feat: rename & add VAM mini pipeline configs

- configs/vam/vam_t2a_all_mini.yaml (Stage 1: T2A, mode=all)
- configs/vam/vam_a2a_audio_proj_mini.yaml (Stage 2: A2A, mode=audio_proj)
- configs/vam/vam_a2a_all_mini.yaml (Stage 3: A2A, mode=all, low LR)

configs/vam/vam_a2a_all_mini.yaml

@@ -0,0 +1,54 @@
+# VAM SFT 快速验证配置（Stage 3: A2A, mode=all, low LR）
+# 参考 minimind-o mini pipeline: mode=all, from_weight=sft_omni, lr=2e-5, max_seq_len=768
+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: 2.0e-5
+  accumulation_steps: 1
+  grad_clip: 1.0
+  dtype: bfloat16
+  num_workers: 0
+  save_interval: 2000
+  log_interval: 50
+  from_weight: sft_omni
+  model_dir: checkpoint/vam_a2a_audio_proj_mini   # 加载 Stage 2 输出
+  from_resume: 0
+  freeze_backbone: none
+  mode: all
+  max_samples: 2000
+
+paths:
+  save_dir: checkpoint/vam_a2a_all_mini
+  data_path: dataset/vam/sft_a2a_mini.parquet
+  audio_encoder_dir: checkpoint/sensevoice
+  vision_dir: checkpoint/siglip
+  tokenizer_dir: checkpoint/omni/native_hf

configs/vam/vam_a2a_audio_proj_mini.yaml

@@ -0,0 +1,54 @@
+# VAM SFT 快速验证配置（Stage 2: A2A, mode=audio_proj）
+# 参考 minimind-o mini pipeline: mode=audio_proj, from_weight=sft_omni, lr=5e-4, max_seq_len=640
+model:
+  hidden_size: 768
+  num_hidden_layers: 8
+  use_moe: 0
+  vocab_size: 6400
+  max_seq_len: 640
+  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: 8
+  learning_rate: 5.0e-4
+  accumulation_steps: 1
+  grad_clip: 1.0
+  dtype: bfloat16
+  num_workers: 0
+  save_interval: 2000
+  log_interval: 50
+  from_weight: sft_omni
+  model_dir: checkpoint/vam_t2a_all_mini   # 加载 Stage 1 输出
+  from_resume: 0
+  freeze_backbone: none
+  mode: audio_proj                         # 仅训练 audio_proj
+  max_samples: 2000
+
+paths:
+  save_dir: checkpoint/vam_a2a_audio_proj_mini
+  data_path: dataset/vam/sft_a2a_mini.parquet
+  audio_encoder_dir: checkpoint/sensevoice
+  vision_dir: checkpoint/siglip
+  tokenizer_dir: checkpoint/omni/native_hf

configs/vam/{vam_sft_mini.yaml → vam_t2a_all_mini.yaml}

@@ -1,4 +1,4 @@
-# VAM SFT 快速验证配置（Stage 1: T2A）
+# VAM SFT 快速验证配置（Stage 1: T2A, mode=all）
 # 参考 minimind-o mini pipeline: mode=all, from_weight=llm, lr=5e-4, max_seq_len=512
 model:
   hidden_size: 768
@@ -47,7 +47,7 @@ train:
   max_samples: 2000
 
 paths:
-  save_dir: checkpoint/vam_sft_mini
+  save_dir: checkpoint/vam_t2a_all_mini
   data_path: dataset/vam/sft_t2a_mini.parquet
   audio_encoder_dir: checkpoint/sensevoice
   vision_dir: checkpoint/siglip