Skip to content

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

docs: clarify Fun-ASR-Nano HF CTC boundary

LauraGPT4604a0bc6498
diff --git a/docs/vllm_guide.md b/docs/vllm_guide.md
index 759e37a45..c244e0737 100644
--- a/docs/vllm_guide.md
+++ b/docs/vllm_guide.md
@@ -59,9 +59,13 @@ APIs are not interchangeable:
Use the official `FunAudioLLM/Fun-ASR-Nano-2512` checkpoint from
[ModelScope](https://modelscope.cn/models/FunAudioLLM/Fun-ASR-Nano-2512) or
[Hugging Face](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512). The
-official Hugging Face repository does contain the complete `model.pt` at its
-root. The `Qwen3-0.6B/` subdirectory intentionally contains only the LLM config
-and tokenizer; it is not a standalone model download.
+current Hugging Face `model.pt` has incomplete CTC weights, so it remains
+usable for transcription but FunASR disables the affected CTC path rather than
+returning unreliable timestamps or speaker diarization. Use the ModelScope
+checkpoint when your deployment requires timestamps or speaker diarization;
+the publication repair is tracked in [#3496](https://github.com/modelscope/FunASR/issues/3496).
+The `Qwen3-0.6B/` subdirectory intentionally contains only the LLM config and
+tokenizer; it is not a standalone model download.
```python
from funasr.auto.auto_model_vllm import AutoModelVLLM
diff --git a/docs/vllm_guide_zh.md b/docs/vllm_guide_zh.md
index 192870565..cfc0e4cab 100644
--- a/docs/vllm_guide_zh.md
+++ b/docs/vllm_guide_zh.md
@@ -59,9 +59,11 @@ cd /path/to/FunASR && pip install -e .
请从 [ModelScope](https://modelscope.cn/models/FunAudioLLM/Fun-ASR-Nano-2512)
或 [Hugging Face](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512)
-下载官方 `FunAudioLLM/Fun-ASR-Nano-2512`。官方 Hugging Face 仓库的根目录
-实际包含完整的 `model.pt`;`Qwen3-0.6B/` 子目录按设计只保存 LLM 配置与
-tokenizer,不是可以单独加载的权重目录。
+下载官方 `FunAudioLLM/Fun-ASR-Nano-2512`。当前 Hugging Face 的 `model.pt`
+CTC 权重不完整,因此仍可用于转写,但 FunASR 会禁用受影响的 CTC 路径,避免
+返回不可靠的时间戳或说话人分离结果。部署需要时间戳或说话人分离时,请使用
+ModelScope checkpoint;权重发布修复进度见 [#3496](https://github.com/modelscope/FunASR/issues/3496)。
+`Qwen3-0.6B/` 子目录按设计只保存 LLM 配置与 tokenizer,不是可以单独加载的权重目录。
```python
from funasr.auto.auto_model_vllm import AutoModelVLLM
diff --git a/docs/vllm_guide_zh_v2.md b/docs/vllm_guide_zh_v2.md
index f05646b87..3bc35c304 100644
--- a/docs/vllm_guide_zh_v2.md
+++ b/docs/vllm_guide_zh_v2.md
@@ -59,9 +59,11 @@ cd /path/to/FunASR && pip install -e .
请从 [ModelScope](https://modelscope.cn/models/FunAudioLLM/Fun-ASR-Nano-2512)
或 [Hugging Face](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512)
-下载官方 `FunAudioLLM/Fun-ASR-Nano-2512`。官方 Hugging Face 仓库的根目录
-实际包含完整的 `model.pt`;`Qwen3-0.6B/` 子目录按设计只保存 LLM 配置与
-tokenizer,不是可以单独加载的权重目录。
+下载官方 `FunAudioLLM/Fun-ASR-Nano-2512`。当前 Hugging Face 的 `model.pt`
+CTC 权重不完整,因此仍可用于转写,但 FunASR 会禁用受影响的 CTC 路径,避免
+返回不可靠的时间戳或说话人分离结果。部署需要时间戳或说话人分离时,请使用
+ModelScope checkpoint;权重发布修复进度见 [#3496](https://github.com/modelscope/FunASR/issues/3496)。
+`Qwen3-0.6B/` 子目录按设计只保存 LLM 配置与 tokenizer,不是可以单独加载的权重目录。
```python
from funasr.auto.auto_model_vllm import AutoModelVLLM
diff --git a/tests/test_vllm_model_source_docs.py b/tests/test_vllm_model_source_docs.py
index e77edd1de..273ec773c 100644
--- a/tests/test_vllm_model_source_docs.py
+++ b/tests/test_vllm_model_source_docs.py
@@ -29,3 +29,48 @@ def test_vllm_guides_distinguish_official_and_native_model_paths(relpath):
for marker in required_markers:
assert marker in text, f"{relpath} is missing {marker}"
+
+
+ ("relpath", "required_markers", "stale_claim"),
+ [
+ (
+ "docs/vllm_guide.md",
+ (
+ "https://github.com/modelscope/FunASR/issues/3496",
+ "incomplete CTC weights",
+ "timestamps or speaker diarization",
+ "Use the ModelScope\ncheckpoint",
+ ),
+ "does contain the complete `model.pt` at its root",
+ ),
+ (
+ "docs/vllm_guide_zh.md",
+ (
+ "https://github.com/modelscope/FunASR/issues/3496",
+ "CTC 权重不完整",
+ "时间戳或说话人分离",
+ "ModelScope checkpoint",
+ ),
+ "实际包含完整的 `model.pt`",
+ ),
+ (
+ "docs/vllm_guide_zh_v2.md",
+ (
+ "https://github.com/modelscope/FunASR/issues/3496",
+ "CTC 权重不完整",
+ "时间戳或说话人分离",
+ "ModelScope checkpoint",
+ ),
+ "实际包含完整的 `model.pt`",
+ ),
+ ],
+)
+def test_vllm_guides_document_hf_incomplete_ctc_checkpoint(
+ relpath, required_markers, stale_claim
+):
+ text = (ROOT / relpath).read_text(encoding="utf-8")
+
+ for marker in required_markers:
+ assert marker in text, f"{relpath} is missing {marker}"
+ assert stale_claim not in text, f"{relpath} still claims HF has complete CTC"