fix: disable timestamps when CTC weights are missing (#3211)

Fail closed for native, remote-code, and vLLM Fun-ASR-Nano timestamp inference when checkpoint CTC weights are missing, partial, shape-mismatched, DDP-prefixed, or wrapped.

Fixes #3208.

examples/industrial_data_pretraining/fun_asr_nano/README.md

@@ -64,6 +64,12 @@ pip install -r requirements.txt
 - [ ] Support speaker diarization
 - [x] Support model training
 
+> [!NOTE]
+> Character-level timestamps require a checkpoint with complete `ctc_decoder.*` and
+> `ctc.*` weights. The current `Fun-ASR-MLT-Nano-2512` checkpoint does not include
+> those weights, so FunASR logs a warning and returns text without `timestamps` or
+> `ctc_timestamps` instead of aligning with uninitialized layers.
+
 # Usage 🛠️
 
 ## Inference

examples/industrial_data_pretraining/fun_asr_nano/README_zh.md

@@ -64,6 +64,11 @@ pip install -r requirements.txt
 - [ ] 支持区分说话人识别
 - [x] 支持模型训练
 
+> [!NOTE]
+> 字符级时间戳要求 checkpoint 同时包含完整的 `ctc_decoder.*` 和 `ctc.*` 权重。
+> 当前 `Fun-ASR-MLT-Nano-2512` checkpoint 未包含这些权重，因此 FunASR 会记录警告并仅返回文本，
+> 不再使用未初始化的层生成 `timestamps` 或 `ctc_timestamps`。
+
 # 用法 🛠️
 
 ## 推理

examples/industrial_data_pretraining/fun_asr_nano/model.py

@@ -15,6 +15,10 @@ from funasr.register import tables
 from funasr.train_utils.device_funcs import force_gatherable, to_device
 from funasr.utils.datadir_writer import DatadirWriter
 from funasr.utils.load_utils import extract_fbank, load_audio_text_image_video
+from funasr.models.fun_asr_nano.checkpoint_utils import (
+    disable_incomplete_ctc,
+    normalize_checkpoint_state,
+)
 from funasr.models.fun_asr_nano.device_utils import resolve_autocast_device_type
 
 
@@ -109,6 +113,7 @@ class FunASRNano(nn.Module):
 
         # ctc decoder
         self.ctc_decoder = None
+        self._externally_loaded_ctc_keys = set()
         # TODO: fix table name
         ctc_decoder_class = tables.adaptor_classes.get(kwargs.get("ctc_decoder", None))
         if ctc_decoder_class is not None:
@@ -135,8 +140,15 @@ class FunASRNano(nn.Module):
             self.ctc_decoder = ctc_decoder_class(**ctc_decoder_conf)
             init_param_path = ctc_decoder_conf.get("init_param_path", None)
             if init_param_path is not None:
-                src_state = torch.load(init_param_path, map_location="cpu")
+                src_state = normalize_checkpoint_state(
+                    torch.load(init_param_path, map_location="cpu")
+                )
                 flag = self.ctc_decoder.load_state_dict(src_state, strict=False)
+                self._externally_loaded_ctc_keys.update(
+                    f"ctc_decoder.{key}"
+                    for key in self.ctc_decoder.state_dict()
+                    if key in src_state
+                )
                 logging.info(f"Loading ctc_decoder ckpt: {init_param_path}, status: {flag}")
             freeze = ctc_decoder_conf.get("freeze", False)
             if freeze:
@@ -160,6 +172,11 @@ class FunASRNano(nn.Module):
         rank = int(os.environ.get("RANK", 0))
         logging.info(f"rank: {rank}, model is builded.")
 
+    def on_pretrained_model_loaded(self, loaded_keys):
+        """Fail closed when a checkpoint configures CTC without trained weights."""
+        loaded_keys = set(loaded_keys).union(getattr(self, "_externally_loaded_ctc_keys", ()))
+        disable_incomplete_ctc(self, loaded_keys, log=logging)
+
     def forward(
         self,
         speech: torch.Tensor = None,