blumia
bd921b1ac5
Add a second backend (sherpa-onnx) alongside llama.cpp, so native streaming models can be selected by ID from a JSON registry. All such models share one adapter: only the model files and a few options differ. Models now live under the app data directory and are fetched by scripts/fetch-models.sh. Adding a model is a data change, not code. Measured on the same audio: the streaming backend is ~35x cheaper in CPU than Qwen3-ASR (0.10-0.12 vs 4.27 core-seconds per audio second), shows the first caption at 0.48-1.44s, and matches its accuracy for zh/en. 新增 sherpa-onnx 流式后端，与 llama.cpp 并存；模型按注册表 id 从应用 数据目录加载。同族模型共用一个适配器，新增模型只改 data/models.json。 实测流式后端 CPU 约为 Qwen3-ASR 的 1/35，首个字幕 0.48~1.44s，中英 准确率持平。注意预编译库是旧 C++ ABI，必须走其 C API。
172 lines
7.0 KiB
C++
172 lines
7.0 KiB
C++
#pragma once
|
||
|
||
#include <cstdio>
|
||
#include <memory>
|
||
#include <string>
|
||
|
||
#include <QCommandLineOption>
|
||
#include <QCommandLineParser>
|
||
|
||
#include "iasr.h"
|
||
#include "modelstore.h"
|
||
#include "streamingasr.h"
|
||
#include "qwen3asr.h"
|
||
#if defined(LC_HAVE_SHERPA)
|
||
#include "sherpaasr.h"
|
||
#endif
|
||
|
||
// 两个可执行程序（GUI / CLI）共用的命令行选项定义与解析。
|
||
namespace lc::app {
|
||
|
||
// 构造命令行选项。
|
||
//
|
||
// Qt 自己不会把默认值印进 --help（构造函数里传的默认值只作为 value() 的兜底），
|
||
// 所以这里在给了默认值时把「（默认 X）」拼到说明后面。
|
||
inline QCommandLineOption makeOption(const QString & name, const QString & description,
|
||
const QString & valueName = QString(),
|
||
const QString & defaultValue = QString())
|
||
{
|
||
QString desc = description;
|
||
if (!defaultValue.isEmpty()) {
|
||
desc += QStringLiteral("（默认 %1）").arg(defaultValue);
|
||
}
|
||
if (valueName.isEmpty()) return QCommandLineOption(name, desc);
|
||
return QCommandLineOption(name, desc, valueName, defaultValue);
|
||
}
|
||
|
||
inline void addModelOptions(QCommandLineParser & p)
|
||
{
|
||
p.addOption(makeOption("model", "主模型 GGUF 路径（架构须为 qwen3vl）", "path"));
|
||
p.addOption(makeOption("mmproj", "mmproj GGUF 路径（clip + qwen3a）", "path"));
|
||
p.addOption(makeOption("threads", "推理线程数", "n", "8"));
|
||
p.addOption(makeOption("max-tokens", "单段最多生成 token 数", "n", "256"));
|
||
p.addOption(makeOption("verbose", "输出 llama.cpp/mtmd 的详细日志"));
|
||
}
|
||
|
||
inline void addStreamOptions(QCommandLineParser & p)
|
||
{
|
||
p.addOption(makeOption("stream-chunk", "流式 chunk / 提交粒度，单位秒", "sec", "2.0"));
|
||
p.addOption(makeOption("stream-window", "滑窗上限，单位秒；仅 window 模式，0=不滑窗", "sec", "16"));
|
||
p.addOption(makeOption("stream-segment", "单段最长时长，单位秒；0=不限", "sec", "30"));
|
||
p.addOption(makeOption("unfixed-tokens", "前缀回滚丢弃的 token 数", "n", "5"));
|
||
p.addOption(makeOption("unfixed-chunks", "冷启动不带前缀的 chunk 数", "n", "2"));
|
||
p.addOption(makeOption("stream-mode", "流式后端: incremental / append / window", "mode",
|
||
"incremental"));
|
||
}
|
||
|
||
inline AsrOptions asrOptions(const QCommandLineParser & p)
|
||
{
|
||
AsrOptions o;
|
||
o.modelPath = p.value("model").toStdString();
|
||
o.mmprojPath = p.value("mmproj").toStdString();
|
||
o.threads = p.value("threads").toInt();
|
||
o.maxTokens = p.value("max-tokens").toInt();
|
||
o.verbose = p.isSet("verbose");
|
||
return o;
|
||
}
|
||
|
||
inline StreamOptions streamOptions(const QCommandLineParser & p)
|
||
{
|
||
StreamOptions o;
|
||
o.chunkSec = p.value("stream-chunk").toDouble();
|
||
o.windowSec = p.value("stream-window").toDouble();
|
||
o.maxUtteranceSec = p.value("stream-segment").toDouble();
|
||
o.unfixedTokenNum = p.value("unfixed-tokens").toInt();
|
||
o.unfixedChunkNum = p.value("unfixed-chunks").toInt();
|
||
o.maxTokensPerStep = p.value("max-tokens").toInt();
|
||
o.incremental = p.value("stream-mode") == QLatin1String("incremental");
|
||
o.appendOnly = p.value("stream-mode") == QLatin1String("append");
|
||
return o;
|
||
}
|
||
|
||
inline bool loadModel(Qwen3Asr & asr, const AsrOptions & opt)
|
||
{
|
||
if (opt.modelPath.empty() || opt.mmprojPath.empty()) {
|
||
std::fprintf(stderr, "必须提供 --model 和 --mmproj\n");
|
||
return false;
|
||
}
|
||
std::string err;
|
||
if (!asr.load(opt, &err)) {
|
||
std::fprintf(stderr, "加载失败: %s\n", err.c_str());
|
||
return false;
|
||
}
|
||
return true;
|
||
}
|
||
|
||
// ---- 推理后端选择 ----
|
||
//
|
||
// llama 后端：Qwen3-ASR 的 GGUF 模型（--model / --mmproj）
|
||
// sherpa 后端：流式 ONNX 模型，按注册表 id 从模型存储目录取（--asr-model <id>）
|
||
inline void addBackendOptions(QCommandLineParser & p)
|
||
{
|
||
p.addOption(makeOption("asr-backend", "推理后端: llama（Qwen3-ASR GGUF）或 sherpa（流式 ONNX）",
|
||
"kind", "llama"));
|
||
p.addOption(makeOption("asr-model", "模型注册表里的 id（sherpa 后端用它定位文件）", "id"));
|
||
p.addOption(makeOption("model-dir", "覆盖模型存储目录（默认 ~/.local/share/live-caption/models）", "dir"));
|
||
p.addOption(makeOption("language", "语言提示，多语言模型用（zh / en / ja …）；留空=自动", "lang"));
|
||
}
|
||
|
||
// 按命令行选择并加载后端；失败时返回 nullptr 并把原因写进 err
|
||
inline std::unique_ptr<IAsr> makeAsr(const QCommandLineParser & p,
|
||
const AsrOptions & asrOpt, std::string * err)
|
||
{
|
||
const QString dir = p.value("model-dir");
|
||
if (!dir.isEmpty()) models::ModelStore::setRootOverride(dir);
|
||
|
||
const QString backend = p.value("asr-backend");
|
||
const QString modelId = p.value("asr-model");
|
||
const QString lang = p.value("language");
|
||
|
||
const models::ModelInfo * info =
|
||
modelId.isEmpty() ? nullptr : models::ModelStore::find(modelId);
|
||
if (!modelId.isEmpty() && !info) {
|
||
if (err) *err = "注册表里没有模型: " + modelId.toStdString();
|
||
return nullptr;
|
||
}
|
||
|
||
const bool useSherpa = (backend == QLatin1String("sherpa"))
|
||
|| (info && info->backend == models::Backend::SherpaOnline);
|
||
|
||
if (useSherpa) {
|
||
#if defined(LC_HAVE_SHERPA)
|
||
if (!info) {
|
||
if (err) *err = "sherpa 后端需要 --asr-model <id>（见 data/models.json）";
|
||
return nullptr;
|
||
}
|
||
QStringList missing;
|
||
if (!models::ModelStore::isInstalled(*info, &missing)) {
|
||
if (err) {
|
||
*err = "模型未安装，缺失: " + missing.join(QStringLiteral(", ")).toStdString()
|
||
+ "\n 可运行: scripts/fetch-models.sh " + info->id.toStdString();
|
||
}
|
||
return nullptr;
|
||
}
|
||
SherpaAsr::Options o;
|
||
o.encoder = info->filePath(QStringLiteral("encoder")).toStdString();
|
||
o.decoder = info->filePath(QStringLiteral("decoder")).toStdString();
|
||
o.joiner = info->filePath(QStringLiteral("joiner")).toStdString();
|
||
o.tokens = info->filePath(QStringLiteral("tokens")).toStdString();
|
||
o.threads = 1; // 这类模型单线程最快（多线程反而被同步开销拖慢）
|
||
if (info->languages.size() > 1 && !lang.isEmpty()) {
|
||
o.language = lang.toStdString(); // 多语言模型（Nemotron）需要语言提示
|
||
}
|
||
auto a = std::make_unique<SherpaAsr>();
|
||
if (!a->load(o, err)) return nullptr;
|
||
return a;
|
||
#else
|
||
if (err) *err = "编译时未找到 sherpa-onnx（-DSHERPA_ONNX_ROOT=...）";
|
||
return nullptr;
|
||
#endif
|
||
}
|
||
|
||
auto a = std::make_unique<Qwen3Asr>();
|
||
if (asrOpt.modelPath.empty() || asrOpt.mmprojPath.empty()) {
|
||
if (err) *err = "llama 后端需要 --model 和 --mmproj（或改用 --asr-backend sherpa --asr-model <id>）";
|
||
return nullptr;
|
||
}
|
||
if (!a->load(asrOpt, err)) return nullptr;
|
||
return a;
|
||
}
|
||
|
||
} // namespace lc::app
|