str.split(",") if styles_str else []) if s.strip()]
                text_w = [float(x) for x in style_weights_str.split(",")] if style_weights_str else []

                asset_manager.ensure_assets_loaded(get_mrt())
                websocket._style_tgt = build_style_vector(
                    websocket._mrt,
                    text_styles=text_list,
                    text_weights=text_w,
                    loop_embed=None,
                    loop_weight=None,
                    mean_weight=float(websocket._rt_mean),
                    centroid_weights=websocket._rt_centroid_weights,
                )
                # Optionally allow live changes to ramp:
                if "style_ramp_seconds" in msg:
                    try: websocket._style_ramp_s = float(msg["style_ramp_seconds"])
                    except: pass
                await send_json({"type":"status","updated":"rt-knobs+style"})

            # --- BUFFER STATUS (from frontend for adaptive pacing - acknowledged silently) ---
            elif mtype == "buffer_status":
                # Frontend reports its buffer level; could be used for adaptive pacing
                # For now we just acknowledge receipt without action
                pass

            # --- STOP ---
            elif mtype == "stop":
                websocket._rt_running = False
                task = getattr(websocket, "_rt_task", None)
                if task is not None:
                    task.cancel()
                    try: await task
                    except asyncio.CancelledError: pass
                await send_json({"type":"stopped"})
                break

            # --- PING/PONG ---
            elif mtype == "ping":
                await send_json({"type":"pong"})

            else:
                await send_json({"type":"error","error":f"Unknown type {mtype}"})

    except WebSocketDisconnect:
        pass
    except Exception as e:
        try:
            await send_json({"type":"error","error":str(e)})
        except Exception:
            pass
    finally:
        # Ensure streaming loop is stopped
        if hasattr(websocket, "_rt_running"):
            websocket._rt_running = False
        task = getattr(websocket, "_rt_task", None)
        if task is not None:
            task.cancel()
            try: await asyncio.wait_for(task, timeout=1.0)
            except: pass
        try:
            if websocket.client_state != WebSocketState.DISCONNECTED:
                await websocket.close()
        except Exception:
            pass

# --- FastRTC Gradio Integration (only in serve mode) ---
if SPACE_MODE == "serve":
    try:
        from fastrtc_magenta import create_magenta_stream, FASTRTC_AVAILABLE
        if FASTRTC_AVAILABLE:
            magenta_stream = create_magenta_stream(
                get_mrt_fn=get_mrt,
                build_style_fn=build_style_vector,
                asset_manager=asset_manager,
                concurrency_limit=1,
                time_limit=3600,
            )
            magenta_stream.mount(app, path="/rtc")
            app = gr.mount_gradio_app(app, magenta_stream.ui, path="/gradio")
            print("✓ FastRTC Gradio interface available at /gradio")
    except Exception as e:
        print(f"⚠ FastRTC integration skipped: {e}")


@app.get("/ping")
def ping():
    return {"ok": True}

@app.get("/", response_class=Response)
def read_root():
    """Root endpoint that explains what this API does"""
    try:
        html_file = Path(__file__).parent / "documentation.html"
        html_content = html_file.read_text(encoding='utf-8')
    except FileNotFoundError:
        # Fallback if file is missing
        html_content = """
        <!DOCTYPE html>
        <html><body>
        <h1>MagentaRT Research API</h1>
        <p>Documentation file not found. Please check documentation.html</p>
        </body></html>
        """
    return Response(content=html_content, media_type="text/html")

@app.get("/lil_demo_540p.mp4")
def demo_video():
    return FileResponse(Path(__file__).parent / "lil_demo_540p.mp4", media_type="video/mp4")

@app.get("/tester", response_class=HTMLResponse)
def tester():
    html_path = Path(__file__).parent / "magentaRT_rt_tester.html"
    return HTMLResponse(
        html_path.read_text(encoding="utf-8"),
        headers={"Cache-Control": "no-store"}  # avoid sticky caches while iterating
    )

@app.get("/magenta_prompts.js")
def magenta_prompts_js():
    js_path = Path(__file__).parent / "magenta_prompts.js"
    return FileResponse(js_path, media_type="text/javascript",
                        headers={"Cache-Control": "no-store"})