feat(multimodal): 实现本地模型管线与请求用量配置

backend/app/local_models/__init__.py

@@ -0,0 +1 @@
+"""Optional local inference; importing this package does not load model libraries."""

backend/app/local_models/catalog.py

@@ -0,0 +1,31 @@
+"""Reviewed model identities. Runtime never resolves a moving model revision."""
+from dataclasses import asdict, dataclass
+
+
+@dataclass(frozen=True)
+class ModelSpec:
+    key: str
+    name: str
+    capability: str
+    repository: str
+    revision: str
+    license: str
+    source: str = "huggingface"
+    dimensions: int | None = None
+
+    def public(self):
+        return asdict(self)
+
+
+CATALOG = {
+    spec.key: spec for spec in [
+        ModelSpec("bekko", "Bekko Embedding v1 A8M", "embedding", "hotchpotch/bekko-embedding-v1-a8m",
+                  "c721113d59a1d91b447450324f51c4b3332c924a", "MIT", dimensions=384),
+        ModelSpec("granite", "Granite Embedding 97M Multilingual r2", "embedding", "ibm-granite/granite-embedding-97m-multilingual-r2",
+                  "835ad14087e140460703cf0fae09f97d469d65c2", "Apache-2.0", dimensions=384),
+        ModelSpec("qwen3-asr", "Qwen3 ASR 0.6B", "transcription", "Qwen/Qwen3-ASR-0.6B",
+                  "5eb144179a02acc5e5ba31e748d22b0cf3e303b0", "Apache-2.0"),
+        ModelSpec("eres2netv2", "ERes2NetV2 中文声纹", "speaker_matching", "iic/speech_eres2netv2_sv_zh-cn_16k-common",
+                  "3317286545c587ae682dbc166831d9448780eebb", "Apache-2.0", source="modelscope", dimensions=192),
+    ]
+}

backend/app/local_models/manager.py

@@ -0,0 +1,178 @@
+"""Explicit resumable downloads; inference itself never fetches weights."""
+from __future__ import annotations
+
+import asyncio
+import hashlib
+import json
+import shutil
+from pathlib import Path
+from urllib.parse import quote
+
+import httpx
+
+from app.config import get_settings
+from app.errors import ApiError
+from app.local_models.catalog import CATALOG
+
+_downloads: dict[tuple[str, str], asyncio.Task] = {}
+
+
+def model_path(key: str) -> Path:
+    if key not in CATALOG:
+        raise ApiError(404, "MODEL_NOT_FOUND", "Unknown local model.")
+    return get_settings().data_dir / "models" / key / CATALOG[key].revision
+
+
+def state_path(key):
+    return model_path(key) / "install-state.json"
+
+
+def read_state(key):
+    try:
+        state = json.loads(state_path(key).read_text(encoding="utf-8"))
+    except (OSError, ValueError):
+        state = {"status": "not_installed", "downloaded_bytes": 0, "total_bytes": None}
+    if state["status"] == "downloading" and task_key(key) not in _downloads:
+        state.update(status="interrupted", error_code="DOWNLOAD_INTERRUPTED")
+    return state
+
+
+def write_state(key, state):
+    path = state_path(key)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    temporary = path.with_suffix(".tmp")
+    temporary.write_text(json.dumps(state), encoding="utf-8")
+    temporary.replace(path)
+
+
+def task_key(key):
+    return str(model_path(key)), key
+
+
+def describe():
+    return {"items": [{**spec.public(), **read_state(key)} for key, spec in CATALOG.items()]}
+
+
+async def download(key):
+    model_path(key)
+    if task_key(key) not in _downloads and read_state(key)["status"] != "installed":
+        write_state(key, {"status": "downloading", "downloaded_bytes": 0, "total_bytes": None})
+        task = asyncio.create_task(_download(key))
+        _downloads[task_key(key)] = task
+        task.add_done_callback(lambda done: _downloads.pop(task_key(key), None))
+    return read_state(key)
+
+
+async def cancel_download(key):
+    task = _downloads.get(task_key(key))
+    if task:
+        task.cancel()
+        await asyncio.gather(task, return_exceptions=True)
+    state = read_state(key)
+    if state["status"] == "downloading":
+        state["status"] = "interrupted"
+        write_state(key, state)
+    return state
+
+
+async def delete(key):
+    from app.local_models.runtime import runtime
+    if runtime.in_use(key):
+        raise ApiError(409, "MODEL_IN_USE", "Model is serving an active request.")
+    await cancel_download(key)
+    path = model_path(key).resolve()
+    root = (get_settings().data_dir / "models").resolve()
+    if not path.is_relative_to(root) or path == root:
+        raise ApiError(400, "INVALID_MODEL_PATH", "Model path escapes storage.")
+    if path.exists():
+        shutil.rmtree(path)
+    return read_state(key)
+
+
+async def _manifest(client, spec):
+    if spec.source == "huggingface":
+        response = await client.get(f"https://huggingface.co/api/models/{spec.repository}/revision/{spec.revision}?blobs=true")
+        response.raise_for_status()
+        files = []
+        for item in response.json()["siblings"]:
+            name = item["rfilename"]
+            if name.startswith(("onnx/", "openvino/", ".")) or not name.endswith((".json", ".txt", ".safetensors", ".md")):
+                continue
+            lfs = item.get("lfs") or {}
+            files.append({"path": name, "size": item["size"], "hash": lfs.get("sha256") or item["blobId"],
+                          "algorithm": "sha256" if lfs else "git-blob",
+                          "url": f"https://huggingface.co/{spec.repository}/resolve/{spec.revision}/{quote(name)}"})
+        return files
+    response = await client.get(f"https://modelscope.cn/api/v1/models/{spec.repository}/repo/files",
+                                params={"Revision": spec.revision, "Recursive": "true"})
+    response.raise_for_status()
+    return [{"path": f["Path"], "size": f["Size"], "hash": f["Sha256"], "algorithm": "sha256",
+             "url": f"https://modelscope.cn/api/v1/models/{spec.repository}/repo?Revision={spec.revision}&FilePath={quote(f['Path'])}"}
+            for f in response.json()["Data"]["Files"]
+            if f["Path"] in {"configuration.json", "pretrained_eres2netv2.ckpt", "README.md"}]
+
+
+def valid_file(path, entry):
+    if not path.is_file() or path.stat().st_size != entry["size"]:
+        return False
+    digest = hashlib.sha256() if entry["algorithm"] == "sha256" else hashlib.sha1()
+    if entry["algorithm"] == "git-blob":
+        digest.update(f"blob {entry['size']}\0".encode())
+    with path.open("rb") as stream:
+        for chunk in iter(lambda: stream.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest() == entry["hash"]
+
+
+async def _download(key):
+    spec, root = CATALOG[key], model_path(key).resolve()
+    state = {"status": "downloading", "downloaded_bytes": 0, "total_bytes": None}
+    try:
+        async with httpx.AsyncClient(timeout=60, follow_redirects=True) as client:
+            manifest = await _manifest(client, spec)
+            if not manifest or not any(f["path"].endswith((".safetensors", ".ckpt")) for f in manifest):
+                raise ValueError("Missing weights in model manifest")
+            state["total_bytes"] = sum(f["size"] for f in manifest)
+            root.mkdir(parents=True, exist_ok=True)
+            if shutil.disk_usage(root).free < state["total_bytes"] + 100 * 1024 * 1024:
+                raise ApiError(507, "MODEL_DISK_FULL", "Insufficient free disk space.")
+            complete = 0
+            for entry in manifest:
+                path = (root / entry["path"]).resolve()
+                if not path.is_relative_to(root):
+                    raise ValueError("Invalid model manifest path")
+                path.parent.mkdir(parents=True, exist_ok=True)
+                if await asyncio.to_thread(valid_file, path, entry):
+                    complete += entry["size"]
+                    continue
+                partial = path.with_suffix(path.suffix + ".partial")
+                offset = partial.stat().st_size if partial.exists() else 0
+                if offset >= entry["size"]:
+                    partial.unlink()
+                    offset = 0
+                async with client.stream("GET", entry["url"], headers={"Range": f"bytes={offset}-"} if offset else {}) as response:
+                    response.raise_for_status()
+                    if offset and response.status_code != 206:
+                        offset = 0
+                    if response.status_code == 206 and not response.headers.get("content-range", "").startswith(f"bytes {offset}-"):
+                        raise ValueError("Invalid download range")
+                    with partial.open("ab" if offset else "wb") as stream:
+                        async for chunk in response.aiter_bytes(1024 * 1024):
+                            offset += len(chunk)
+                            if offset > entry["size"]:
+                                raise ValueError("Download exceeds manifest size")
+                            stream.write(chunk)
+                            state["downloaded_bytes"] = complete + offset
+                            write_state(key, state)
+                if not await asyncio.to_thread(valid_file, partial, entry):
+                    partial.unlink(missing_ok=True)
+                    raise ApiError(422, "MODEL_CHECKSUM_FAILED", "Model file checksum did not match.")
+                partial.replace(path)
+                complete += entry["size"]
+            (root / "verified-manifest.json").write_text(json.dumps(manifest), encoding="utf-8")
+            state.update(status="installed", downloaded_bytes=complete)
+    except asyncio.CancelledError:
+        state.update(status="interrupted", error_code="DOWNLOAD_CANCELLED")
+    except Exception as exc:
+        state.update(status="failed", error_code=exc.code if isinstance(exc, ApiError) else "MODEL_DOWNLOAD_FAILED")
+    write_state(key, state)

backend/app/local_models/runtime.py

@@ -0,0 +1,198 @@
+"""Bounded, cancellable model subprocesses with CPU as the default device."""
+from __future__ import annotations
+
+import asyncio
+import json
+import os
+from contextlib import closing
+from contextvars import ContextVar
+from pathlib import Path
+from typing import Literal
+
+from pydantic import BaseModel, Field
+
+from app.config import BACKEND_DIR
+from app.database.db import connect
+from app.errors import ApiError
+from app.local_models.catalog import CATALOG
+from app.local_models.manager import model_path, read_state
+from app.providers.base import ProviderError
+
+
+class RuntimeConfig(BaseModel):
+    device: Literal["cpu", "cuda"] = "cpu"
+    cpu_threads: int = Field(default=2, ge=1, le=32)
+    memory_limit_mb: int = Field(default=8192, ge=1024, le=131072)
+    gpu_memory_limit_mb: int = Field(default=4096, ge=512, le=65536)
+    timeout_seconds: int = Field(default=1800, ge=30, le=14400)
+    embedding_model: Literal["bekko", "granite"] = "bekko"
+    version: int = Field(default=1, ge=1)
+
+
+runtime_context = ContextVar("runtime_config", default=None)
+runtime_progress = ContextVar("runtime_progress", default=None)
+
+
+def configuration():
+    if runtime_context.get() is not None:
+        return runtime_context.get()
+    with closing(connect()) as conn:
+        conn.execute("CREATE TABLE IF NOT EXISTS local_runtime_config (id INTEGER PRIMARY KEY CHECK(id=1), config_json TEXT NOT NULL)")
+        row = conn.execute("SELECT config_json FROM local_runtime_config WHERE id=1").fetchone()
+    return RuntimeConfig.model_validate_json(row[0]) if row else RuntimeConfig()
+
+
+def configure(request):
+    from app.database.db import transaction
+    configuration()
+    with closing(connect()) as conn, transaction(conn):
+        row = conn.execute("SELECT config_json FROM local_runtime_config WHERE id=1").fetchone()
+        previous = RuntimeConfig.model_validate_json(row[0]) if row else RuntimeConfig()
+        if request.version != previous.version:
+            raise ApiError(409, "VERSION_CONFLICT", "Local runtime settings changed; reload first.")
+        request = request.model_copy(update={"version": request.version + 1})
+        conn.execute("INSERT OR REPLACE INTO local_runtime_config VALUES (1,?)", (request.model_dump_json(),))
+    return request
+
+
+def interpreter():
+    return Path(os.getenv("APP_MODEL_PYTHON", str(BACKEND_DIR / ".venv-models" / ("Scripts/python.exe" if os.name == "nt" else "bin/python"))))
+
+
+class Runtime:
+    def __init__(self):
+        self.active = {}
+        self.active_files = {}
+        self.waiters = []
+        self.counter = 0
+        self.diagnostics = []
+
+    def in_use(self, key):
+        return key in self.active.values()
+
+    def media_in_use(self, path):
+        target = str(Path(path).resolve())
+        return any(target in paths for paths in self.active_files.values())
+
+    async def infer(self, key, operation, payload, *, priority=10):
+        if read_state(key)["status"] != "installed":
+            raise ProviderError("LOCAL_MODEL_NOT_INSTALLED", "请先在模型配置中下载本地模型。")
+        if not interpreter().is_file():
+            raise ProviderError("LOCAL_RUNTIME_NOT_INSTALLED", "请先运行本地模型 CPU/CUDA 安装脚本。")
+        config = configuration()
+        self.counter += 1
+        ticket = (priority, self.counter)
+        self.waiters.append(ticket)
+        process = None
+        attempt = None
+        try:
+            # One resident model at a time prevents overlapping CPU/GPU allocations.
+            while self.active or ticket != min(self.waiters):
+                await asyncio.sleep(0.05)
+            self.waiters.remove(ticket)
+            self.active[ticket] = key
+            self.active_files[ticket] = {str(Path(payload[name]).resolve()) for name in ("source", "reference") if payload.get(name)}
+            # Deletion may have occurred while this request was queued.
+            if read_state(key)["status"] != "installed":
+                raise ProviderError("LOCAL_MODEL_NOT_INSTALLED", "模型文件已被删除。")
+            from app.services.usage_service import UsageAttempt
+            attempt = UsageAttempt("local-models", CATALOG[key].repository, "local", operation, source="local")
+            env = {**os.environ, "HF_HUB_OFFLINE": "1", "TRANSFORMERS_OFFLINE": "1",
+                   "HF_HUB_DISABLE_TELEMETRY": "1", "OMP_NUM_THREADS": str(config.cpu_threads),
+                   "PYTHONIOENCODING": "utf-8"}
+            process = await asyncio.create_subprocess_exec(str(interpreter()), str(Path(__file__).with_name("worker.py")),
+                stdin=asyncio.subprocess.PIPE, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.DEVNULL,
+                env=env, limit=16 * 1024 * 1024, **({"creationflags": 0x08000000} if os.name == "nt" else {}))
+            request = {"key": key, "operation": operation, "model_path": str(model_path(key).resolve()),
+                       "config": config.model_dump(), "payload": payload}
+            async def receive():
+                process.stdin.write(json.dumps(request).encode())
+                await process.stdin.drain()
+                process.stdin.close()
+                final = None
+                while line := await process.stdout.readline():
+                    if len(line) > 16 * 1024 * 1024:
+                        raise ProviderError("LOCAL_MODEL_INVALID_RESPONSE", "本地模型输出超限。")
+                    message = json.loads(line)
+                    if "progress" in message:
+                        callback = runtime_progress.get()
+                        if callback:
+                            callback(message)
+                    else:
+                        final = message
+                await process.wait()
+                return final
+            try:
+                result = await asyncio.wait_for(receive(), config.timeout_seconds)
+            except TimeoutError as exc:
+                raise ProviderError("LOCAL_MODEL_TIMEOUT", "本地模型处理超时。") from exc
+            if process.returncode != 0:
+                raise ProviderError("LOCAL_MODEL_PROCESS_FAILED", "本地模型进程退出，请检查依赖与资源预算。")
+            if not isinstance(result, dict):
+                raise ProviderError("LOCAL_MODEL_INVALID_RESPONSE", "本地模型进程未返回有效结果。")
+            if "error_code" in result:
+                raise ProviderError(result["error_code"], result.get("message", "本地推理失败。"))
+            attempt.observe(result)
+            attempt.completed = True
+            self.diagnostics.append({"model": CATALOG[key].repository, "revision": CATALOG[key].revision,
+                                     **result.get("diagnostics", {})})
+            self.diagnostics = self.diagnostics[-100:]
+            return result["result"]
+        finally:
+            if ticket in self.waiters:
+                self.waiters.remove(ticket)
+            if process is not None and process.returncode is None:
+                process.kill()
+                await process.wait()
+            self.active.pop(ticket, None)
+            self.active_files.pop(ticket, None)
+            if attempt:
+                attempt.persist()
+
+
+runtime = Runtime()
+
+
+class LocalEmbedding:
+    dim = 384
+
+    @property
+    def model_id(self):
+        spec = CATALOG[configuration().embedding_model]
+        return f"{spec.repository}@{spec.revision}"
+
+    @property
+    def version(self):
+        return CATALOG[configuration().embedding_model].revision
+
+    @property
+    def available(self):
+        return read_state(configuration().embedding_model)["status"] == "installed" and interpreter().is_file()
+
+    async def embed_documents(self, texts):
+        return await runtime.infer(configuration().embedding_model, "embedding", {"texts": texts}, priority=0)
+
+    async def embed_query(self, query):
+        return (await self.embed_documents([query]))[0]
+
+
+class LocalSpeech:
+    @property
+    def available(self):
+        return self.available_for("transcription")
+
+    def available_for(self, capability):
+        key = "qwen3-asr" if capability == "transcription" else "eres2netv2"
+        return read_state(key)["status"] == "installed" and interpreter().is_file()
+
+    async def transcribe(self, source, language):
+        from app.providers.routing import RoutedTranscript
+        from app.contracts import TranscriptSegment
+        result = await runtime.infer("qwen3-asr", "transcription", {"source": str(source.resolve()), "language": language})
+        return RoutedTranscript(text=result["text"], source="local",
+                                segments=[TranscriptSegment(**s) for s in result["segments"]])
+
+    async def match(self, source, reference):
+        result = await runtime.infer("eres2netv2", "speaker_matching",
+            {"source": str(source.resolve()), "reference": str(reference.resolve())}, priority=0)
+        return result["score"]

backend/app/local_models/worker.py

@@ -0,0 +1,175 @@
+"""One offline inference process. Heavy libraries stay out of the API process."""
+from __future__ import annotations
+
+import contextlib
+import json
+import os
+import sys
+import threading
+import time
+
+
+def decode(path, *, limit_seconds=3600):
+    import av
+    import numpy as np
+    frames = []
+    samples = 0
+    with av.open(path, options={"protocol_whitelist": "file,pipe"}) as container:
+        if not container.streams.audio:
+            raise ValueError("Media has no audio track")
+        resampler = av.AudioResampler(format="fltp", layout="mono", rate=16000)
+        for frame in container.decode(audio=0):
+            for output in resampler.resample(frame):
+                audio = output.to_ndarray().reshape(-1)
+                samples += len(audio)
+                if samples > limit_seconds * 16000:
+                    raise ValueError("Audio exceeds one hour")
+                frames.append(audio)
+        for output in resampler.resample(None):
+            frames.append(output.to_ndarray().reshape(-1))
+    if not frames:
+        raise ValueError("Audio is empty")
+    audio = np.concatenate(frames).astype(np.float32)
+    if not np.isfinite(audio).all() or len(audio) < 1600:
+        raise ValueError("Invalid or too short audio")
+    return audio
+
+
+def speech_regions(audio):
+    """Energy-based segmentation, not word alignment; retain original sample offsets."""
+    import numpy as np
+    window = 480
+    energies = [float(np.sqrt(np.mean(audio[i:i + window] ** 2))) for i in range(0, len(audio), window)]
+    threshold = max(0.002, float(np.percentile(energies, 20)) * 2)
+    active = [i for i, energy in enumerate(energies) if energy >= threshold]
+    if not active:
+        return []
+    regions, start, previous = [], active[0], active[0]
+    for index in active[1:]:
+        if index - previous > 20 or (index - start) * window >= 20 * 16000:
+            regions.append((max(0, start * window - 2400), min(len(audio), (previous + 1) * window + 2400)))
+            start = index
+        previous = index
+    regions.append((max(0, start * window - 2400), min(len(audio), (previous + 1) * window + 2400)))
+    return regions
+
+
+def speaker_model(path, device):
+    import torch
+    from modelscope.models.audio.sv.ERes2NetV2 import ERes2NetV2
+    from pathlib import Path
+    model = ERes2NetV2(baseWidth=26, scale=2, expansion=2, embed_dim=192)
+    weights = torch.load(Path(path) / "pretrained_eres2netv2.ckpt", map_location="cpu", weights_only=True)
+    model.load_state_dict(weights, strict=True)
+    return model.to(device).eval()
+
+
+def voice_embedding(model, audio, device):
+    import torch
+    import torchaudio.compliance.kaldi as kaldi
+    if len(audio) < 16000:
+        raise ValueError("Speaker comparison needs at least one second of audio")
+    features = kaldi.fbank(torch.from_numpy(audio).unsqueeze(0), num_mel_bins=80, sample_frequency=16000)
+    features -= features.mean(dim=0, keepdim=True)
+    with torch.inference_mode():
+        vector = model(features.unsqueeze(0).to(device)).flatten()
+        return torch.nn.functional.normalize(vector, dim=0)
+
+
+def run(request):
+    import torch
+    import psutil
+    config, payload = request["config"], request["payload"]
+    torch.set_num_threads(config["cpu_threads"])
+    requested = config["device"]
+    device = "cuda:0" if requested == "cuda" and torch.cuda.is_available() else "cpu"
+    if device != "cpu":
+        total = torch.cuda.get_device_properties(0).total_memory
+        torch.cuda.set_per_process_memory_fraction(min(1.0, config["gpu_memory_limit_mb"] * 1024 ** 2 / total))
+    process = psutil.Process()
+    peak = [0]
+    stop = threading.Event()
+
+    def monitor():
+        while not stop.wait(0.2):
+            used = process.memory_info().rss
+            peak[0] = max(peak[0], used)
+            if used > config["memory_limit_mb"] * 1024 ** 2:
+                os._exit(75)
+
+    threading.Thread(target=monitor, daemon=True).start()
+    started = time.monotonic()
+    path, operation = request["model_path"], request["operation"]
+    try:
+        usage = {}
+        if operation == "embedding":
+            from sentence_transformers import SentenceTransformer
+            model = SentenceTransformer(path, device=device, local_files_only=True, trust_remote_code=False,
+                                        model_kwargs={"attn_implementation": "sdpa"})
+            loaded = time.monotonic()
+            result = model.encode(payload["texts"], batch_size=4, normalize_embeddings=True, show_progress_bar=False).tolist()
+            # Count the tokenizer's actual encoded input, not characters or words.
+            usage = {"input_tokens": int(model.tokenize(payload["texts"])["attention_mask"].sum())}
+        elif operation == "transcription":
+            from qwen_asr import Qwen3ASRModel
+            model = Qwen3ASRModel.from_pretrained(path, dtype=torch.float32 if device == "cpu" else torch.float16,
+                device_map=device, attn_implementation="sdpa", max_inference_batch_size=1, max_new_tokens=512)
+            loaded = time.monotonic()
+            audio = decode(payload["source"])
+            regions = speech_regions(audio)
+            language = {"zh": "Chinese", "en": "English", "ja": "Japanese", "yue": "Cantonese"}.get(payload.get("language"), payload.get("language"))
+            segments = []
+            for start, end in regions:
+                output = model.transcribe(audio=(audio[start:end], 16000), language=language)[0]
+                if output.text.strip():
+                    segments.append({"segment_id": f"segment_{len(segments) + 1}", "start_time": start / 16000,
+                                     "end_time": end / 16000, "text": output.text, "language": output.language})
+                    sys.__stdout__.write(json.dumps({"progress": end / len(audio), "segment": segments[-1]}, ensure_ascii=False) + "\n")
+                    sys.__stdout__.flush()
+            result = {"text": "\n".join(s["text"] for s in segments), "segments": segments}
+        elif operation == "speaker_matching":
+            model = speaker_model(path, device)
+            loaded = time.monotonic()
+            first = voice_embedding(model, decode(payload["source"]), device)
+            second = voice_embedding(model, decode(payload["reference"]), device)
+            # Similarity, not a calibrated identity probability.
+            result = {"score": max(0.0, min(1.0, float(torch.dot(first, second))))}
+        elif operation == "diarization":
+            model = speaker_model(path, device)
+            loaded = time.monotonic()
+            audio = decode(payload["source"])
+            centroids, speakers = [], []
+            for segment in payload["segments"]:
+                sample = audio[int(segment["start_time"] * 16000):int(segment["end_time"] * 16000)]
+                if len(sample) < 16000:
+                    speakers.append(None)
+                    continue
+                vector = voice_embedding(model, sample, device)
+                similarities = [float(torch.dot(vector, c)) for c in centroids]
+                best = max(range(len(similarities)), key=similarities.__getitem__) if similarities else None
+                if best is None or similarities[best] < 0.36:
+                    best = len(centroids)
+                    centroids.append(vector)
+                speakers.append(f"speaker_{best + 1}")
+            result = {"speakers": speakers}
+        else:
+            raise ValueError("Unknown inference operation")
+        return {"result": result, "usage": usage, "diagnostics": {"requested_device": requested, "actual_device": device,
+                "fallback_reason": "CUDA_UNAVAILABLE" if requested == "cuda" and device == "cpu" else None,
+                "load_seconds": loaded - started, "inference_seconds": time.monotonic() - loaded,
+                "peak_memory_bytes": max(peak[0], process.memory_info().rss), "operation": operation}}
+    finally:
+        stop.set()
+
+
+if __name__ == "__main__":
+    request = json.loads(sys.stdin.buffer.read())
+    # Third-party progress/logging must never corrupt the protocol or leak into API errors.
+    with contextlib.redirect_stdout(sys.stderr):
+        try:
+            response = run(request)
+        except (ImportError, ModuleNotFoundError):
+            response = {"error_code": "LOCAL_RUNTIME_DEPENDENCY_MISSING", "message": "本地模型运行依赖不完整，请重新运行安装脚本。"}
+        except Exception:
+            response = {"error_code": "LOCAL_INFERENCE_FAILED", "message": "本地推理失败，请检查媒体格式、模型和设备配置。"}
+    sys.stdout.buffer.write((json.dumps(response, ensure_ascii=False, allow_nan=False) + "\n").encode("utf-8"))