feat: add VAM T2A config variant using omni-v initialization

configs/vam/vam_t2a_all_mini_omni-v.yaml

@@ -0,0 +1,54 @@
+# VAM SFT 快速验证配置（Stage 1: T2A, mode=all, 从 omni-v 初始化）
+# 从 checkpoint/omni-v/omni-v.pth 加载（已具备视觉能力），继续训练音频
+model:
+  hidden_size: 768
+  num_hidden_layers: 8
+  use_moe: 0
+  vocab_size: 6400
+  max_seq_len: 768
+  num_attention_heads: 8
+  num_key_value_heads: 4
+  dropout: 0.0
+  # Talker（语音生成）相关
+  num_talker_hidden_layers: 4
+  talker_hidden_size: 768
+  # 音频相关
+  audio_special_token: "<|audio_pad|>"
+  audio_ids: [16]
+  audio_hidden_size: 512
+  audio_vocab_size: 2112
+  audio_pad_token: 2049
+  audio_stop_token: 2050
+  audio_spk_token: 2051
+  spk_emb_size: 192
+  think_end_ids: [26, 234, 234]
+  # 视觉相关
+  image_special_token: "<|image_pad|>"
+  image_ids: [12]
+  image_hidden_size: 768
+  image_token_len: 64
+  bridge_layer: 3
+
+train:
+  epochs: 1
+  batch_size: 4
+  learning_rate: 5.0e-4
+  accumulation_steps: 1
+  grad_clip: 1.0
+  dtype: bfloat16
+  num_workers: 0
+  save_interval: 5000
+  log_interval: 50
+  from_weight: omni-v
+  model_dir: checkpoint/omni-v      # 加载 checkpoint/omni-v/omni-v.pth（已有 vision_proj）
+  from_resume: 0
+  freeze_backbone: none
+  mode: all
+  max_samples: 2000
+
+paths:
+  save_dir: checkpoint/vam_t2a_all_mini_omni-v
+  data_path: dataset/vam/sft_t2a_mini.parquet
+  audio_encoder_dir: checkpoint/sensevoice
+  vision_dir: checkpoint/siglip
+  tokenizer_dir: checkpoint/omni/native_hf