NemotronLabs-VoiceChat-11B outputs empty text + silent audio — zero-weight tensors in checkpoint
#8
by bliunlpr - opened
The NVIDIA-NemotronLabs-VoiceChat-11B checkpoint produces empty text and silent audio for any input. The model outputs 196 consecutive pad tokens (id=12). Inspection of model.safetensors reveals 21 fully-zero tensors, including 4 complete LLM layers (0, 1, 10, 11) and function_head.
python examples/speechlm2/offline_voicechat_infer.py
--checkpoint /path/to/NVIDIA-NemotronLabs-VoiceChat-11B
--audio /path/to/sample.wav
Output: text='', tokens_text = 196× pad (id=12), audio silent.
Notably, the model collapses to pad at step 3 during the prompt region (teacher forcing with correct prompt tokens), indicating the weights themselves are broken — not just a generation feedback loop.
Questions
- Is this a known issue with the published checkpoint, or a download problem? (File opens cleanly, all 1632 keys readable — doesn't look truncated. Is there an official checksum?)
- Were layers 0/1/10/11,
embed_tokens, andfunction_headintentionally zeroed during VoiceChat training, or is this a checkpoint export bug? - Is there a corrected checkpoint, or a recommended way to restore the missing weights from
Nemotron-Nano-9B-v2?
Happy to share the diagnostic scripts. Thanks!