Fun-ASR-Nano Hugging Face checkpoint is missing 86 CTC tensors required for timestamps and diarization

[oliver-mee]

Summary

The current Hugging Face and ModelScope copies of FunAudioLLM/Fun-ASR-Nano-2512 do not contain the same model.pt.

The Hugging Face file is smaller and lacks every ctc_decoder.* and ctc.* tensor. ASR text still works, but the CTC timestamps required by FunASR's native speaker diarization path are unavailable.

Verified on 2026-08-14 with FunASR 1.4.2.

Artefact comparison

source	model.pt size	state-dict tensors	CTC tensors
Hugging Face, commit 272c57b82523ada6fd87095e955f8e29100979ab	1,971,149,431 bytes	1,261	0
ModelScope	2,127,426,538 bytes	1,347	84 ctc_decoder.* + 2 ctc.*

Complete ModelScope checkpoint SHA-256:

81fec8616083c69377f3ceef36aba3655660ee0ca69a5d4a1e9810cd340ca499

Complete namespace counts:

audio_encoder  914
audio_adaptor   36
llm            311
ctc_decoder     84
ctc              2

Minimal inspection

from collections import Counter
from huggingface_hub import hf_hub_download
import os
import torch

path = hf_hub_download("FunAudioLLM/Fun-ASR-Nano-2512", "model.pt")
state = torch.load(path, map_location="cpu", mmap=True, weights_only=False)

print(os.path.getsize(path))
print(len(state))
print(Counter(key.split(".", 1)[0] for key in state))

Effect

Using the Hugging Face checkpoint through AutoModel produces text, but no Nano CTC timestamps. That in turn prevents the native spk_model="cam++" diarization path from producing valid speaker-attributed segments.

Replacing only model.pt with the complete ModelScope checkpoint restores character timestamps and native diarization. With the complete checkpoint, AutoModel works on both CPU and Apple MPS.

Expected result

Please synchronise the complete ModelScope model.pt to the Hugging Face repository, or otherwise document why the published artefacts intentionally differ.

It may also be useful for the loader to fail closed when a Nano checkpoint has no CTC tensors, rather than allowing text-only inference to make the checkpoint appear complete.

Related: #3208 and #3211.

[LauraGPT]

Independent verification confirms this report.

Artifact evidence

source	exact object	bytes	state-dict entries	CTC tensors
Hugging Face at 272c57b82523ada6fd87095e955f8e29100979ab	55ae0d2fee369f0f11cce0795f6927934ad17cf11b278a7e56a51272074160bb	1,971,149,431	1,261	0
ModelScope complete artifact	81fec8616083c69377f3ceef36aba3655660ee0ca69a5d4a1e9810cd340ca499	2,127,426,538	1,347	84 ctc_decoder.* + 2 ctc.*

I downloaded the ModelScope file independently with byte ranges, reassembled it to exactly 2,127,426,538 bytes, and matched the server's X-Linked-Etag. The complete namespace counts are 914 audio_encoder, 36 audio_adaptor, 311 llm, 84 ctc_decoder, and 2 ctc.

Public-package runtime proof

Using the public PyPI wheel funasr==1.4.2 on CPU:

• the complete checkpoint loaded with all keys matched;
• the official Chinese sample returned 开饭时间：早上九点至下午五点。;
• CTC output contained 15 character timestamps;
• fsmn-vad + cam++ produced sentence_info with spk: 0.

The existing loader protection from #3211 already disables incomplete CTC modules rather than running random timestamp weights, so no additional code fix is required for fail-closed behavior. The remaining fix is the model artifact sync.

Publication state

A complete atomic HF update was prepared: replace model.pt and add a bilingual model-card integrity contract with the exact size, SHA-256, and tensor counts. No HF content was changed: the current fine-grained token has an empty model-specific permission override, and backup-branch creation, direct LFS upload, create_pr=True preupload, and discussion creation all returned HTTP 403 before mutation.

This issue should remain open as a model-publication blocker. A FunAudioLLM model owner needs to upload the complete ModelScope model.pt. After that, I will re-download the HF object, verify SHA-256/state dict, rerun the public 1.4.2 CPU timestamp and diarization tests, and close the issue.