Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

docs(site): publish llama.cpp v0.2.0 downloads (#3483)

zhifu gao9e983c6c4ab4
diff --git a/web-pages/product-site/data/deployments.json b/web-pages/product-site/data/deployments.json
index 96d39d107..325575551 100644
--- a/web-pages/product-site/data/deployments.json
+++ b/web-pages/product-site/data/deployments.json
@@ -1,6 +1,6 @@
{
"schema_version": 1,
- "verified": "2026-08-04",
+ "verified": "2026-08-11",
"deployments": [
{
"id": "vllm",
@@ -157,16 +157,28 @@
"models": ["SenseVoiceSmall-GGUF", "Paraformer-GGUF", "Fun-ASR-Nano-GGUF", "FSMN-VAD-GGUF"],
"operating_systems": ["Linux", "macOS", "Windows"],
"interfaces": ["CLI", "local HTTP server"],
- "tested": {"funasr": "runtime-llamacpp-v0.1.9", "runtime": "pinned llama.cpp backend", "verified": "2026-07-26"},
+ "tested": {"funasr": "runtime-llamacpp-v0.2.0", "runtime": "llama.cpp@803b7fca", "verified": "2026-08-11"},
"commands": {
- "install": ["cmake -B build -DCMAKE_BUILD_TYPE=Release", "cmake --build build -j"],
- "launch": ["./build/bin/llama-funasr-sensevoice -m funasr-gguf/sensevoice-small-q8.gguf --vad funasr-gguf/fsmn-vad.gguf -a sample.wav"],
- "health": ["./build/bin/llama-funasr-sensevoice --help"],
- "smoke": ["./runtime/llama.cpp/tests/run_regression.sh"]
+ "install": ["curl -fLO https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.2.0/funasr-llamacpp-linux-x64.tar.gz", "echo \"15e6407143b4fb91d90bb37f2a41c64c4d48ea0fbe6404b88a9b70269c84f240 funasr-llamacpp-linux-x64.tar.gz\" | sha256sum -c -", "mkdir funasr-llamacpp && tar -xzf funasr-llamacpp-linux-x64.tar.gz -C funasr-llamacpp && cd funasr-llamacpp && bash download-funasr-model.sh sensevoice ./funasr-gguf f16"],
+ "launch": ["cd funasr-llamacpp && ./llama-funasr-sensevoice -m funasr-gguf/sensevoice-small-f16.gguf --vad funasr-gguf/fsmn-vad.gguf -a sample.wav"],
+ "health": ["cd funasr-llamacpp && ./llama-funasr-sensevoice --help"],
+ "smoke": ["cd funasr-llamacpp && ./llama-funasr-sensevoice -m funasr-gguf/sensevoice-small-f16.gguf --vad funasr-gguf/fsmn-vad.gguf -a sample.wav | tee transcript.txt && test -s transcript.txt"]
},
+ "downloads": [
+ {"operating_system": "Linux", "architecture": "arm64", "backend": "CPU", "archive": "funasr-llamacpp-linux-arm64.tar.gz", "url": "https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.2.0/funasr-llamacpp-linux-arm64.tar.gz", "sha256": "c78987b2384c6aef339aea1bcd0e130070455d6394fa7ab7ca26840ead10d5da"},
+ {"operating_system": "Linux", "architecture": "x64", "backend": "CPU", "archive": "funasr-llamacpp-linux-x64.tar.gz", "url": "https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.2.0/funasr-llamacpp-linux-x64.tar.gz", "sha256": "15e6407143b4fb91d90bb37f2a41c64c4d48ea0fbe6404b88a9b70269c84f240"},
+ {"operating_system": "Linux", "architecture": "x64 AVX2", "backend": "CPU", "archive": "funasr-llamacpp-linux-x64-avx2.tar.gz", "url": "https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.2.0/funasr-llamacpp-linux-x64-avx2.tar.gz", "sha256": "02e10e9a46ea76a040c45d431efe51a3324e64f08c24d38e18c8a4d2781490cd"},
+ {"operating_system": "Linux", "architecture": "x64", "backend": "Vulkan", "archive": "funasr-llamacpp-linux-x64-vulkan.tar.gz", "url": "https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.2.0/funasr-llamacpp-linux-x64-vulkan.tar.gz", "sha256": "caf71b8c0b4c3249cebc4175e5406d3c588eb9e8966a00d571d4cc5070405385"},
+ {"operating_system": "macOS", "architecture": "arm64", "backend": "CPU", "archive": "funasr-llamacpp-macos-arm64.tar.gz", "url": "https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.2.0/funasr-llamacpp-macos-arm64.tar.gz", "sha256": "416cbb289e31cb7575365d382155074e922fd061807a37b9ca0247dabd9bc6f9"},
+ {"operating_system": "Windows", "architecture": "x64", "backend": "CPU", "archive": "funasr-llamacpp-windows-x64.zip", "url": "https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.2.0/funasr-llamacpp-windows-x64.zip", "sha256": "297c962346d7e30d7a7c2c860dfaab3ff07d01fddf15e6fc5212ca9545441a51"},
+ {"operating_system": "Windows", "architecture": "x64 AVX2", "backend": "CPU", "archive": "funasr-llamacpp-windows-x64-avx2.zip", "url": "https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.2.0/funasr-llamacpp-windows-x64-avx2.zip", "sha256": "4db0f11f603c324a63545cd7009cdd45bb45576efe282cec22796b5fd42d8ea1"},
+ {"operating_system": "Windows", "architecture": "x64", "backend": "Vulkan", "archive": "funasr-llamacpp-windows-x64-vulkan.zip", "url": "https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.2.0/funasr-llamacpp-windows-x64-vulkan.zip", "sha256": "90b45240c6ccc9177c25490a11848de60a406e129391c8736b14521c0c28cdcb"},
+ {"operating_system": "Windows", "architecture": "x64", "backend": "CUDA", "archive": "funasr-llamacpp-windows-x64-cuda.zip", "url": "https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.2.0/funasr-llamacpp-windows-x64-cuda.zip", "sha256": "7f2f9ef4d7e0291b284a295ec74bbeca9ea635a7f5f42d0ad06eb780c0d6efc1"}
+ ],
"evidence": [
{"label": "llama.cpp runtime", "url": "https://github.com/modelscope/FunASR/blob/main/runtime/llama.cpp/README.md"},
- {"label": "runtime release", "url": "https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.9"},
+ {"label": "runtime v0.2.0 release", "url": "https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.0"},
+ {"label": "nine-platform release workflow", "url": "https://github.com/modelscope/FunASR/actions/runs/31458121788"},
{"label": "regression tests", "url": "https://github.com/modelscope/FunASR/tree/main/runtime/llama.cpp/tests"}
],
"benchmarks": [
@@ -200,27 +212,27 @@
"translations": {
"zh": {
"name": "llama.cpp / GGUF 独立运行",
- "summary": "使用跨平台预编译包或源码构建,在 CPU、桌面 GPU 和边缘设备上运行 FunASR GGUF 模型。",
- "fit": ["不依赖 Python ML 环境", "桌面应用和离线边缘部署", "需要 Linux、macOS、Windows 发布包"],
+ "summary": "使用 v0.2.0 的九个预编译包或源码构建,在 CPU、Vulkan、CUDA 和边缘设备上运行 FunASR GGUF 模型。",
+ "fit": ["不依赖 Python ML 环境", "桌面应用和离线边缘部署", "需要 Linux、macOS、Windows 的 CPU、Vulkan 或 CUDA 发布包"],
"not_fit": ["需要 vLLM 式大批量 GPU 调度", "未验证目标 GPU 架构的通用预编译 CUDA 包", "必须使用完整 Python 模型生态的流程"],
"selection_reason": "GGUF 运行时和独立二进制优先满足可移植、离线和低依赖部署。",
- "primary_limitation": "预编译 GPU 包只覆盖标注的后端和架构;其他设备需要从源码构建并实机验证。",
+ "primary_limitation": "预编译 GPU 包只覆盖标注的后端和架构;Windows AMD Vulkan 的崩溃修复仍需问题报告者实机复测,其他设备也必须在目标硬件验证。",
"status_label": "生产验证",
- "operations": ["按发布资产 SHA-256 校验下载", "模型和二进制使用同一发布清单", "保留旧二进制和模型目录用于回滚"],
+ "operations": ["按页面列出的 SHA-256 校验九个发布资产", "模型和二进制使用同一发布清单", "保留旧二进制和模型目录用于回滚"],
"security": ["默认只读取本地音频和模型", "HTTP 服务绑定内网地址并限制上传大小", "不要从不可信地址加载 GGUF"],
- "troubleshooting": ["CPU 路径先用 q8 模型验证", "CUDA 和 Vulkan 必须匹配本机驱动", "用仓库回归脚本区分模型问题与构建问题"]
+ "troubleshooting": ["CPU 路径先用 q8 模型验证", "CUDA 和 Vulkan 必须匹配本机驱动", "Windows AMD Vulkan 异常时记录 GPU、驱动、GGML_VK_MAX_NODES_PER_SUBMIT 和 GGML_VK_SERIALIZE_SUBMISSIONS 后回报"]
},
"en": {
"name": "llama.cpp / GGUF standalone",
- "summary": "Use cross-platform release packages or source builds to run FunASR GGUF models on CPUs, desktop GPUs, and edge devices.",
- "fit": ["No Python ML environment", "Desktop applications and offline edge deployment", "Linux, macOS, and Windows release packages"],
+ "summary": "Use the nine v0.2.0 release packages or source builds to run FunASR GGUF models on CPU, Vulkan, CUDA, and edge devices.",
+ "fit": ["No Python ML environment", "Desktop applications and offline edge deployment", "CPU, Vulkan, or CUDA packages for Linux, macOS, and Windows"],
"not_fit": ["vLLM-style large GPU batch scheduling", "A universal prebuilt CUDA package for an unverified GPU architecture", "Workflows that require the full Python model ecosystem"],
"selection_reason": "GGUF runtimes and standalone binaries prioritize portability, offline use, and low dependency count.",
- "primary_limitation": "Prebuilt GPU packages cover only the documented backend and architecture; other devices require a source build and hardware validation.",
+ "primary_limitation": "Prebuilt GPU packages cover only the documented backend and architecture; the Windows AMD Vulkan crash fix still awaits reporter hardware confirmation, and every other target also requires hardware validation.",
"status_label": "Production verified",
- "operations": ["Verify release assets with SHA-256", "Keep models and binaries on the same release manifest", "Retain the previous binary and model directory for rollback"],
+ "operations": ["Verify all nine release assets against the listed SHA-256 values", "Keep models and binaries on the same release manifest", "Retain the previous binary and model directory for rollback"],
"security": ["Read local audio and models by default", "Bind the HTTP server to a private address and limit uploads", "Do not load GGUF files from untrusted sources"],
- "troubleshooting": ["Validate the q8 CPU path first", "Match CUDA or Vulkan to the local driver", "Use repository regression scripts to separate model and build failures"]
+ "troubleshooting": ["Validate the q8 CPU path first", "Match CUDA or Vulkan to the local driver", "For Windows AMD Vulkan failures, record the GPU, driver, GGML_VK_MAX_NODES_PER_SUBMIT, and GGML_VK_SERIALIZE_SUBMISSIONS before reporting"]
}
}
},
diff --git a/web-pages/product-site/registry.py b/web-pages/product-site/registry.py
index cb668a0a5..b485314b0 100644
--- a/web-pages/product-site/registry.py
+++ b/web-pages/product-site/registry.py
@@ -3,6 +3,7 @@
from __future__ import annotations
import json
+import re
from pathlib import Path
from typing import Any
from urllib.parse import urlparse
@@ -45,6 +46,7 @@ BENCHMARK_FIELDS = (
'source',
'verified',
)
+DOWNLOAD_FIELDS = ('operating_system', 'architecture', 'backend', 'archive', 'url', 'sha256')
def load_registry(path: Path) -> dict[str, Any]:
@@ -161,6 +163,30 @@ def validate_registry(data: dict[str, Any]) -> list[str]:
f'{label}: production-verified entry requires {field_name}'
)
+ downloads = entry.get('downloads', [])
+ if not isinstance(downloads, list):
+ errors.append(f'{label}: downloads must be a list')
+ else:
+ for download_index, download in enumerate(downloads):
+ if not isinstance(download, dict):
+ errors.append(f'{label}: download {download_index} must be an object')
+ continue
+ for field in DOWNLOAD_FIELDS:
+ if not download.get(field):
+ errors.append(
+ f'{label}: download {download_index} requires {field}'
+ )
+ if not _is_https_url(download.get('url')):
+ errors.append(
+ f'{label}: download URL must use https (item {download_index})'
+ )
+ digest = download.get('sha256')
+ if not isinstance(digest, str) or not re.fullmatch(r'[0-9a-f]{64}', digest):
+ errors.append(
+ f'{label}: download SHA-256 must be 64 lowercase hex characters '
+ f'(item {download_index})'
+ )
+
benchmarks = entry.get('benchmarks', [])
if not isinstance(benchmarks, list):
errors.append(f'{label}: benchmarks must be a list')
diff --git a/web-pages/product-site/templates/deploy-detail.html b/web-pages/product-site/templates/deploy-detail.html
index 0b1502cba..400e4ba37 100644
--- a/web-pages/product-site/templates/deploy-detail.html
+++ b/web-pages/product-site/templates/deploy-detail.html
@@ -46,6 +46,39 @@
</div>
</section>
+{% set downloads = entry.get('downloads', []) %}
+{% if downloads %}
+<section class="section content-section" data-section="downloads">
+ <div class="section-heading compact-heading">
+ <p class="eyebrow">{{ '发布资产' if language == 'zh' else 'Release assets' }}</p>
+ <h2>{{ '按操作系统和后端下载已校验的预编译包' if language == 'zh' else 'Download a verified package for your operating system and backend' }}</h2>
+ <p>{{ 'SHA-256 来自公开 GitHub Release;下载后先核对摘要,再在目标硬件完成 smoke test。' if language == 'zh' else 'SHA-256 values come from the public GitHub Release. Verify the digest, then run a smoke test on the target hardware.' }}</p>
+ </div>
+ <div class="table-wrap download-table">
+ <table>
+ <thead><tr>
+ <th>{{ '系统' if language == 'zh' else 'OS' }}</th>
+ <th>{{ '架构' if language == 'zh' else 'Architecture' }}</th>
+ <th>{{ '后端' if language == 'zh' else 'Backend' }}</th>
+ <th>{{ '下载' if language == 'zh' else 'Download' }}</th>
+ <th>SHA-256</th>
+ </tr></thead>
+ <tbody>
+ {% for item in downloads %}
+ <tr data-download-asset>
+ <td>{{ item.operating_system }}</td>
+ <td>{{ item.architecture }}</td>
+ <td>{{ item.backend }}</td>
+ <td><a href="{{ item.url }}">{{ item.archive }}</a></td>
+ <td><code data-field="sha256">{{ item.sha256 }}</code></td>
+ </tr>
+ {% endfor %}
+ </tbody>
+ </table>
+ </div>
+</section>
+{% endif %}
+
<section id="commands" class="section content-section command-section" data-section="commands">
<div class="section-heading compact-heading">
<p class="eyebrow">{{ '运行路径' if language == 'zh' else 'Run path' }}</p>
diff --git a/web-pages/product-site/tests/browser/product-site.spec.ts b/web-pages/product-site/tests/browser/product-site.spec.ts
index 0470f1f8a..18326fa76 100644
--- a/web-pages/product-site/tests/browser/product-site.spec.ts
+++ b/web-pages/product-site/tests/browser/product-site.spec.ts
@@ -90,6 +90,45 @@ test('selector, language peers, copy, and compatibility routes work', async ({ b
await context.close();
});
+for (const viewport of [
+ { name: 'mobile', width: 390, height: 844 },
+ { name: 'desktop', width: 1440, height: 900 },
+]) {
+ test(`llama.cpp v0.2.0 download matrix is stable at ${viewport.name}`, async ({ page }, testInfo) => {
+ await page.setViewportSize(viewport);
+ await page.goto('/deploy/llama-cpp.html');
+
+ const section = page.locator('[data-section="downloads"]');
+ await expect(section.locator('[data-download-asset]')).toHaveCount(9);
+ await expect(section.locator('a[href*="runtime-llamacpp-v0.2.0"]')).toHaveCount(9);
+ await expect(page.getByText('Windows AMD Vulkan', { exact: false }).first()).toBeVisible();
+
+ await section.evaluate((node) => node.scrollIntoView({ block: 'start' }));
+
+ const layout = await page.evaluate(() => {
+ const tableWrap = document.querySelector<HTMLElement>('.download-table');
+ const header = document.querySelector<HTMLElement>('.site-header');
+ const heading = document.querySelector<HTMLElement>('[data-section="downloads"] .section-heading');
+ return {
+ overflow: document.documentElement.scrollWidth - document.documentElement.clientWidth,
+ tableClientWidth: tableWrap?.clientWidth ?? 0,
+ tableScrollWidth: tableWrap?.scrollWidth ?? 0,
+ headerBottom: header?.getBoundingClientRect().bottom ?? 0,
+ headingTop: heading?.getBoundingClientRect().top ?? 0,
+ };
+ });
+ expect(layout.overflow).toBeLessThanOrEqual(1);
+ expect(layout.tableClientWidth).toBeGreaterThan(0);
+ expect(layout.tableScrollWidth).toBeGreaterThanOrEqual(layout.tableClientWidth);
+ expect(layout.headingTop).toBeGreaterThanOrEqual(layout.headerBottom + 8);
+
+ await page.screenshot({
+ path: testInfo.outputPath(`llama-cpp-downloads-${viewport.name}.png`),
+ fullPage: true,
+ });
+ });
+}
+
for (const viewport of [
{ name: 'mobile', width: 390, height: 844 },
{ name: 'desktop', width: 1440, height: 900 },
diff --git a/web-pages/product-site/tests/test_output.py b/web-pages/product-site/tests/test_output.py
index 748e9fb37..2c3af8357 100644
--- a/web-pages/product-site/tests/test_output.py
+++ b/web-pages/product-site/tests/test_output.py
@@ -58,6 +58,25 @@ def test_detail_commands_come_from_registry(built_site):
assert command in rendered
+ ('relative', 'boundary'),
+ (
+ ('deploy/llama-cpp.html', 'Windows AMD'),
+ ('en/deploy/llama-cpp.html', 'Windows AMD'),
+ ),
+)
+def test_llama_cpp_pages_render_v020_download_matrix(built_site, relative, boundary):
+ soup = read_soup(built_site / relative)
+ section = soup.select_one('[data-section="downloads"]')
+
+ assert section
+ rows = section.select('[data-download-asset]')
+ assert len(rows) == 9
+ assert all(row.select_one('a[href*="runtime-llamacpp-v0.2.0"]') for row in rows)
+ assert all(len(row.select_one('[data-field="sha256"]').get_text(strip=True)) == 64 for row in rows)
+ assert boundary in soup.get_text(' ', strip=True)
+
+
def test_benchmark_rows_have_complete_conditions(built_site):
registry = load_registry(SITE_ROOT / 'data' / 'deployments.json')
records = [record for entry in registry['deployments'] for record in entry['benchmarks']]
diff --git a/web-pages/product-site/tests/test_registry.py b/web-pages/product-site/tests/test_registry.py
index dd755f9bc..3bd14e362 100644
--- a/web-pages/product-site/tests/test_registry.py
+++ b/web-pages/product-site/tests/test_registry.py
@@ -118,6 +118,51 @@ def test_sensevoice_tensorrt_contract_tracks_merged_native_runtime(valid_registr
assert 'tensorrt version' in limitation
+def test_llama_cpp_contract_tracks_v020_release_assets(valid_registry):
+ entry = next(item for item in valid_registry['deployments'] if item['id'] == 'llama-cpp')
+
+ assert entry['tested'] == {
+ 'funasr': 'runtime-llamacpp-v0.2.0',
+ 'runtime': 'llama.cpp@803b7fca',
+ 'verified': '2026-08-11',
+ }
+ assert len(entry['downloads']) == 9
+ assert {item['archive'] for item in entry['downloads']} == {
+ 'funasr-llamacpp-linux-arm64.tar.gz',
+ 'funasr-llamacpp-linux-x64.tar.gz',
+ 'funasr-llamacpp-linux-x64-avx2.tar.gz',
+ 'funasr-llamacpp-linux-x64-vulkan.tar.gz',
+ 'funasr-llamacpp-macos-arm64.tar.gz',
+ 'funasr-llamacpp-windows-x64.zip',
+ 'funasr-llamacpp-windows-x64-avx2.zip',
+ 'funasr-llamacpp-windows-x64-vulkan.zip',
+ 'funasr-llamacpp-windows-x64-cuda.zip',
+ }
+ assert all(item['url'].startswith(
+ 'https://github.com/modelscope/FunASR/releases/download/runtime-llamacpp-v0.2.0/'
+ ) for item in entry['downloads'])
+ assert all(len(item['sha256']) == 64 for item in entry['downloads'])
+ assert any('actions/runs/31458121788' in item['url'] for item in entry['evidence'])
+ assert any(
+ 'download-funasr-model.sh sensevoice ./funasr-gguf f16' in command
+ for command in entry['commands']['install']
+ )
+ assert 'sensevoice-small-f16.gguf' in entry['commands']['launch'][0]
+ assert 'AMD' in entry['translations']['en']['primary_limitation']
+
+
+def test_download_assets_require_https_and_sha256(valid_registry):
+ data = copy.deepcopy(valid_registry)
+ entry = next(item for item in data['deployments'] if item['id'] == 'llama-cpp')
+ entry['downloads'][0]['url'] = 'http://example.com/runtime.tar.gz'
+ entry['downloads'][1]['sha256'] = 'not-a-sha256'
+
+ errors = validate_registry(data)
+
+ assert 'llama-cpp: download URL must use https (item 0)' in errors
+ assert 'llama-cpp: download SHA-256 must be 64 lowercase hex characters (item 1)' in errors
+
+
def test_production_entry_requires_evidence(valid_registry):
data = copy.deepcopy(valid_registry)
entry = next(item for item in data['deployments'] if item['maturity'] == 'production-verified')