fix VAM: funasr+pyarrow deadlock, librosa+numpy compat, reduce iter_batches batch_size

Root cause: funasr.AutoModel (loaded by SenseVoice encoder) and pyarrow
have a C library initialization conflict causing iter_batches to hang.

Fix:
- lazy-import pyarrow in VAMDataset/VLMDataset.__init__ (after model loads)
- remove unused pyarrow imports from dataset/common.py
- replace librosa.resample with torchaudio.resample (avoids numba+numpy 2.5
incompatibility)
- reduce iter_batches batch_size from 65536 to 4096 (lower memory)

Config updates:
- point Stage 2 (audio_proj) to omni-v Stage 1 checkpoint
- add explicit device: cuda to omni-v config

configs/vam/vam_a2a_audio_proj_mini.yaml

@@ -40,7 +40,7 @@ train:
   save_interval: 2000
   log_interval: 50
   from_weight: sft_omni
-  model_dir: checkpoint/vam_t2a_all_mini   # 加载 Stage 1 输出
+  model_dir: checkpoint/vam_t2a_all_mini_omni-v   # 加载 Stage 1 输出（omni-v variant）
   from_resume: 0
   freeze_backbone: none
   mode: audio_proj                         # 仅训练 audio_proj

configs/vam/vam_t2a_all_mini_omni-v.yaml

@@ -36,6 +36,7 @@ train:
   accumulation_steps: 1
   grad_clip: 1.0
   dtype: bfloat16
+  device: cuda
   num_workers: 0
   save_interval: 5000
   log_interval: 50

src/dataset/common.py

@@ -7,8 +7,6 @@ import random
 from datasets import load_dataset, Features, Sequence, Value
 from PIL import Image
 from datasets import Dataset as HFDataset
-import pyarrow as pa
-import pyarrow.parquet as pq
 
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
 

src/dataset/vam.py

@@ -4,8 +4,6 @@ import random
 import torch
 from torch.utils.data import Dataset
 from PIL import Image
-import pyarrow as pa
-import pyarrow.parquet as pq
 
 from dataset.common import pre_processing_chat, post_processing_chat
 
@@ -20,11 +18,13 @@ class VAMDataset(Dataset):
                  scheduled_sampling=0.05,
                  image_token_len=64, max_samples=None):
         super().__init__()
+        import pyarrow as pa
+        import pyarrow.parquet as pq
         tables = []
         total = 0
         for p in data_path.split(','):
             pf = pq.ParquetFile(p.strip())
-            for batch in pf.iter_batches(batch_size=65536):
+            for batch in pf.iter_batches(batch_size=4096):
                 tables.append(batch)
                 total += batch.num_rows
                 if max_samples is not None and total >= max_samples:
@@ -119,12 +119,17 @@ class VAMDataset(Dataset):
         import soundfile as sf
         import numpy as np
         import io
-        import librosa
+        import torch
         if not audio_bytes: return None, 0
         wav, sr = sf.read(io.BytesIO(audio_bytes))
         if wav.ndim > 1: wav = wav.mean(axis=1)
-        if sr != 16000: wav = librosa.resample(wav.astype(float), orig_sr=sr, target_sr=16000)
-        wav = self.augment_wav(wav.astype(np.float32))
+        wav = wav.astype(np.float32)
+        if sr != 16000:
+            import torchaudio.functional as AF
+            wav_t = torch.from_numpy(wav).unsqueeze(0)
+            wav_t = AF.resample(wav_t, sr, 16000)
+            wav = wav_t.squeeze(0).numpy()
+        wav = self.augment_wav(wav)
         inputs = self.audio_processor(wav, sampling_rate=16000, return_tensors="pt", return_attention_mask=True)
         valid_len = inputs.attention_mask.sum().item()
         return self.augment_mel(inputs.input_features.squeeze(0)), valid_len

src/dataset/vlm.py

@@ -1,6 +1,5 @@
 import io
 import json
-import pyarrow.parquet as pq
 import torch
 from torch.utils.data import Dataset
 from PIL import Image
@@ -12,6 +11,7 @@ from dataset.common import VLM, pre_processing_chat, post_processing_chat
 class VLMDataset(Dataset):
     def __init__(self, parquet_path, tokenizer, preprocess=None, max_length=512, image_special_token='<|image_pad|>', image_token_len=64, max_samples=None):
         super().__init__()
+        import pyarrow.parquet as pq
         pf = pq.ParquetFile(parquet_path)
         total = pf.metadata.num_rows
         if max_samples is not None and max_samples < total: