Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

am.mvn

@@ -0,0 +1,8 @@
+<Nnet> 
+<Splice> 560 560
+[ 0 ]
+<AddShift> 560 560 
+<LearnRateCoef> 0 [ -8.311879 -8.600912 -9.615928 -10.43595 -11.21292 -11.88333 -12.36243 -12.63706 -12.8818 -12.83066 -12.89103 -12.95666 -13.19763 -13.40598 -13.49113 -13.5546 -13.55639 -13.51915 -13.68284 -13.53289 -13.42107 -13.65519 -13.50713 -13.75251 -13.76715 -13.87408 -13.73109 -13.70412 -13.56073 -13.53488 -13.54895 -13.56228 -13.59408 -13.62047 -13.64198 -13.66109 -13.62669 -13.58297 -13.57387 -13.4739 -13.53063 -13.48348 -13.61047 -13.64716 -13.71546 -13.79184 -13.90614 -14.03098 -14.18205 -14.35881 -14.48419 -14.60172 -14.70591 -14.83362 -14.92122 -15.00622 -15.05122 -15.03119 -14.99028 -14.92302 -14.86927 -14.82691 -14.7972 -14.76909 -14.71356 -14.61277 -14.51696 -14.42252 -14.36405 -14.30451 -14.23161 -14.19851 -14.16633 -14.15649 -14.10504 -13.99518 -13.79562 -13.3996 -12.7767 -11.71208 -8.311879 -8.600912 -9.615928 -10.43595 -11.21292 -11.88333 -12.36243 -12.63706 -12.8818 -12.83066 -12.89103 -12.95666 -13.19763 -13.40598 -13.49113 -13.5546 -13.55639 -13.51915 -13.68284 -13.53289 -13.42107 -13.65519 -13.50713 -13.75251 -13.76715 -13.87408 -13.73109 -13.70412 -13.56073 -13.53488 -13.54895 -13.56228 -13.59408 -13.62047 -13.64198 -13.66109 -13.62669 -13.58297 -13.57387 -13.4739 -13.53063 -13.48348 -13.61047 -13.64716 -13.71546 -13.79184 -13.90614 -14.03098 -14.18205 -14.35881 -14.48419 -14.60172 -14.70591 -14.83362 -14.92122 -15.00622 -15.05122 -15.03119 -14.99028 -14.92302 -14.86927 -14.82691 -14.7972 -14.76909 -14.71356 -14.61277 -14.51696 -14.42252 -14.36405 -14.30451 -14.23161 -14.19851 -14.16633 -14.15649 -14.10504 -13.99518 -13.79562 -13.3996 -12.7767 -11.71208 -8.311879 -8.600912 -9.615928 -10.43595 -11.21292 -11.88333 -12.36243 -12.63706 -12.8818 -12.83066 -12.89103 -12.95666 -13.19763 -13.40598 -13.49113 -13.5546 -13.55639 -13.51915 -13.68284 -13.53289 -13.42107 -13.65519 -13.50713 -13.75251 -13.76715 -13.87408 -13.73109 -13.70412 -13.56073 -13.53488 -13.54895 -13.56228 -13.59408 -13.62047 -13.64198 -13.66109 -13.62669 -13.58297 -13.57387 -13.4739 -13.53063 -13.48348 -13.61047 -13.64716 -13.71546 -13.79184 -13.90614 -14.03098 -14.18205 -14.35881 -14.48419 -14.60172 -14.70591 -14.83362 -14.92122 -15.00622 -15.05122 -15.03119 -14.99028 -14.92302 -14.86927 -14.82691 -14.7972 -14.76909 -14.71356 -14.61277 -14.51696 -14.42252 -14.36405 -14.30451 -14.23161 -14.19851 -14.16633 -14.15649 -14.10504 -13.99518 -13.79562 -13.3996 -12.7767 -11.71208 -8.311879 -8.600912 -9.615928 -10.43595 -11.21292 -11.88333 -12.36243 -12.63706 -12.8818 -12.83066 -12.89103 -12.95666 -13.19763 -13.40598 -13.49113 -13.5546 -13.55639 -13.51915 -13.68284 -13.53289 -13.42107 -13.65519 -13.50713 -13.75251 -13.76715 -13.87408 -13.73109 -13.70412 -13.56073 -13.53488 -13.54895 -13.56228 -13.59408 -13.62047 -13.64198 -13.66109 -13.62669 -13.58297 -13.57387 -13.4739 -13.53063 -13.48348 -13.61047 -13.64716 -13.71546 -13.79184 -13.90614 -14.03098 -14.18205 -14.35881 -14.48419 -14.60172 -14.70591 -14.83362 -14.92122 -15.00622 -15.05122 -15.03119 -14.99028 -14.92302 -14.86927 -14.82691 -14.7972 -14.76909 -14.71356 -14.61277 -14.51696 -14.42252 -14.36405 -14.30451 -14.23161 -14.19851 -14.16633 -14.15649 -14.10504 -13.99518 -13.79562 -13.3996 -12.7767 -11.71208 -8.311879 -8.600912 -9.615928 -10.43595 -11.21292 -11.88333 -12.36243 -12.63706 -12.8818 -12.83066 -12.89103 -12.95666 -13.19763 -13.40598 -13.49113 -13.5546 -13.55639 -13.51915 -13.68284 -13.53289 -13.42107 -13.65519 -13.50713 -13.75251 -13.76715 -13.87408 -13.73109 -13.70412 -13.56073 -13.53488 -13.54895 -13.56228 -13.59408 -13.62047 -13.64198 -13.66109 -13.62669 -13.58297 -13.57387 -13.4739 -13.53063 -13.48348 -13.61047 -13.64716 -13.71546 -13.79184 -13.90614 -14.03098 -14.18205 -14.35881 -14.48419 -14.60172 -14.70591 -14.83362 -14.92122 -15.00622 -15.05122 -15.03119 -14.99028 -14.92302 -14.86927 -14.82691 -14.7972 -14.76909 -14.71356 -14.61277 -14.51696 -14.42252 -14.36405 -14.30451 -14.23161 -14.19851 -14.16633 -14.15649 -14.10504 -13.99518 -13.79562 -13.3996 -12.7767 -11.71208 -8.311879 -8.600912 -9.615928 -10.43595 -11.21292 -11.88333 -12.36243 -12.63706 -12.8818 -12.83066 -12.89103 -12.95666 -13.19763 -13.40598 -13.49113 -13.5546 -13.55639 -13.51915 -13.68284 -13.53289 -13.42107 -13.65519 -13.50713 -13.75251 -13.76715 -13.87408 -13.73109 -13.70412 -13.56073 -13.53488 -13.54895 -13.56228 -13.59408 -13.62047 -13.64198 -13.66109 -13.62669 -13.58297 -13.57387 -13.4739 -13.53063 -13.48348 -13.61047 -13.64716 -13.71546 -13.79184 -13.90614 -14.03098 -14.18205 -14.35881 -14.48419 -14.60172 -14.70591 -14.83362 -14.92122 -15.00622 -15.05122 -15.03119 -14.99028 -14.92302 -14.86927 -14.82691 -14.7972 -14.76909 -14.71356 -14.61277 -14.51696 -14.42252 -14.36405 -14.30451 -14.23161 -14.19851 -14.16633 -14.15649 -14.10504 -13.99518 -13.79562 -13.3996 -12.7767 -11.71208 -8.311879 -8.600912 -9.615928 -10.43595 -11.21292 -11.88333 -12.36243 -12.63706 -12.8818 -12.83066 -12.89103 -12.95666 -13.19763 -13.40598 -13.49113 -13.5546 -13.55639 -13.51915 -13.68284 -13.53289 -13.42107 -13.65519 -13.50713 -13.75251 -13.76715 -13.87408 -13.73109 -13.70412 -13.56073 -13.53488 -13.54895 -13.56228 -13.59408 -13.62047 -13.64198 -13.66109 -13.62669 -13.58297 -13.57387 -13.4739 -13.53063 -13.48348 -13.61047 -13.64716 -13.71546 -13.79184 -13.90614 -14.03098 -14.18205 -14.35881 -14.48419 -14.60172 -14.70591 -14.83362 -14.92122 -15.00622 -15.05122 -15.03119 -14.99028 -14.92302 -14.86927 -14.82691 -14.7972 -14.76909 -14.71356 -14.61277 -14.51696 -14.42252 -14.36405 -14.30451 -14.23161 -14.19851 -14.16633 -14.15649 -14.10504 -13.99518 -13.79562 -13.3996 -12.7767 -11.71208 ]
+<Rescale> 560 560
+<LearnRateCoef> 0 [ 0.155775 0.154484 0.1527379 0.1518718 0.1506028 0.1489256 0.147067 0.1447061 0.1436307 0.1443568 0.1451849 0.1455157 0.1452821 0.1445717 0.1439195 0.1435867 0.1436018 0.1438781 0.1442086 0.1448844 0.1454756 0.145663 0.146268 0.1467386 0.1472724 0.147664 0.1480913 0.1483739 0.1488841 0.1493636 0.1497088 0.1500379 0.1502916 0.1505389 0.1506787 0.1507102 0.1505992 0.1505445 0.1505938 0.1508133 0.1509569 0.1512396 0.1514625 0.1516195 0.1516156 0.1515561 0.1514966 0.1513976 0.1512612 0.151076 0.1510596 0.1510431 0.151077 0.1511168 0.1511917 0.151023 0.1508045 0.1505885 0.1503493 0.1502373 0.1501726 0.1500762 0.1500065 0.1499782 0.150057 0.1502658 0.150469 0.1505335 0.1505505 0.1505328 0.1504275 0.1502438 0.1499674 0.1497118 0.1494661 0.1493102 0.1493681 0.1495501 0.1499738 0.1509654 0.155775 0.154484 0.1527379 0.1518718 0.1506028 0.1489256 0.147067 0.1447061 0.1436307 0.1443568 0.1451849 0.1455157 0.1452821 0.1445717 0.1439195 0.1435867 0.1436018 0.1438781 0.1442086 0.1448844 0.1454756 0.145663 0.146268 0.1467386 0.1472724 0.147664 0.1480913 0.1483739 0.1488841 0.1493636 0.1497088 0.1500379 0.1502916 0.1505389 0.1506787 0.1507102 0.1505992 0.1505445 0.1505938 0.1508133 0.1509569 0.1512396 0.1514625 0.1516195 0.1516156 0.1515561 0.1514966 0.1513976 0.1512612 0.151076 0.1510596 0.1510431 0.151077 0.1511168 0.1511917 0.151023 0.1508045 0.1505885 0.1503493 0.1502373 0.1501726 0.1500762 0.1500065 0.1499782 0.150057 0.1502658 0.150469 0.1505335 0.1505505 0.1505328 0.1504275 0.1502438 0.1499674 0.1497118 0.1494661 0.1493102 0.1493681 0.1495501 0.1499738 0.1509654 0.155775 0.154484 0.1527379 0.1518718 0.1506028 0.1489256 0.147067 0.1447061 0.1436307 0.1443568 0.1451849 0.1455157 0.1452821 0.1445717 0.1439195 0.1435867 0.1436018 0.1438781 0.1442086 0.1448844 0.1454756 0.145663 0.146268 0.1467386 0.1472724 0.147664 0.1480913 0.1483739 0.1488841 0.1493636 0.1497088 0.1500379 0.1502916 0.1505389 0.1506787 0.1507102 0.1505992 0.1505445 0.1505938 0.1508133 0.1509569 0.1512396 0.1514625 0.1516195 0.1516156 0.1515561 0.1514966 0.1513976 0.1512612 0.151076 0.1510596 0.1510431 0.151077 0.1511168 0.1511917 0.151023 0.1508045 0.1505885 0.1503493 0.1502373 0.1501726 0.1500762 0.1500065 0.1499782 0.150057 0.1502658 0.150469 0.1505335 0.1505505 0.1505328 0.1504275 0.1502438 0.1499674 0.1497118 0.1494661 0.1493102 0.1493681 0.1495501 0.1499738 0.1509654 0.155775 0.154484 0.1527379 0.1518718 0.1506028 0.1489256 0.147067 0.1447061 0.1436307 0.1443568 0.1451849 0.1455157 0.1452821 0.1445717 0.1439195 0.1435867 0.1436018 0.1438781 0.1442086 0.1448844 0.1454756 0.145663 0.146268 0.1467386 0.1472724 0.147664 0.1480913 0.1483739 0.1488841 0.1493636 0.1497088 0.1500379 0.1502916 0.1505389 0.1506787 0.1507102 0.1505992 0.1505445 0.1505938 0.1508133 0.1509569 0.1512396 0.1514625 0.1516195 0.1516156 0.1515561 0.1514966 0.1513976 0.1512612 0.151076 0.1510596 0.1510431 0.151077 0.1511168 0.1511917 0.151023 0.1508045 0.1505885 0.1503493 0.1502373 0.1501726 0.1500762 0.1500065 0.1499782 0.150057 0.1502658 0.150469 0.1505335 0.1505505 0.1505328 0.1504275 0.1502438 0.1499674 0.1497118 0.1494661 0.1493102 0.1493681 0.1495501 0.1499738 0.1509654 0.155775 0.154484 0.1527379 0.1518718 0.1506028 0.1489256 0.147067 0.1447061 0.1436307 0.1443568 0.1451849 0.1455157 0.1452821 0.1445717 0.1439195 0.1435867 0.1436018 0.1438781 0.1442086 0.1448844 0.1454756 0.145663 0.146268 0.1467386 0.1472724 0.147664 0.1480913 0.1483739 0.1488841 0.1493636 0.1497088 0.1500379 0.1502916 0.1505389 0.1506787 0.1507102 0.1505992 0.1505445 0.1505938 0.1508133 0.1509569 0.1512396 0.1514625 0.1516195 0.1516156 0.1515561 0.1514966 0.1513976 0.1512612 0.151076 0.1510596 0.1510431 0.151077 0.1511168 0.1511917 0.151023 0.1508045 0.1505885 0.1503493 0.1502373 0.1501726 0.1500762 0.1500065 0.1499782 0.150057 0.1502658 0.150469 0.1505335 0.1505505 0.1505328 0.1504275 0.1502438 0.1499674 0.1497118 0.1494661 0.1493102 0.1493681 0.1495501 0.1499738 0.1509654 0.155775 0.154484 0.1527379 0.1518718 0.1506028 0.1489256 0.147067 0.1447061 0.1436307 0.1443568 0.1451849 0.1455157 0.1452821 0.1445717 0.1439195 0.1435867 0.1436018 0.1438781 0.1442086 0.1448844 0.1454756 0.145663 0.146268 0.1467386 0.1472724 0.147664 0.1480913 0.1483739 0.1488841 0.1493636 0.1497088 0.1500379 0.1502916 0.1505389 0.1506787 0.1507102 0.1505992 0.1505445 0.1505938 0.1508133 0.1509569 0.1512396 0.1514625 0.1516195 0.1516156 0.1515561 0.1514966 0.1513976 0.1512612 0.151076 0.1510596 0.1510431 0.151077 0.1511168 0.1511917 0.151023 0.1508045 0.1505885 0.1503493 0.1502373 0.1501726 0.1500762 0.1500065 0.1499782 0.150057 0.1502658 0.150469 0.1505335 0.1505505 0.1505328 0.1504275 0.1502438 0.1499674 0.1497118 0.1494661 0.1493102 0.1493681 0.1495501 0.1499738 0.1509654 0.155775 0.154484 0.1527379 0.1518718 0.1506028 0.1489256 0.147067 0.1447061 0.1436307 0.1443568 0.1451849 0.1455157 0.1452821 0.1445717 0.1439195 0.1435867 0.1436018 0.1438781 0.1442086 0.1448844 0.1454756 0.145663 0.146268 0.1467386 0.1472724 0.147664 0.1480913 0.1483739 0.1488841 0.1493636 0.1497088 0.1500379 0.1502916 0.1505389 0.1506787 0.1507102 0.1505992 0.1505445 0.1505938 0.1508133 0.1509569 0.1512396 0.1514625 0.1516195 0.1516156 0.1515561 0.1514966 0.1513976 0.1512612 0.151076 0.1510596 0.1510431 0.151077 0.1511168 0.1511917 0.151023 0.1508045 0.1505885 0.1503493 0.1502373 0.1501726 0.1500762 0.1500065 0.1499782 0.150057 0.1502658 0.150469 0.1505335 0.1505505 0.1505328 0.1504275 0.1502438 0.1499674 0.1497118 0.1494661 0.1493102 0.1493681 0.1495501 0.1499738 0.1509654 ]
+</Nnet> 

chn_jpn_yue_eng_ko_spectok.bpe.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa87f86064c3730d799ddf7af3c04659151102cba548bce325cf06ba4da4e6a8
+size 377341

config.json

@@ -0,0 +1,35 @@
+{
+  "_name_or_path": "/home/chenyifu/new_rl/dissrc/vita_audio_rl/VITA-Audio/models/VITA-Audio-Plus-Vanilla",
+  "architectures": [
+    "Qwen2MTPSenseVoiceForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_qwen2.Qwen2MTPSenseVoiceConfig",
+    "AutoModelForCausalLM": "modeling_qwen2.Qwen2MTPSenseVoiceForCausalLM"
+  },
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 3584,
+  "initializer_range": 0.02,
+  "intermediate_size": 18944,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 28,
+  "model_type": "qwen2_mtp_sensevoice",
+  "mtp_loss_weight": 1.0,
+  "num_attention_heads": 28,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 4,
+  "num_nextn_predict_layers": 0,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.48.3",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 168072
+}

config.yaml

@@ -0,0 +1,98 @@
+encoder: SenseVoiceEncoderSmall
+encoder_conf:
+    output_size: 512
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 50
+    tp_blocks: 20
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.1
+    input_layer: pe
+    pos_enc_class: SinusoidalPositionEncoder
+    normalize_before: true
+    kernel_size: 11
+    sanm_shfit: 0
+    selfattention_layer_type: sanm
+
+
+model: SenseVoiceSmall
+model_conf:
+    length_normalized_loss: true
+    sos: 1
+    eos: 2
+    ignore_id: -1
+
+tokenizer: SentencepiecesTokenizer
+tokenizer_conf:
+  bpemodel: null
+  unk_symbol: <unk>
+  split_with_space: true
+
+frontend: WavFrontend
+frontend_conf:
+    fs: 16000
+    window: hamming
+    n_mels: 80
+    frame_length: 25
+    frame_shift: 10
+    lfr_m: 7
+    lfr_n: 6
+    cmvn_file: null
+
+
+dataset: SenseVoiceCTCDataset
+dataset_conf:
+  index_ds: IndexDSJsonl
+  batch_sampler: EspnetStyleBatchSampler
+  data_split_num: 32
+  batch_type: token
+  batch_size: 14000
+  max_token_length: 2000
+  min_token_length: 60
+  max_source_length: 2000
+  min_source_length: 60
+  max_target_length: 200
+  min_target_length: 0
+  shuffle: true
+  num_workers: 4
+  sos: ${model_conf.sos}
+  eos: ${model_conf.eos}
+  IndexDSJsonl: IndexDSJsonl
+  retry: 20
+
+train_conf:
+  accum_grad: 1
+  grad_clip: 5
+  max_epoch: 20
+  keep_nbest_models: 10
+  avg_nbest_model: 10
+  log_interval: 100
+  resume: true
+  validate_interval: 10000
+  save_checkpoint_interval: 10000
+
+optim: adamw
+optim_conf:
+  lr: 0.00002
+scheduler: warmuplr
+scheduler_conf:
+  warmup_steps: 25000
+
+specaug: SpecAugLFR
+specaug_conf:
+    apply_time_warp: false
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 30
+    lfr_rate: 6
+    num_freq_mask: 1
+    apply_time_mask: true
+    time_mask_width_range:
+    - 0
+    - 12
+    num_time_mask: 1
+

configuration.json

@@ -0,0 +1,14 @@
+{
+  "framework": "pytorch",
+  "task" : "auto-speech-recognition",
+  "model": {"type" : "funasr"},
+  "pipeline": {"type":"funasr-pipeline"},
+  "model_name_in_hub": {
+    "ms":"", 
+    "hf":""},
+  "file_path_metas": {
+    "config":"config.yaml",
+    "tokenizer_conf": {"bpemodel": "chn_jpn_yue_eng_ko_spectok.bpe.model"},
+    "frontend_conf":{"cmvn_file": "am.mvn"}}
+}
+

configuration_qwen2.py

@@ -0,0 +1,201 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Qwen2 model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class Qwen2MTPSenseVoiceConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen2Model`]. It is used to instantiate a
+    Qwen2 model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Qwen2-7B-beta [Qwen/Qwen2-7B-beta](https://huggingface.co/Qwen/Qwen2-7B-beta).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 151936):
+            Vocabulary size of the Qwen2 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Qwen2Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 22016):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 32):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 28):
+            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import Qwen2Model, Qwen2Config
+
+    >>> # Initializing a Qwen2 style configuration
+    >>> configuration = Qwen2Config()
+
+    >>> # Initializing a model from the Qwen2-7B style configuration
+    >>> model = Qwen2Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "qwen2_mtp_sensevoice"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    # Default tensor parallel plan for base model `Qwen2`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        intermediate_size=22016,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        attention_dropout=0.0,
+        num_nextn_predict_layers=1,
+        mtp_loss_weight=1.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window if use_sliding_window else None
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_dropout = attention_dropout
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+
+        self.num_nextn_predict_layers = num_nextn_predict_layers
+        self.mtp_loss_weight = mtp_loss_weight
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

generation_config.json

@@ -0,0 +1,20 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "mtp_inference_mode": [
+    1,
+    10,
+    4,
+    10
+  ],
+  "pad_token_id": 151643,
+  "repetition_penalty": 1.05,
+  "temperature": 0.7,
+  "top_k": 20,
+  "top_p": 0.8,
+  "transformers_version": "4.48.3"
+}

global_step2000/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a2387a018ef5689c634156404c67099e176675468cd179eb35f9647b7b69091
+size 46470262140

global_step2000/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79af6dea103a515d6712f03014f08e0b0e6773a4c92ab1a80593f9bb441d9c5f
+size 46470262140

global_step2000/zero_pp_rank_0_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98fbfdf66674f6a6520a0ce7f4b17cb4be253584f7ddec5cf9f0992775743ad9
+size 234981668

global_step2000/zero_pp_rank_1_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f3aa576a45f1f6989b7c23ac4a8e00e9b27b1b15ae8055219cfa85213727e51
+size 234981668

latest

@@ -0,0 +1 @@
+global_step2000

model-00001-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f46aed22f63b265b1e7dde506810c9271639d09b70cb67df2a1616fffab0557
+size 4992406120

model-00002-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1948ecf43e46ecb3b1829428825a2bca7ee28622e3e2274ced246cd25fe4e150
+size 4932751008

model-00003-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:73ba64678d56ffe0fbaaf72b950ecc5fe51893846bd7af615eb98d107fdc057f
+size 4828352366

model-00004-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca2741b6f7580e663f61ca77128860cae1f058414027f2ac2091aa4eded4545b
+size 1204740224

modeling_qwen2.py

@@ -0,0 +1,1641 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/qwen2/modular_qwen2.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_qwen2.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+from typing import Callable, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, StaticCache
+from transformers.generation import GenerationMixin
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import (
+    LossKwargs,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_qwen2 import Qwen2MTPSenseVoiceConfig as Qwen2Config
+
+from .modeling_sensevoice import AudioEncoder
+from .resampler_projector import ResamplerProjector
+
+
+logger = logging.get_logger(__name__)
+logger.setLevel(logging.INFO)
+
+_CHECKPOINT_FOR_DOC = "meta-qwen2/Qwen2-2-7b-hf"
+_CONFIG_FOR_DOC = "Qwen2Config"
+
+
+def fixed_cross_entropy(source, target, num_items_in_batch: int = None, ignore_index: int = -100, **kwargs):
+    reduction = "sum" if num_items_in_batch is not None else "mean"
+    loss = nn.functional.cross_entropy(source, target, ignore_index=ignore_index, reduction=reduction)
+    if reduction == "sum":
+        loss = loss / num_items_in_batch
+    return loss
+
+
+def ForCausalLMLoss(
+    logits, labels, vocab_size: int, num_items_in_batch: int = None, ignore_index: int = -100, **kwargs
+):
+    # Upcast to float if we need to compute the loss to avoid potential precision issues
+    # logits = logits.float()
+    labels = labels.to(logits.device)
+    # Shift so that tokens < n predict n
+    shift_logits = logits[..., :-1, :].contiguous()
+    shift_labels = labels[..., 1:].contiguous()
+
+    # Flatten the tokens
+    shift_logits = shift_logits.view(-1, vocab_size)
+    shift_labels = shift_labels.view(-1)
+    # Enable model parallelism
+    shift_labels = shift_labels.to(shift_logits.device)
+    loss = fixed_cross_entropy(shift_logits, shift_labels, num_items_in_batch, ignore_index, **kwargs)
+    return loss
+
+
+def compute_kl_loss(logits, labels):
+    # import pdb;pdb.set_trace()
+    *_, vocab_size = logits.shape
+    # Convert logits to log probabilities
+    log_probs = torch.nn.functional.log_softmax(logits, dim=-1)
+    # Convert labels to probabilities
+    target_probs = torch.nn.functional.softmax(labels, dim=-1)
+    # Define the KL Divergence loss function
+    loss_fct = nn.KLDivLoss(reduction='batchmean')
+    # Compute the loss
+    loss = loss_fct(log_probs.view(-1, vocab_size), target_probs.view(-1, vocab_size))
+    return loss
+
+
+class Qwen2MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+class Qwen2Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Qwen2Config, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=True)
+        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=True)
+        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=True)
+        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        sliding_window = None
+        if (
+            self.config.use_sliding_window
+            and getattr(self.config, "sliding_window", None) is not None
+            and self.layer_idx >= self.config.max_window_layers
+        ):
+            sliding_window = self.config.sliding_window
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
+                logger.warning_once(
+                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+                )
+            else:
+                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            sliding_window=sliding_window,  # main diff with Llama
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class Qwen2RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Qwen2RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class Qwen2DecoderLayer(nn.Module):
+    def __init__(self, config: Qwen2Config, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = Qwen2Attention(config=config, layer_idx=layer_idx)
+        self.mlp = Qwen2MLP(config)
+        self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        if config.sliding_window and config._attn_implementation != "flash_attention_2":
+            logger.warning_once(
+                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
+                "unexpected results may be encountered."
+            )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        return outputs
+
+
+class Qwen2RotaryEmbedding(nn.Module):
+    def __init__(self, config: Qwen2Config, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len)
+            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
+            # This .to() is needed if the model has been moved to a device after being initialized (because
+            # the buffer is automatically moved, but not the original copy)
+            self.original_inv_freq = self.original_inv_freq.to(device)
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+
+        # Core RoPE block
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+QWEN2_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`Qwen2Config`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
+    QWEN2_START_DOCSTRING,
+)
+class Qwen2PreTrainedModel(PreTrainedModel):
+    config_class = Qwen2Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Qwen2DecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+QWEN2_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance, see our
+            [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+
+
+@add_start_docstrings(
+    "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
+    QWEN2_START_DOCSTRING,
+)
+class Qwen2Model(Qwen2PreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen2DecoderLayer`]
+
+    Args:
+        config: Qwen2Config
+    """
+
+    def __init__(self, config: Qwen2Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [Qwen2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Qwen2RotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        self.audio_model = AudioEncoder(config)
+        self.audio_projection = ResamplerProjector(512, config.hidden_size)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        audios: Optional[torch.FloatTensor] = None,
+        audio_indices: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        layer_idxs = None,
+        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        if (past_key_values is None or len(past_key_values) == 0) and audios is not None:
+            audio_embeds, audio_lengths = self.audio_model(audios)
+            # if torch.distributed.get_rank() == 0:
+            #     print(f"audio_embeds {audio_embeds.size()}")
+            assert audio_embeds.shape[0] == len(audios)
+            fake_audios = None
+
+            audio_embeds = self.audio_projection(audio_embeds)
+
+            # torch.set_printoptions(threshold=100_000)
+            # if torch.distributed.get_rank() == 0:
+            #     print(f"audio_embeds {audio_embeds.size()}")
+            #     print(f"audio_embeds {audio_embeds.sum()}")
+            #     print(f"audios {[x.size() for x in audios]}")
+            #     print(f"audios {[x.sum() for x in audios]}")
+            #     print(f"input_ids {input_ids.size()}")
+            #     print(f"input_ids {input_ids.sum()}")
+            #     # print(f"input_ids {input_ids}")
+            #     print(f"audio_indices {[x.size() for x in audio_indices]}")
+            #     print(f"audio_indices {[x.sum() for x in audio_indices]}")
+            #     # print(f"audio_indices {audio_indices}")
+
+        elif self.training:
+            device = self.get_input_embeddings().weight.data.device
+            dtype = self.get_input_embeddings().weight.data.dtype
+            fake_audios = torch.ones((1, 1, 560), dtype=dtype, device=device)
+            audio_embeds, audio_lengths = self.audio_model(fake_audios)
+            audio_embeds = self.audio_projection(audio_embeds)
+
+        else:
+            fake_audios = None
+            audio_embeds = None
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if fake_audios is not None:
+            inputs_embeds = inputs_embeds + audio_embeds.mean() * 0.0
+        elif audio_embeds is not None:
+            inputs_embeds = inputs_embeds.clone()
+            for audio_embeds_, audio_lengths_, audio_indices_ in zip(audio_embeds, audio_lengths, audio_indices,):
+                # print(f"{audio_embeds_.size()=} {audio_lengths_=} {audio_indices_.size()=}")
+                audio_embeds_ = audio_embeds_[:audio_lengths_, ...]
+                audio_embeds_ = audio_embeds_.to(inputs_embeds.device)
+                indices_b, indices_s = audio_indices_.to(inputs_embeds.device).unbind(dim=0)
+                inputs_embeds[indices_b.view(-1), indices_s.view(-1)] = audio_embeds_.view(-1, audio_embeds_.shape[-1])
+            # inputs_embeds = inputs_embeds + audio_embeds.mean() * 0.0
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+
+        if layer_idxs is None:
+            layer_idxs = list(range(self.config.num_hidden_layers))
+        layers = [self.layers[layer_idx] for layer_idx in layer_idxs]
+
+        for decoder_layer in layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                    **flash_attn_kwargs,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                    **flash_attn_kwargs,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        output = BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+        return output if return_dict else output.to_tuple()
+
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and (attention_mask == 0.0).any():
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        **kwargs,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
+                `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache,
+                to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            device (`torch.device`):
+                The device to plcae the 4D attention mask on.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
+
+        return causal_mask
+
+
+class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...
+
+
+class Qwen2MTPSenseVoiceForCausalLM(Qwen2PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Qwen2Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.mtp_projs = nn.ModuleList(
+            [nn.Linear(2 * config.hidden_size, config.hidden_size, bias=False) for _ in range(self.config.num_nextn_predict_layers)]
+        )
+
+        self.mtp_embed_norms = nn.ModuleList([Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) for _ in range(self.config.num_nextn_predict_layers)])
+        self.mtp_hidden_norms = nn.ModuleList([Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) for _ in range(self.config.num_nextn_predict_layers)])
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def mtp_forward(
+        self,
+        mtp_idx,
+        input_ids: torch.LongTensor = None,
+        hidden_states: torch.Tensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        kl_labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        num_logits_to_keep: int = 0,
+        **kwargs: Unpack[KwargsForCausalLM],
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.model.embed_tokens(input_ids)
+            # inputs_embeds = inputs_embeds.to(hidden_states.device)
+
+        inputs_embeds = torch.cat(
+            (
+                self.mtp_embed_norms[mtp_idx](inputs_embeds),
+                self.mtp_hidden_norms[mtp_idx](hidden_states),
+            ),
+            dim=-1,
+        )
+
+        inputs_embeds = self.mtp_projs[mtp_idx](inputs_embeds)
+
+        outputs = self.model(
+            input_ids=None,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            layer_idxs=[self.config.num_hidden_layers - self.config.num_nextn_predict_layers + mtp_idx],
+            **kwargs,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+
+        if labels is not None:
+            loss = []
+            # ce_loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+            ce_loss = ForCausalLMLoss(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+            loss += [ce_loss]
+
+            if False:
+                kl_logits = logits.contiguous()
+                kl_labels = kl_labels.contiguous()
+                kl_loss = compute_kl_loss(kl_logits, kl_labels)
+
+                kl_loss_weight = 1
+                loss += [kl_loss_weight * kl_loss]
+
+            # if self.training and torch.distributed.is_initialized() and torch.distributed.get_rank() == 0:
+            #     with torch.no_grad():
+            #         logger.info(f"\tMTP {mtp_idx=} {loss=}")
+        else:
+            loss = None
+
+        return outputs, logits, loss
+
+    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        audios: Optional[torch.FloatTensor] = None,
+        audio_indices: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        num_logits_to_keep: int = 0,
+        **kwargs: Unpack[KwargsForCausalLM],
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+            num_logits_to_keep (`int`, *optional*):
+                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Qwen2ForCausalLM
+
+        >>> model = Qwen2ForCausalLM.from_pretrained("meta-qwen2/Qwen2-2-7b-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("meta-qwen2/Qwen2-2-7b-hf")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # ===============================================================================================
+        if not self.training:
+            if input_ids is not None:
+                num_input_tokens = input_ids.size(1)
+            if inputs_embeds is not None:
+                num_input_tokens = inputs_embeds.size(1)
+
+            if use_cache:
+                if self.input_ids is None and self.inputs_embeds is None:
+                    if input_ids is not None:
+                        self.input_ids = input_ids
+                    if inputs_embeds is not None:
+                        self.inputs_embeds = inputs_embeds
+                    if position_ids is not None:
+                        self.position_ids = position_ids
+
+                else:
+                    if input_ids is not None:
+                        self.input_ids = torch.cat([self.input_ids, input_ids], dim=1)
+                    if inputs_embeds is not None:
+                        self.inputs_embeds = torch.cat([self.inputs_embeds, inputs_embeds], dim=1)
+                    if position_ids is not None:
+                        self.position_ids = torch.cat([self.position_ids, position_ids], dim=1)
+
+            else:
+                self.input_ids = input_ids
+                self.inputs_embeds = inputs_embeds
+                self.position_ids = position_ids
+
+            self.attention_mask = attention_mask
+
+            if self.num_prefill_tokens < 0:
+                self.num_prefill_tokens = self.input_ids.size(1)
+            num_decode_tokens = self.input_ids.size(1) - self.num_prefill_tokens
+
+            if self.mtp_inference_mode[num_decode_tokens] == "M":
+                self.mtp_idx = -1
+            elif self.mtp_inference_mode[num_decode_tokens] == "m":
+                if self.mtp_inference_mode[num_decode_tokens - 1] == "M":
+                    self.mtp_idx = 0
+                else:
+                    pass
+
+            # if True:
+            if False:
+                print("=" * 100)
+                print(f"{self.mtp_idx=}")
+                print(f"{self.num_prefill_tokens=}")
+                print(f"{num_decode_tokens=}")
+                print(f"{self.mtp_inference_mode=}")
+                if self.input_ids is not None:
+                    print(f"{self.input_ids.size()=}")
+                if self.inputs_embeds is not None:
+                    print(f"{self.inputs_embeds.size()=}")
+                if self.hidden_states[self.mtp_idx] is not None:
+                    print(f"{self.hidden_states[self.mtp_idx].size()=}")
+
+
+            if self.mtp_idx > -1 and self.mtp_idx < self.config.num_nextn_predict_layers and num_input_tokens == 1:
+                layer_idx = self.config.num_hidden_layers - self.config.num_nextn_predict_layers + self.mtp_idx
+
+                if use_cache:
+                    if len(past_key_values.key_cache) > layer_idx:
+                        num_seen_tokens = past_key_values.key_cache[layer_idx].size(2)
+                    else:
+                        num_seen_tokens = 0
+                else:
+                    num_seen_tokens = 0
+
+                hidden_states = self.hidden_states[self.mtp_idx][:, num_seen_tokens:, :]
+
+                if self.input_ids is not None:
+                    input_ids = self.input_ids[:, num_seen_tokens + self.mtp_idx +  1:]
+                if self.inputs_embeds is not None:
+                    inputs_embeds = self.inputs_embeds[:, num_seen_tokens + self.mtp_idx + 1:, :]
+                if self.position_ids is not None:
+                    position_ids = self.position_ids[:, num_seen_tokens + self.mtp_idx + 1:]
+                attention_mask = self.attention_mask[:, num_seen_tokens + self.mtp_idx + 1:]
+
+                if False:
+                # if True:
+                    print("=" * 100)
+                    print(f"{self.mtp_idx=}")
+                    print(f"{layer_idx=}")
+                    if input_ids is not None:
+                        print(f"{input_ids.size()=} {input_ids=}")
+                    if inputs_embeds is not None:
+                        print(f"{inputs_embeds.size()=} {inputs_embeds=}")
+                    print(f"{hidden_states.size()=} {hidden_states=}")
+                    if attention_mask is not None:
+                        print(f"{attention_mask.size()=} {attention_mask=}")
+                    if position_ids is not None:
+                        print(f"{position_ids.size()=} {position_ids=}")
+                    if use_cache and len(past_key_values.key_cache) > layer_idx:
+                        print(f"{past_key_values.key_cache[layer_idx].size()=}")
+                    print(f"{use_cache=}")
+                    print(f"{num_logits_to_keep=}")
+                    print(f"{output_attentions=}")
+                    print(f"{output_hidden_states=}")
+                    print(f"{cache_position=}")
+
+                mtp_outputs, logits, _ = self.mtp_forward(
+                    self.mtp_idx,
+                    input_ids=input_ids,
+                    hidden_states=hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_values=past_key_values,
+                    inputs_embeds=inputs_embeds,
+                    labels=None,
+                    kl_labels=None,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    output_hidden_states=output_hidden_states,
+                    return_dict=return_dict,
+                    cache_position=cache_position,
+                    num_logits_to_keep=num_logits_to_keep,
+                    **kwargs,
+                )
+                hidden_states = mtp_outputs.last_hidden_state
+
+                self.mtp_idx += 1
+                if use_cache:
+                    if self.hidden_states[self.mtp_idx] is None:
+                        self.hidden_states[self.mtp_idx] = hidden_states
+                    else:
+                        self.hidden_states[self.mtp_idx] = torch.cat([self.hidden_states[self.mtp_idx], hidden_states], dim=1)
+
+                else:
+                    self.hidden_states[self.mtp_idx] = hidden_states
+
+                return CausalLMOutputWithPast(
+                    loss=None,
+                    logits=logits,
+                    past_key_values=past_key_values,
+                    hidden_states=mtp_outputs.hidden_states,
+                    attentions=mtp_outputs.attentions,
+                )
+
+            if use_cache and past_key_values is not None:
+                if len(past_key_values.key_cache) > 0:
+                    # print(f"{past_key_values.key_cache[0].size()=}")
+                    num_seen_tokens = past_key_values.key_cache[0].size(2)
+                else:
+                    num_seen_tokens = 0
+            else:
+                num_seen_tokens = 0
+
+            if self.input_ids is not None:
+                input_ids = self.input_ids[:, num_seen_tokens:]
+            if self.inputs_embeds is not None:
+                inputs_embeds = self.inputs_embeds[:, num_seen_tokens:, :]
+            if self.position_ids is not None:
+                position_ids = self.position_ids[:, num_seen_tokens:]
+            attention_mask = attention_mask
+
+        # ===============================================================================================
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            audios=audios,
+            audio_indices=audio_indices,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            layer_idxs=list(range(self.config.num_hidden_layers - self.config.num_nextn_predict_layers)),
+            **kwargs,
+        )
+
+        hidden_states = outputs[0]
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+            # loss = ForCausalLMLoss(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+            # if self.training and torch.distributed.is_initialized() and torch.distributed.get_rank() == 0:
+            #     with torch.no_grad():
+            #         logger.info(f"STP {loss=}")
+
+        # ===============================================================================================
+        if labels is not None and self.config.num_nextn_predict_layers > 0:
+
+            if self.lm_head.weight.requires_grad and False:
+                if inputs_embeds is None:
+                    inputs_embeds = self.model.embed_tokens(input_ids)
+
+                inputs_embeds = inputs_embeds
+                hidden_states = hidden_states
+                kl_labels = logits
+
+            else:
+                with torch.no_grad():
+                    if inputs_embeds is None:
+                        inputs_embeds = self.model.embed_tokens(input_ids)
+
+                    inputs_embeds = inputs_embeds.detach()
+                    hidden_states = hidden_states.detach()
+                    kl_labels = logits.detach()
+
+            if self.lm_head.weight.requires_grad:
+                pass
+            else:
+                loss = 0.0
+
+            for mtp_idx in range(self.config.num_nextn_predict_layers):
+
+                # SFT with data packing
+                if True:
+                    mtp_mask = position_ids > mtp_idx
+                    # input_ids = input_ids[mtp_mask].unsqueeze(0)
+                    inputs_embeds = inputs_embeds[mtp_mask].unsqueeze(0)
+                    if attention_mask is not None:
+                        attention_mask = attention_mask[mtp_mask].unsqueeze(0)
+                    if position_ids is not None:
+                        position_ids = position_ids[mtp_mask].unsqueeze(0)
+                    labels = labels[mtp_mask].unsqueeze(0)
+                    kl_labels = kl_labels[mtp_mask].unsqueeze(0)
+
+                    mtp_mask = torch.cat((mtp_mask[:, 1:], mtp_mask[:, :1]), dim=1)
+                    hidden_states = hidden_states[mtp_mask].unsqueeze(0)
+
+                    cu_seq_lens_q, cu_seq_lens_k, max_length_q, max_length_k = prepare_fa2_from_position_ids_for_mtp(position_ids, mtp_idx)
+                    # kwargs["cu_seq_lens_q"] = cu_seq_lens_q
+                    # kwargs["cu_seq_lens_k"] = cu_seq_lens_k
+                    # kwargs["max_length_q"] = max_length_q
+                    # kwargs["max_length_k"] = max_length_k
+
+                    # print(f"{cu_seq_lens_q}")
+                    # print(f"{cu_seq_lens_k}")
+                    # print(f"{max_length_q}")
+                    # print(f"{max_length_k}")
+
+                mtp_outputs, _, mtp_loss = self.mtp_forward(
+                    mtp_idx,
+                    input_ids=None,
+                    hidden_states=hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_values=past_key_values,
+                    inputs_embeds=inputs_embeds,
+                    labels=labels,
+                    kl_labels=kl_labels,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    output_hidden_states=output_hidden_states,
+                    return_dict=return_dict,
+                    cache_position=cache_position,
+                    num_logits_to_keep=num_logits_to_keep,
+                    cu_seq_lens_q=cu_seq_lens_q,
+                    cu_seq_lens_k=cu_seq_lens_k,
+                    max_length_q=max_length_q,
+                    max_length_k=max_length_k,
+                    **kwargs,
+                )
+
+                loss += sum(mtp_loss) / self.config.num_nextn_predict_layers * self.config.mtp_loss_weight
+
+                hidden_states = mtp_outputs.last_hidden_state
+
+        if not self.training:
+            self.mtp_idx = 0
+
+            if use_cache:
+                if self.hidden_states[self.mtp_idx] is None:
+                    self.hidden_states[self.mtp_idx] = hidden_states
+
+                else:
+                    self.hidden_states[self.mtp_idx] = torch.cat([self.hidden_states[self.mtp_idx], hidden_states], dim=1)
+
+            else:
+                self.hidden_states[self.mtp_idx] = hidden_states
+
+        # ===============================================================================================
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def _prepare_mtp_for_generation(
+        self,
+        mtp_inference_mode,
+        max_new_tokens,
+    ):
+
+        self.input_ids = None
+        self.inputs_embeds = None
+        self.hidden_states = [None] * (self.config.num_nextn_predict_layers + 1)
+        self.position_ids = None
+        self.attention_mask = None
+
+        self.mtp_idx = -1
+        self.num_prefill_tokens = -1
+
+        assert isinstance(mtp_inference_mode, list)
+        assert len(mtp_inference_mode) >= 2
+        assert len(mtp_inference_mode) % 2 == 0
+
+        main_nums = mtp_inference_mode[::2]
+        mtp_nums = mtp_inference_mode[1::2]
+
+        mtp_inference_mode = []
+        while len(mtp_inference_mode) < max_new_tokens:
+
+            if len(mtp_nums) > 1:
+                mtp_num = mtp_nums.pop(0)
+            else:
+                mtp_num = mtp_nums[0]
+
+            if len(main_nums) > 1:
+                main_num = main_nums.pop(0)
+            else:
+                main_num = main_nums[0]
+
+            mtp_inference_mode += "M" * main_num + "m" * mtp_num
+
+        self.mtp_inference_mode = mtp_inference_mode
+
+    def _prepare_cache_for_generation(self, *args, **kwargs):
+
+        generation_config = args[0]
+        mtp_inference_mode = getattr(generation_config, "mtp_inference_mode", [1, self.config.num_nextn_predict_layers])
+        max_new_tokens = generation_config.max_new_tokens
+
+        self._prepare_mtp_for_generation(mtp_inference_mode, max_new_tokens)
+
+        return super()._prepare_cache_for_generation(*args, **kwargs)
+
+
+@add_start_docstrings(
+    """
+    The Qwen2 Model transformer with a sequence classification head on top (linear layer).
+
+    [`Qwen2ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    QWEN2_START_DOCSTRING,
+)
+class Qwen2ForSequenceClassification(Qwen2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = Qwen2Model(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
+            else:
+                sequence_lengths = -1
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config)
+
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The Qwen2 Model transformer with a token classification head on top (a linear layer on top of the hidden-states
+    output) e.g. for Named-Entity-Recognition (NER) tasks.
+    """,
+    QWEN2_START_DOCSTRING,
+)
+class Qwen2ForTokenClassification(Qwen2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = Qwen2Model(config)
+        if getattr(config, "classifier_dropout", None) is not None:
+            classifier_dropout = config.classifier_dropout
+        elif getattr(config, "hidden_dropout", None) is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.score = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = outputs[0]
+        sequence_output = self.dropout(sequence_output)
+        logits = self.score(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.config)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+The Qwen2 Model transformer with a span classification head on top for extractive question-answering tasks like
+SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    QWEN2_START_DOCSTRING,
+)
+class Qwen2ForQuestionAnswering(Qwen2PreTrainedModel):
+    base_model_prefix = "transformer"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = Qwen2Model(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, 2)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.transformer.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.transformer.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.transformer(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        loss = None
+        if start_positions is not None and end_positions is not None:
+            loss = self.loss_function(start_logits, end_logits, start_positions, end_positions, **kwargs)
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+def prepare_fa2_from_position_ids_for_mtp(position_ids, mtp_idx):
+    position_ids = position_ids.flatten()
+    indices_q = torch.arange(position_ids.size(0), device=position_ids.device, dtype=torch.int32)
+
+    cu_seq_lens = torch.cat(
+        (
+            indices_q[position_ids == mtp_idx + 1],
+            torch.tensor(position_ids.size(), device=position_ids.device, dtype=torch.int32),
+        )
+    )
+
+    max_length = position_ids.max() + 1 - 1 - mtp_idx
+
+    return cu_seq_lens, cu_seq_lens, max_length, max_length

modeling_sensevoice.py

@@ -0,0 +1,1260 @@
+
+import time
+import torch
+from torch import nn
+import torch.nn.functional as F
+from typing import Iterable, Optional
+
+from funasr.register import tables
+from funasr.models.ctc.ctc import CTC
+from funasr.utils.datadir_writer import DatadirWriter
+from funasr.models.paraformer.search import Hypothesis
+from funasr.train_utils.device_funcs import force_gatherable
+from funasr.losses.label_smoothing_loss import LabelSmoothingLoss
+from funasr.metrics.compute_acc import compute_accuracy, th_accuracy
+from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank
+# from utils.ctc_alignment import ctc_forced_align
+
+def ctc_forced_align(
+    log_probs: torch.Tensor,
+    targets: torch.Tensor,
+    input_lengths: torch.Tensor,
+    target_lengths: torch.Tensor,
+    blank: int = 0,
+    ignore_id: int = -1,
+) -> torch.Tensor:
+    """Align a CTC label sequence to an emission.
+
+    Args:
+        log_probs (Tensor): log probability of CTC emission output.
+            Tensor of shape `(B, T, C)`. where `B` is the batch size, `T` is the input length,
+            `C` is the number of characters in alphabet including blank.
+        targets (Tensor): Target sequence. Tensor of shape `(B, L)`,
+            where `L` is the target length.
+        input_lengths (Tensor):
+            Lengths of the inputs (max value must each be <= `T`). 1-D Tensor of shape `(B,)`.
+        target_lengths (Tensor):
+            Lengths of the targets. 1-D Tensor of shape `(B,)`.
+        blank_id (int, optional): The index of blank symbol in CTC emission. (Default: 0)
+        ignore_id (int, optional): The index of ignore symbol in CTC emission. (Default: -1)
+    """
+    targets[targets == ignore_id] = blank
+
+    batch_size, input_time_size, _ = log_probs.size()
+    bsz_indices = torch.arange(batch_size, device=input_lengths.device)
+
+    _t_a_r_g_e_t_s_ = torch.cat(
+        (
+            torch.stack((torch.full_like(targets, blank), targets), dim=-1).flatten(start_dim=1),
+            torch.full_like(targets[:, :1], blank),
+        ),
+        dim=-1,
+    )
+    diff_labels = torch.cat(
+        (
+            torch.as_tensor([[False, False]], device=targets.device).expand(batch_size, -1),
+            _t_a_r_g_e_t_s_[:, 2:] != _t_a_r_g_e_t_s_[:, :-2],
+        ),
+        dim=1,
+    )
+
+    neg_inf = torch.tensor(float("-inf"), device=log_probs.device, dtype=log_probs.dtype)
+    padding_num = 2
+    padded_t = padding_num + _t_a_r_g_e_t_s_.size(-1)
+    best_score = torch.full((batch_size, padded_t), neg_inf, device=log_probs.device, dtype=log_probs.dtype)
+    best_score[:, padding_num + 0] = log_probs[:, 0, blank]
+    best_score[:, padding_num + 1] = log_probs[bsz_indices, 0, _t_a_r_g_e_t_s_[:, 1]]
+
+    backpointers = torch.zeros((batch_size, input_time_size, padded_t), device=log_probs.device, dtype=targets.dtype)
+
+    for t in range(1, input_time_size):
+        prev = torch.stack(
+            (best_score[:, 2:], best_score[:, 1:-1], torch.where(diff_labels, best_score[:, :-2], neg_inf))
+        )
+        prev_max_value, prev_max_idx = prev.max(dim=0)
+        best_score[:, padding_num:] = log_probs[:, t].gather(-1, _t_a_r_g_e_t_s_) + prev_max_value
+        backpointers[:, t, padding_num:] = prev_max_idx
+
+    l1l2 = best_score.gather(
+        -1, torch.stack((padding_num + target_lengths * 2 - 1, padding_num + target_lengths * 2), dim=-1)
+    )
+
+    path = torch.zeros((batch_size, input_time_size), device=best_score.device, dtype=torch.long)
+    path[bsz_indices, input_lengths - 1] = padding_num + target_lengths * 2 - 1 + l1l2.argmax(dim=-1)
+
+    for t in range(input_time_size - 1, 0, -1):
+        target_indices = path[:, t]
+        prev_max_idx = backpointers[bsz_indices, t, target_indices]
+        path[:, t - 1] += target_indices - prev_max_idx
+
+    alignments = _t_a_r_g_e_t_s_.gather(dim=-1, index=(path - padding_num).clamp(min=0))
+    return alignments
+
+class SinusoidalPositionEncoder(torch.nn.Module):
+    """ """
+
+    def __int__(self, d_model=80, dropout_rate=0.1):
+        pass
+
+    def encode(
+        self, positions: torch.Tensor = None, depth: int = None, dtype: torch.dtype = torch.float32
+    ):
+        batch_size = positions.size(0)
+        positions = positions.type(dtype)
+        device = positions.device
+        log_timescale_increment = torch.log(torch.tensor([10000], dtype=dtype, device=device)) / (
+            depth / 2 - 1
+        )
+        inv_timescales = torch.exp(
+            torch.arange(depth / 2, device=device).type(dtype) * (-log_timescale_increment)
+        )
+        inv_timescales = torch.reshape(inv_timescales, [batch_size, -1])
+        scaled_time = torch.reshape(positions, [1, -1, 1]) * torch.reshape(
+            inv_timescales, [1, 1, -1]
+        )
+        encoding = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=2)
+        return encoding.type(dtype)
+
+    def forward(self, x):
+        batch_size, timesteps, input_dim = x.size()
+        positions = torch.arange(1, timesteps + 1, device=x.device)[None, :]
+        position_encoding = self.encode(positions, input_dim, x.dtype).to(x.device)
+
+        return x + position_encoding
+
+
+class PositionwiseFeedForward(torch.nn.Module):
+    """Positionwise feed forward layer.
+
+    Args:
+        idim (int): Input dimenstion.
+        hidden_units (int): The number of hidden units.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(self, idim, hidden_units, dropout_rate, activation=torch.nn.ReLU()):
+        """Construct an PositionwiseFeedForward object."""
+        super(PositionwiseFeedForward, self).__init__()
+        self.w_1 = torch.nn.Linear(idim, hidden_units)
+        self.w_2 = torch.nn.Linear(hidden_units, idim)
+        self.dropout = torch.nn.Dropout(dropout_rate)
+        self.activation = activation
+
+    def forward(self, x):
+        """Forward function."""
+        return self.w_2(self.dropout(self.activation(self.w_1(x))))
+
+
+class MultiHeadedAttentionSANM(nn.Module):
+    """Multi-Head Attention layer.
+
+    Args:
+        n_head (int): The number of heads.
+        n_feat (int): The number of features.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(
+        self,
+        n_head,
+        in_feat,
+        n_feat,
+        dropout_rate,
+        kernel_size,
+        sanm_shfit=0,
+        lora_list=None,
+        lora_rank=8,
+        lora_alpha=16,
+        lora_dropout=0.1,
+    ):
+        """Construct an MultiHeadedAttention object."""
+        super().__init__()
+        assert n_feat % n_head == 0
+        # We assume d_v always equals d_k
+        self.d_k = n_feat // n_head
+        self.h = n_head
+        # self.linear_q = nn.Linear(n_feat, n_feat)
+        # self.linear_k = nn.Linear(n_feat, n_feat)
+        # self.linear_v = nn.Linear(n_feat, n_feat)
+
+        self.linear_out = nn.Linear(n_feat, n_feat)
+        self.linear_q_k_v = nn.Linear(in_feat, n_feat * 3)
+        self.attn = None
+        self.dropout = nn.Dropout(p=dropout_rate)
+
+        self.fsmn_block = nn.Conv1d(
+            n_feat, n_feat, kernel_size, stride=1, padding=0, groups=n_feat, bias=False
+        )
+        # padding
+        left_padding = (kernel_size - 1) // 2
+        if sanm_shfit > 0:
+            left_padding = left_padding + sanm_shfit
+        right_padding = kernel_size - 1 - left_padding
+        self.pad_fn = nn.ConstantPad1d((left_padding, right_padding), 0.0)
+
+    def forward_fsmn(self, inputs, mask, mask_shfit_chunk=None):
+        b, t, d = inputs.size()
+        if mask is not None:
+            mask = torch.reshape(mask, (b, -1, 1))
+            if mask_shfit_chunk is not None:
+                mask = mask * mask_shfit_chunk
+            inputs = inputs * mask
+
+        x = inputs.transpose(1, 2)
+        x = self.pad_fn(x)
+        x = self.fsmn_block(x)
+        x = x.transpose(1, 2)
+        x += inputs
+        x = self.dropout(x)
+        if mask is not None:
+            x = x * mask
+        return x
+
+    def forward_qkv(self, x):
+        """Transform query, key and value.
+
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+
+        Returns:
+            torch.Tensor: Transformed query tensor (#batch, n_head, time1, d_k).
+            torch.Tensor: Transformed key tensor (#batch, n_head, time2, d_k).
+            torch.Tensor: Transformed value tensor (#batch, n_head, time2, d_k).
+
+        """
+        b, t, d = x.size()
+        q_k_v = self.linear_q_k_v(x)
+        q, k, v = torch.split(q_k_v, int(self.h * self.d_k), dim=-1)
+        q_h = torch.reshape(q, (b, t, self.h, self.d_k)).transpose(
+            1, 2
+        )  # (batch, head, time1, d_k)
+        k_h = torch.reshape(k, (b, t, self.h, self.d_k)).transpose(
+            1, 2
+        )  # (batch, head, time2, d_k)
+        v_h = torch.reshape(v, (b, t, self.h, self.d_k)).transpose(
+            1, 2
+        )  # (batch, head, time2, d_k)
+
+        return q_h, k_h, v_h, v
+
+    def forward_attention(self, value, scores, mask, mask_att_chunk_encoder=None):
+        """Compute attention context vector.
+
+        Args:
+            value (torch.Tensor): Transformed value (#batch, n_head, time2, d_k).
+            scores (torch.Tensor): Attention score (#batch, n_head, time1, time2).
+            mask (torch.Tensor): Mask (#batch, 1, time2) or (#batch, time1, time2).
+
+        Returns:
+            torch.Tensor: Transformed value (#batch, time1, d_model)
+                weighted by the attention score (#batch, time1, time2).
+
+        """
+        n_batch = value.size(0)
+        if mask is not None:
+            if mask_att_chunk_encoder is not None:
+                mask = mask * mask_att_chunk_encoder
+
+            mask = mask.unsqueeze(1).eq(0)  # (batch, 1, *, time2)
+
+            min_value = -float(
+                "inf"
+            )  # float(numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min)
+            scores = scores.masked_fill(mask, min_value)
+            attn = torch.softmax(scores, dim=-1).masked_fill(
+                mask, 0.0
+            )  # (batch, head, time1, time2)
+        else:
+            attn = torch.softmax(scores, dim=-1)  # (batch, head, time1, time2)
+
+        p_attn = self.dropout(attn)
+        x = torch.matmul(p_attn, value)  # (batch, head, time1, d_k)
+        x = (
+            x.transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k)
+        )  # (batch, time1, d_model)
+
+        return self.linear_out(x)  # (batch, time1, d_model)
+
+    def forward(self, x, mask, mask_shfit_chunk=None, mask_att_chunk_encoder=None):
+        """Compute scaled dot product attention.
+
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
+                (#batch, time1, time2).
+
+        Returns:
+            torch.Tensor: Output tensor (#batch, time1, d_model).
+
+        """
+        q_h, k_h, v_h, v = self.forward_qkv(x)
+        fsmn_memory = self.forward_fsmn(v, mask, mask_shfit_chunk)
+        q_h = q_h * self.d_k ** (-0.5)
+        scores = torch.matmul(q_h, k_h.transpose(-2, -1))
+        att_outs = self.forward_attention(v_h, scores, mask, mask_att_chunk_encoder)
+        return att_outs + fsmn_memory
+
+    def forward_chunk(self, x, cache=None, chunk_size=None, look_back=0):
+        """Compute scaled dot product attention.
+
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
+                (#batch, time1, time2).
+
+        Returns:
+            torch.Tensor: Output tensor (#batch, time1, d_model).
+
+        """
+        q_h, k_h, v_h, v = self.forward_qkv(x)
+        if chunk_size is not None and look_back > 0 or look_back == -1:
+            if cache is not None:
+                k_h_stride = k_h[:, :, : -(chunk_size[2]), :]
+                v_h_stride = v_h[:, :, : -(chunk_size[2]), :]
+                k_h = torch.cat((cache["k"], k_h), dim=2)
+                v_h = torch.cat((cache["v"], v_h), dim=2)
+
+                cache["k"] = torch.cat((cache["k"], k_h_stride), dim=2)
+                cache["v"] = torch.cat((cache["v"], v_h_stride), dim=2)
+                if look_back != -1:
+                    cache["k"] = cache["k"][:, :, -(look_back * chunk_size[1]) :, :]
+                    cache["v"] = cache["v"][:, :, -(look_back * chunk_size[1]) :, :]
+            else:
+                cache_tmp = {
+                    "k": k_h[:, :, : -(chunk_size[2]), :],
+                    "v": v_h[:, :, : -(chunk_size[2]), :],
+                }
+                cache = cache_tmp
+        fsmn_memory = self.forward_fsmn(v, None)
+        q_h = q_h * self.d_k ** (-0.5)
+        scores = torch.matmul(q_h, k_h.transpose(-2, -1))
+        att_outs = self.forward_attention(v_h, scores, None)
+        return att_outs + fsmn_memory, cache
+
+
+class LayerNorm(nn.LayerNorm):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward(self, input):
+        output = F.layer_norm(
+            input.float(),
+            self.normalized_shape,
+            self.weight.float() if self.weight is not None else None,
+            self.bias.float() if self.bias is not None else None,
+            self.eps,
+        )
+        return output.type_as(input)
+
+
+def sequence_mask(lengths, maxlen=None, dtype=torch.float32, device=None):
+    if maxlen is None:
+        maxlen = lengths.max()
+    row_vector = torch.arange(0, maxlen, 1).to(lengths.device)
+    matrix = torch.unsqueeze(lengths, dim=-1)
+    mask = row_vector < matrix
+    mask = mask.detach()
+
+    return mask.to(dtype).to(device) if device is not None else mask.to(dtype)
+    # return mask.type(dtype).to(device) if device is not None else mask.type(dtype)
+
+
+class EncoderLayerSANM(nn.Module):
+    def __init__(
+        self,
+        in_size,
+        size,
+        self_attn,
+        feed_forward,
+        dropout_rate,
+        normalize_before=True,
+        concat_after=False,
+        stochastic_depth_rate=0.0,
+    ):
+        """Construct an EncoderLayer object."""
+        super(EncoderLayerSANM, self).__init__()
+        self.self_attn = self_attn
+        self.feed_forward = feed_forward
+        self.norm1 = LayerNorm(in_size)
+        self.norm2 = LayerNorm(size)
+        self.dropout = nn.Dropout(dropout_rate)
+        self.in_size = in_size
+        self.size = size
+        self.normalize_before = normalize_before
+        self.concat_after = concat_after
+        if self.concat_after:
+            self.concat_linear = nn.Linear(size + size, size)
+        self.stochastic_depth_rate = stochastic_depth_rate
+        self.dropout_rate = dropout_rate
+
+    def forward(self, x, mask, cache=None, mask_shfit_chunk=None, mask_att_chunk_encoder=None):
+        """Compute encoded features.
+
+        Args:
+            x_input (torch.Tensor): Input tensor (#batch, time, size).
+            mask (torch.Tensor): Mask tensor for the input (#batch, time).
+            cache (torch.Tensor): Cache tensor of the input (#batch, time - 1, size).
+
+        Returns:
+            torch.Tensor: Output tensor (#batch, time, size).
+            torch.Tensor: Mask tensor (#batch, time).
+
+        """
+        skip_layer = False
+        # with stochastic depth, residual connection `x + f(x)` becomes
+        # `x <- x + 1 / (1 - p) * f(x)` at training time.
+        stoch_layer_coeff = 1.0
+        if self.training and self.stochastic_depth_rate > 0:
+            skip_layer = torch.rand(1).item() < self.stochastic_depth_rate
+            stoch_layer_coeff = 1.0 / (1 - self.stochastic_depth_rate)
+
+        if skip_layer:
+            if cache is not None:
+                x = torch.cat([cache, x], dim=1)
+            return x, mask
+
+        residual = x
+        if self.normalize_before:
+            x = self.norm1(x)
+
+        if self.concat_after:
+            x_concat = torch.cat(
+                (
+                    x,
+                    self.self_attn(
+                        x,
+                        mask,
+                        mask_shfit_chunk=mask_shfit_chunk,
+                        mask_att_chunk_encoder=mask_att_chunk_encoder,
+                    ),
+                ),
+                dim=-1,
+            )
+            if self.in_size == self.size:
+                x = residual + stoch_layer_coeff * self.concat_linear(x_concat)
+            else:
+                x = stoch_layer_coeff * self.concat_linear(x_concat)
+        else:
+            if self.in_size == self.size:
+                x = residual + stoch_layer_coeff * self.dropout(
+                    self.self_attn(
+                        x,
+                        mask,
+                        mask_shfit_chunk=mask_shfit_chunk,
+                        mask_att_chunk_encoder=mask_att_chunk_encoder,
+                    )
+                )
+            else:
+                x = stoch_layer_coeff * self.dropout(
+                    self.self_attn(
+                        x,
+                        mask,
+                        mask_shfit_chunk=mask_shfit_chunk,
+                        mask_att_chunk_encoder=mask_att_chunk_encoder,
+                    )
+                )
+        if not self.normalize_before:
+            x = self.norm1(x)
+
+        residual = x
+        if self.normalize_before:
+            x = self.norm2(x)
+        x = residual + stoch_layer_coeff * self.dropout(self.feed_forward(x))
+        if not self.normalize_before:
+            x = self.norm2(x)
+
+        return x, mask, cache, mask_shfit_chunk, mask_att_chunk_encoder
+
+    def forward_chunk(self, x, cache=None, chunk_size=None, look_back=0):
+        """Compute encoded features.
+
+        Args:
+            x_input (torch.Tensor): Input tensor (#batch, time, size).
+            mask (torch.Tensor): Mask tensor for the input (#batch, time).
+            cache (torch.Tensor): Cache tensor of the input (#batch, time - 1, size).
+
+        Returns:
+            torch.Tensor: Output tensor (#batch, time, size).
+            torch.Tensor: Mask tensor (#batch, time).
+
+        """
+
+        residual = x
+        if self.normalize_before:
+            x = self.norm1(x)
+
+        if self.in_size == self.size:
+            attn, cache = self.self_attn.forward_chunk(x, cache, chunk_size, look_back)
+            x = residual + attn
+        else:
+            x, cache = self.self_attn.forward_chunk(x, cache, chunk_size, look_back)
+
+        if not self.normalize_before:
+            x = self.norm1(x)
+
+        residual = x
+        if self.normalize_before:
+            x = self.norm2(x)
+        x = residual + self.feed_forward(x)
+        if not self.normalize_before:
+            x = self.norm2(x)
+
+        return x, cache
+
+
+@tables.register("encoder_classes", "SenseVoiceEncoderSmall")
+class SenseVoiceEncoderSmall(nn.Module):
+    """
+    Author: Speech Lab of DAMO Academy, Alibaba Group
+    SCAMA: Streaming chunk-aware multihead attention for online end-to-end speech recognition
+    https://arxiv.org/abs/2006.01713
+    """
+
+    def __init__(
+        self,
+        input_size: int,
+        output_size: int = 256,
+        attention_heads: int = 4,
+        linear_units: int = 2048,
+        num_blocks: int = 6,
+        tp_blocks: int = 0,
+        dropout_rate: float = 0.1,
+        positional_dropout_rate: float = 0.1,
+        attention_dropout_rate: float = 0.0,
+        stochastic_depth_rate: float = 0.0,
+        input_layer: Optional[str] = "conv2d",
+        pos_enc_class=SinusoidalPositionEncoder,
+        normalize_before: bool = True,
+        concat_after: bool = False,
+        positionwise_layer_type: str = "linear",
+        positionwise_conv_kernel_size: int = 1,
+        padding_idx: int = -1,
+        kernel_size: int = 11,
+        sanm_shfit: int = 0,
+        selfattention_layer_type: str = "sanm",
+        **kwargs,
+    ):
+        super().__init__()
+        self._output_size = output_size
+
+        self.embed = SinusoidalPositionEncoder()
+
+        self.normalize_before = normalize_before
+
+        positionwise_layer = PositionwiseFeedForward
+        positionwise_layer_args = (
+            output_size,
+            linear_units,
+            dropout_rate,
+        )
+
+        encoder_selfattn_layer = MultiHeadedAttentionSANM
+        encoder_selfattn_layer_args0 = (
+            attention_heads,
+            input_size,
+            output_size,
+            attention_dropout_rate,
+            kernel_size,
+            sanm_shfit,
+        )
+        encoder_selfattn_layer_args = (
+            attention_heads,
+            output_size,
+            output_size,
+            attention_dropout_rate,
+            kernel_size,
+            sanm_shfit,
+        )
+
+        self.encoders0 = nn.ModuleList(
+            [
+                EncoderLayerSANM(
+                    input_size,
+                    output_size,
+                    encoder_selfattn_layer(*encoder_selfattn_layer_args0),
+                    positionwise_layer(*positionwise_layer_args),
+                    dropout_rate,
+                )
+                for i in range(1)
+            ]
+        )
+        self.encoders = nn.ModuleList(
+            [
+                EncoderLayerSANM(
+                    output_size,
+                    output_size,
+                    encoder_selfattn_layer(*encoder_selfattn_layer_args),
+                    positionwise_layer(*positionwise_layer_args),
+                    dropout_rate,
+                )
+                for i in range(num_blocks - 1)
+            ]
+        )
+
+        self.tp_encoders = nn.ModuleList(
+            [
+                EncoderLayerSANM(
+                    output_size,
+                    output_size,
+                    encoder_selfattn_layer(*encoder_selfattn_layer_args),
+                    positionwise_layer(*positionwise_layer_args),
+                    dropout_rate,
+                )
+                for i in range(tp_blocks)
+            ]
+        )
+
+        self.after_norm = LayerNorm(output_size)
+
+        self.tp_norm = LayerNorm(output_size)
+
+    def output_size(self) -> int:
+        return self._output_size
+
+    def forward(
+        self,
+        xs_pad: torch.Tensor,
+        ilens: torch.Tensor,
+    ):
+        """Embed positions in tensor."""
+        masks = sequence_mask(ilens, dtype=torch.bfloat16, device=ilens.device)[:, None, :]
+        # print(f"{masks=}")
+        # print(f"{ilens=}")
+        # print(f"{(masks>0.5).squeeze(1).sum(1).int()=}")
+
+        xs_pad *= self.output_size() ** 0.5
+
+        xs_pad = self.embed(xs_pad)
+
+        # forward encoder1
+        for layer_idx, encoder_layer in enumerate(self.encoders0):
+            encoder_outs = encoder_layer(xs_pad, masks)
+            xs_pad, masks = encoder_outs[0], encoder_outs[1]
+
+        for layer_idx, encoder_layer in enumerate(self.encoders):
+            encoder_outs = encoder_layer(xs_pad, masks)
+            xs_pad, masks = encoder_outs[0], encoder_outs[1]
+
+        xs_pad = self.after_norm(xs_pad)
+
+        # forward encoder2
+        # olens = masks.squeeze(1).sum(1).int()
+        olens = (masks > 0.5).squeeze(1).sum(1).int()
+
+        for layer_idx, encoder_layer in enumerate(self.tp_encoders):
+            encoder_outs = encoder_layer(xs_pad, masks)
+            xs_pad, masks = encoder_outs[0], encoder_outs[1]
+
+        xs_pad = self.tp_norm(xs_pad)
+        return xs_pad, olens
+
+
+@tables.register("model_classes", "SenseVoiceSmall")
+class SenseVoiceSmall(nn.Module):
+    """CTC-attention hybrid Encoder-Decoder model"""
+
+    def __init__(
+        self,
+        specaug: str = None,
+        specaug_conf: dict = None,
+        normalize: str = None,
+        normalize_conf: dict = None,
+        encoder: str = None,
+        encoder_conf: dict = None,
+        ctc_conf: dict = None,
+        input_size: int = 80,
+        vocab_size: int = -1,
+        ignore_id: int = -1,
+        blank_id: int = 0,
+        sos: int = 1,
+        eos: int = 2,
+        length_normalized_loss: bool = False,
+        **kwargs,
+    ):
+
+        super().__init__()
+
+        if specaug is not None:
+            specaug_class = tables.specaug_classes.get(specaug)
+            specaug = specaug_class(**specaug_conf)
+        if normalize is not None:
+            normalize_class = tables.normalize_classes.get(normalize)
+            normalize = normalize_class(**normalize_conf)
+        encoder_class = tables.encoder_classes.get(encoder)
+        encoder = encoder_class(input_size=input_size, **encoder_conf)
+        encoder_output_size = encoder.output_size()
+
+        if ctc_conf is None:
+            ctc_conf = {}
+        ctc = CTC(odim=vocab_size, encoder_output_size=encoder_output_size, **ctc_conf)
+
+        self.blank_id = blank_id
+        self.sos = sos if sos is not None else vocab_size - 1
+        self.eos = eos if eos is not None else vocab_size - 1
+        self.vocab_size = vocab_size
+        self.ignore_id = ignore_id
+        self.specaug = specaug
+        self.normalize = normalize
+        self.encoder = encoder
+        self.error_calculator = None
+
+        self.ctc = ctc
+
+        self.length_normalized_loss = length_normalized_loss
+        self.encoder_output_size = encoder_output_size
+
+        self.lid_dict = {"auto": 0, "zh": 3, "en": 4, "yue": 7, "ja": 11, "ko": 12, "nospeech": 13}
+        self.lid_int_dict = {24884: 3, 24885: 4, 24888: 7, 24892: 11, 24896: 12, 24992: 13}
+        self.textnorm_dict = {"withitn": 14, "woitn": 15}
+        self.textnorm_int_dict = {25016: 14, 25017: 15}
+        self.embed = torch.nn.Embedding(7 + len(self.lid_dict) + len(self.textnorm_dict), input_size)
+        self.emo_dict = {"unk": 25009, "happy": 25001, "sad": 25002, "angry": 25003, "neutral": 25004}
+        
+        self.criterion_att = LabelSmoothingLoss(
+            size=self.vocab_size,
+            padding_idx=self.ignore_id,
+            smoothing=kwargs.get("lsm_weight", 0.0),
+            normalize_length=self.length_normalized_loss,
+        )
+    
+    @staticmethod
+    def from_pretrained(model:str=None, **kwargs):
+        from funasr import AutoModel
+        model, kwargs = AutoModel.build_model(model=model, trust_remote_code=True, **kwargs)
+        
+        return model, kwargs
+
+    def forward(
+        self,
+        speech: torch.Tensor,
+        speech_lengths: torch.Tensor,
+        text: torch.Tensor,
+        text_lengths: torch.Tensor,
+        **kwargs,
+    ):
+        """Encoder + Decoder + Calc loss
+        Args:
+                speech: (Batch, Length, ...)
+                speech_lengths: (Batch, )
+                text: (Batch, Length)
+                text_lengths: (Batch,)
+        """
+        # import pdb;
+        # pdb.set_trace()
+        if len(text_lengths.size()) > 1:
+            text_lengths = text_lengths[:, 0]
+        if len(speech_lengths.size()) > 1:
+            speech_lengths = speech_lengths[:, 0]
+
+        batch_size = speech.shape[0]
+
+        # 1. Encoder
+        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, text)
+
+        loss_ctc, cer_ctc = None, None
+        loss_rich, acc_rich = None, None
+        stats = dict()
+
+        loss_ctc, cer_ctc = self._calc_ctc_loss(
+            encoder_out[:, 4:, :], encoder_out_lens - 4, text[:, 4:], text_lengths - 4
+        )
+
+        loss_rich, acc_rich = self._calc_rich_ce_loss(
+            encoder_out[:, :4, :], text[:, :4]
+        )
+
+        loss = loss_ctc + loss_rich
+        # Collect total loss stats
+        stats["loss_ctc"] = torch.clone(loss_ctc.detach()) if loss_ctc is not None else None
+        stats["loss_rich"] = torch.clone(loss_rich.detach()) if loss_rich is not None else None
+        stats["loss"] = torch.clone(loss.detach()) if loss is not None else None
+        stats["acc_rich"] = acc_rich
+
+        # force_gatherable: to-device and to-tensor if scalar for DataParallel
+        if self.length_normalized_loss:
+            batch_size = int((text_lengths + 1).sum())
+        loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
+        return loss, stats, weight
+
+    def encode(
+        self,
+        speech: torch.Tensor,
+        speech_lengths: torch.Tensor,
+        text: torch.Tensor,
+        **kwargs,
+    ):
+        """Frontend + Encoder. Note that this method is used by asr_inference.py
+        Args:
+                speech: (Batch, Length, ...)
+                speech_lengths: (Batch, )
+                ind: int
+        """
+
+        # Data augmentation
+        if self.specaug is not None and self.training:
+            speech, speech_lengths = self.specaug(speech, speech_lengths)
+
+        # Normalization for feature: e.g. Global-CMVN, Utterance-CMVN
+        if self.normalize is not None:
+            speech, speech_lengths = self.normalize(speech, speech_lengths)
+
+
+        lids = torch.LongTensor([[self.lid_int_dict[int(lid)] if torch.rand(1) > 0.2 and int(lid) in self.lid_int_dict else 0 ] for lid in text[:, 0]]).to(speech.device)
+        language_query = self.embed(lids)
+        
+        styles = torch.LongTensor([[self.textnorm_int_dict[int(style)]] for style in text[:, 3]]).to(speech.device)
+        style_query = self.embed(styles)
+        speech = torch.cat((style_query, speech), dim=1)
+        speech_lengths += 1
+
+        event_emo_query = self.embed(torch.LongTensor([[1, 2]]).to(speech.device)).repeat(speech.size(0), 1, 1)
+        input_query = torch.cat((language_query, event_emo_query), dim=1)
+        speech = torch.cat((input_query, speech), dim=1)
+        speech_lengths += 3
+
+        encoder_out, encoder_out_lens = self.encoder(speech, speech_lengths)
+
+        return encoder_out, encoder_out_lens
+
+    def _calc_ctc_loss(
+        self,
+        encoder_out: torch.Tensor,
+        encoder_out_lens: torch.Tensor,
+        ys_pad: torch.Tensor,
+        ys_pad_lens: torch.Tensor,
+    ):
+        # Calc CTC loss
+        loss_ctc = self.ctc(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens)
+
+        # Calc CER using CTC
+        cer_ctc = None
+        if not self.training and self.error_calculator is not None:
+            ys_hat = self.ctc.argmax(encoder_out).data
+            cer_ctc = self.error_calculator(ys_hat.cpu(), ys_pad.cpu(), is_ctc=True)
+        return loss_ctc, cer_ctc
+
+    def _calc_rich_ce_loss(
+        self,
+        encoder_out: torch.Tensor,
+        ys_pad: torch.Tensor,
+    ):
+        decoder_out = self.ctc.ctc_lo(encoder_out)
+        # 2. Compute attention loss
+        loss_rich = self.criterion_att(decoder_out, ys_pad.contiguous())
+        acc_rich = th_accuracy(
+            decoder_out.view(-1, self.vocab_size),
+            ys_pad.contiguous(),
+            ignore_label=self.ignore_id,
+        )
+
+        return loss_rich, acc_rich
+
+
+    def inference(
+        self,
+        data_in,
+        data_lengths=None,
+        key: list = ["wav_file_tmp_name"],
+        tokenizer=None,
+        frontend=None,
+        **kwargs,
+    ):
+
+
+        meta_data = {}
+        if (
+            isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank"
+        ):  # fbank
+            speech, speech_lengths = data_in, data_lengths
+            if len(speech.shape) < 3:
+                speech = speech[None, :, :]
+            if speech_lengths is None:
+                speech_lengths = speech.shape[1]
+        else:
+            # extract fbank feats
+            time1 = time.perf_counter()
+            audio_sample_list = load_audio_text_image_video(
+                data_in,
+                fs=frontend.fs,
+                audio_fs=kwargs.get("fs", 16000),
+                data_type=kwargs.get("data_type", "sound"),
+                tokenizer=tokenizer,
+            )
+            time2 = time.perf_counter()
+            meta_data["load_data"] = f"{time2 - time1:0.3f}"
+            speech, speech_lengths = extract_fbank(
+                audio_sample_list, data_type=kwargs.get("data_type", "sound"), frontend=frontend
+            )
+            time3 = time.perf_counter()
+            meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
+            meta_data["batch_data_time"] = (
+                speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
+            )
+
+        speech = speech.to(device=kwargs["device"])
+        speech_lengths = speech_lengths.to(device=kwargs["device"])
+
+        language = kwargs.get("language", "auto")
+        language_query = self.embed(
+            torch.LongTensor(
+                [[self.lid_dict[language] if language in self.lid_dict else 0]]
+            ).to(speech.device)
+        ).repeat(speech.size(0), 1, 1)
+        
+        use_itn = kwargs.get("use_itn", False)
+        output_timestamp = kwargs.get("output_timestamp", False)
+
+        textnorm = kwargs.get("text_norm", None)
+        if textnorm is None:
+            textnorm = "withitn" if use_itn else "woitn"
+        textnorm_query = self.embed(
+            torch.LongTensor([[self.textnorm_dict[textnorm]]]).to(speech.device)
+        ).repeat(speech.size(0), 1, 1)
+        speech = torch.cat((textnorm_query, speech), dim=1)
+        speech_lengths += 1
+
+        event_emo_query = self.embed(torch.LongTensor([[1, 2]]).to(speech.device)).repeat(
+            speech.size(0), 1, 1
+        )
+        input_query = torch.cat((language_query, event_emo_query), dim=1)
+        speech = torch.cat((input_query, speech), dim=1)
+        speech_lengths += 3
+
+        # Encoder
+        encoder_out, encoder_out_lens = self.encoder(speech, speech_lengths)
+        if isinstance(encoder_out, tuple):
+            encoder_out = encoder_out[0]
+
+        # c. Passed the encoder result and the beam search
+        ctc_logits = self.ctc.log_softmax(encoder_out)
+        if kwargs.get("ban_emo_unk", False):
+            ctc_logits[:, :, self.emo_dict["unk"]] = -float("inf")
+
+        results = []
+        b, n, d = encoder_out.size()
+        if isinstance(key[0], (list, tuple)):
+            key = key[0]
+        if len(key) < b:
+            key = key * b
+        for i in range(b):
+            x = ctc_logits[i, : encoder_out_lens[i].item(), :]
+            yseq = x.argmax(dim=-1)
+            yseq = torch.unique_consecutive(yseq, dim=-1)
+
+            ibest_writer = None
+            if kwargs.get("output_dir") is not None:
+                if not hasattr(self, "writer"):
+                    self.writer = DatadirWriter(kwargs.get("output_dir"))
+                ibest_writer = self.writer[f"1best_recog"]
+
+            mask = yseq != self.blank_id
+            token_int = yseq[mask].tolist()
+
+            # Change integer-ids to tokens
+            text = tokenizer.decode(token_int)
+            if ibest_writer is not None:
+                ibest_writer["text"][key[i]] = text
+
+            if output_timestamp:
+                from itertools import groupby
+                timestamp = []
+                tokens = tokenizer.text2tokens(text)[4:]
+
+                logits_speech = self.ctc.softmax(encoder_out)[i, 4:encoder_out_lens[i].item(), :]
+
+                pred = logits_speech.argmax(-1).cpu()
+                logits_speech[pred==self.blank_id, self.blank_id] = 0
+
+                align = ctc_forced_align(
+                    logits_speech.unsqueeze(0).float(),
+                    torch.Tensor(token_int[4:]).unsqueeze(0).long().to(logits_speech.device),
+                    (encoder_out_lens-4).long(),
+                    torch.tensor(len(token_int)-4).unsqueeze(0).long().to(logits_speech.device),
+                    ignore_id=self.ignore_id,
+                )
+
+                pred = groupby(align[0, :encoder_out_lens[0]])
+                _start = 0
+                token_id = 0
+                ts_max = encoder_out_lens[i] - 4
+                for pred_token, pred_frame in pred:
+                    _end = _start + len(list(pred_frame))
+                    if pred_token != 0:
+                        ts_left = max((_start*60-30)/1000, 0)
+                        ts_right = min((_end*60-30)/1000, (ts_max*60-30)/1000)
+                        timestamp.append([tokens[token_id], ts_left, ts_right])
+                        token_id += 1
+                    _start = _end
+
+                result_i = {"key": key[i], "text": text, "timestamp": timestamp}
+                results.append(result_i)
+            else:
+                result_i = {"key": key[i], "text": text}
+                results.append(result_i)
+        return results, meta_data
+
+
+    def inference_encode(
+        self,
+        data_in,
+        data_lengths=None,
+        key: list = ["wav_file_tmp_name"],
+        **kwargs,
+    ):
+
+        # fbank
+        speech, speech_lengths = data_in, data_lengths
+        if len(speech.shape) < 3:
+            speech = speech[None, :, :]
+        if speech_lengths is None:
+            speech_lengths = speech.shape[1]
+
+        speech = speech.to(device=kwargs["device"])
+        speech_lengths = speech_lengths.to(device=kwargs["device"])
+
+        language = kwargs.get("language", "auto")
+        language_query = self.embed(
+            torch.LongTensor(
+                [[self.lid_dict[language] if language in self.lid_dict else 0]]
+            ).to(speech.device)
+        ).repeat(speech.size(0), 1, 1)
+        
+        use_itn = kwargs.get("use_itn", False)
+        output_timestamp = kwargs.get("output_timestamp", False)
+
+        textnorm = kwargs.get("text_norm", None)
+        if textnorm is None:
+            textnorm = "withitn" if use_itn else "woitn"
+        textnorm_query = self.embed(
+            torch.LongTensor([[self.textnorm_dict[textnorm]]]).to(speech.device)
+        ).repeat(speech.size(0), 1, 1)
+        speech = torch.cat((textnorm_query, speech), dim=1)
+        speech_lengths += 1
+
+        event_emo_query = self.embed(torch.LongTensor([[1, 2]]).to(speech.device)).repeat(
+            speech.size(0), 1, 1
+        )
+        input_query = torch.cat((language_query, event_emo_query), dim=1)
+        speech = torch.cat((input_query, speech), dim=1)
+        speech_lengths += 3
+
+        # Encoder
+        encoder_out, encoder_out_lens = self.encoder(speech, speech_lengths)
+        if isinstance(encoder_out, tuple):
+            encoder_out = encoder_out[0]
+
+        return encoder_out, encoder_out_lens
+
+    def export_rebuild_model(model, **kwargs):
+        model.device = kwargs.get("device")
+        model.make_pad_mask = sequence_mask(kwargs["max_seq_len"], flip=False)
+        model.forward = types.MethodType(export_forward, model)
+        model.export_dummy_inputs = types.MethodType(export_dummy_inputs, model)
+        model.export_input_names = types.MethodType(export_input_names, model)
+        model.export_output_names = types.MethodType(export_output_names, model)
+        model.export_dynamic_axes = types.MethodType(export_dynamic_axes, model)
+        model.export_name = types.MethodType(export_name, model)
+        return model
+    
+    def export(self, **kwargs):
+        # from export_meta import export_rebuild_model
+
+        if "max_seq_len" not in kwargs:
+            kwargs["max_seq_len"] = 512
+        models = export_rebuild_model(model=self, **kwargs)
+        return models
+
+
+class AudioEncoder(nn.Module):
+
+    def __init__(
+        self,
+        config,
+    ):
+        super().__init__()
+
+        # TODO
+        # model_dir = "/data/models/FunAudioLLM/SenseVoiceSmall/"
+
+        if "_name_or_path" in config:
+            model_dir = config._name_or_path
+        else:
+            import os
+            model_file= os.path.abspath(__file__)
+            model_dir = os.path.dirname(model_file)
+
+        # self.model, self.kwargs = SenseVoiceSmall.from_pretrained(model_dir, device="cpu")
+        self.model, self.kwargs = self.build_model(model=model_dir, trust_remote_code=False,)
+
+    
+    def forward(
+        self,
+        audios,
+    ):
+
+        from torch.nn.utils.rnn import pad_sequence
+        feats_pad = pad_sequence(audios, batch_first=True, padding_value=0.0)
+        # feats_lens = torch.as_tensor([len(x) + 4 for x in audios])
+        feats_lens = torch.as_tensor([len(x) for x in audios])
+
+        feats_pad = feats_pad.to(torch.bfloat16)
+
+        encoder_out, encoder_out_lens = self.model.inference_encode(
+            feats_pad,
+            data_lengths=feats_lens,
+            language="auto", # "zh", "en", "yue", "ja", "ko", "nospeech"
+            use_itn=False,
+            ban_emo_unk=False,
+            **self.kwargs,
+        )
+
+        return encoder_out, encoder_out_lens
+
+        audio_embeds = []
+        for x, y in zip(encoder_out, encoder_out_lens):
+            audio_embeds.append(x[:y, ...])
+
+        audio_embeds = torch.stack(audio_embeds, dim=0)
+
+        return audio_embeds
+
+    # https://github.com/modelscope/FunASR/blob/main/funasr/auto/auto_model.py
+    @staticmethod
+    def build_model(**kwargs):
+        from omegaconf import DictConfig, ListConfig
+        import os
+
+        from funasr.download.download_model_from_hub import download_model
+        from funasr.train_utils.set_all_random_seed import set_all_random_seed
+        from funasr.register import tables
+        from funasr.train_utils.load_pretrained_model import load_pretrained_model
+        from funasr.utils.misc import deep_update
+
+        import logging
+
+        assert "model" in kwargs
+        if "model_conf" not in kwargs:
+            logging.info("download models from model hub: {}".format(kwargs.get("hub", "ms")))
+            kwargs = download_model(**kwargs)
+
+        set_all_random_seed(kwargs.get("seed", 0))
+
+        device = kwargs.get("device", "cuda")
+        if not torch.cuda.is_available() or kwargs.get("ngpu", 1) == 0:
+            device = "cpu"
+            kwargs["batch_size"] = 1
+        kwargs["device"] = device
+
+        torch.set_num_threads(kwargs.get("ncpu", 4))
+
+        # build tokenizer
+        tokenizer = kwargs.get("tokenizer", None)
+        kwargs["tokenizer"] = tokenizer
+        kwargs["vocab_size"] = -1
+
+        if tokenizer is not None:
+            tokenizers = (
+                tokenizer.split(",") if isinstance(tokenizer, str) else tokenizer
+            )  # type of tokenizers is list!!!
+            tokenizers_conf = kwargs.get("tokenizer_conf", {})
+            tokenizers_build = []
+            vocab_sizes = []
+            token_lists = []
+
+            ### === only for kws ===
+            token_list_files = kwargs.get("token_lists", [])
+            seg_dicts = kwargs.get("seg_dicts", [])
+            ### === only for kws ===
+
+            if not isinstance(tokenizers_conf, (list, tuple, ListConfig)):
+                tokenizers_conf = [tokenizers_conf] * len(tokenizers)
+
+            for i, tokenizer in enumerate(tokenizers):
+                tokenizer_class = tables.tokenizer_classes.get(tokenizer)
+                tokenizer_conf = tokenizers_conf[i]
+
+                ### === only for kws ===
+                if len(token_list_files) > 1:
+                    tokenizer_conf["token_list"] = token_list_files[i]
+                if len(seg_dicts) > 1:
+                    tokenizer_conf["seg_dict"] = seg_dicts[i]
+                ### === only for kws ===
+
+                tokenizer = tokenizer_class(**tokenizer_conf)
+                tokenizers_build.append(tokenizer)
+                token_list = tokenizer.token_list if hasattr(tokenizer, "token_list") else None
+                token_list = (
+                    tokenizer.get_vocab() if hasattr(tokenizer, "get_vocab") else token_list
+                )
+                vocab_size = -1
+                if token_list is not None:
+                    vocab_size = len(token_list)
+
+                if vocab_size == -1 and hasattr(tokenizer, "get_vocab_size"):
+                    vocab_size = tokenizer.get_vocab_size()
+                token_lists.append(token_list)
+                vocab_sizes.append(vocab_size)
+
+            if len(tokenizers_build) <= 1:
+                tokenizers_build = tokenizers_build[0]
+                token_lists = token_lists[0]
+                vocab_sizes = vocab_sizes[0]
+
+            kwargs["tokenizer"] = tokenizers_build
+            kwargs["vocab_size"] = vocab_sizes
+            kwargs["token_list"] = token_lists
+
+        # build frontend
+        frontend = kwargs.get("frontend", None)
+        kwargs["input_size"] = None
+        if frontend is not None:
+            frontend_class = tables.frontend_classes.get(frontend)
+            frontend = frontend_class(**kwargs.get("frontend_conf", {}))
+            kwargs["input_size"] = (
+                frontend.output_size() if hasattr(frontend, "output_size") else None
+            )
+        kwargs["frontend"] = frontend
+        # build model
+        model_class = tables.model_classes.get(kwargs["model"])
+        assert model_class is not None, f'{kwargs["model"]} is not registered'
+        model_conf = {}
+        deep_update(model_conf, kwargs.get("model_conf", {}))
+        deep_update(model_conf, kwargs)
+        model = model_class(**model_conf)
+
+        # init_param
+        init_param = kwargs.get("init_param", None)
+        if init_param is not None:
+            if os.path.exists(init_param):
+                logging.info(f"Loading pretrained params from {init_param}")
+                load_pretrained_model(
+                    model=model,
+                    path=init_param,
+                    ignore_init_mismatch=kwargs.get("ignore_init_mismatch", True),
+                    oss_bucket=kwargs.get("oss_bucket", None),
+                    scope_map=kwargs.get("scope_map", []),
+                    excludes=kwargs.get("excludes", None),
+                )
+            else:
+                print(f"error, init_param does not exist!: {init_param}")
+
+        # fp16
+        if kwargs.get("fp16", False):
+            model.to(torch.float16)
+        elif kwargs.get("bf16", False):
+            model.to(torch.bfloat16)
+        # model.to(device)
+
+        if not kwargs.get("disable_log", True):
+            tables.print()
+
+        return model, kwargs
+

modular_qwen2.py

@@ -0,0 +1,134 @@
+from typing import Callable, Optional, Tuple
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from transformers.cache_utils import Cache
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+from transformers.processing_utils import Unpack
+from transformers.utils import logging
+from transformers.models.llama.modeling_llama import (
+    LlamaAttention,
+    LlamaDecoderLayer,
+    LlamaForCausalLM,
+    LlamaForQuestionAnswering,
+    LlamaForSequenceClassification,
+    LlamaForTokenClassification,
+    LlamaMLP,
+    LlamaModel,
+    apply_rotary_pos_emb,
+    eager_attention_forward,
+)
+from .configuration_qwen2 import Qwen2Config
+
+
+logger = logging.get_logger(__name__)
+
+
+class Qwen2MLP(LlamaMLP):
+    def __init__(self, config):
+        super().__init__(config)
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+
+
+class Qwen2Attention(LlamaAttention):
+    def __init__(self, config: Qwen2Config, layer_idx: int):
+        super().__init__(config, layer_idx)
+        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=True)
+        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=True)
+        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=True)
+        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        sliding_window = None
+        if (
+            self.config.use_sliding_window
+            and getattr(self.config, "sliding_window", None) is not None
+            and self.layer_idx >= self.config.max_window_layers
+        ):
+            sliding_window = self.config.sliding_window
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
+                logger.warning_once(
+                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+                )
+            else:
+                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            sliding_window=sliding_window,  # main diff with Llama
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class Qwen2DecoderLayer(LlamaDecoderLayer):
+    def __init__(self, config: Qwen2Config, layer_idx: int):
+        super().__init__()
+        self.self_attn = Qwen2Attention(config=config, layer_idx=layer_idx)
+        self.mlp = Qwen2MLP(config)
+        if config.sliding_window and config._attn_implementation != "flash_attention_2":
+            logger.warning_once(
+                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
+                "unexpected results may be encountered."
+            )
+
+
+class Qwen2Model(LlamaModel):
+    pass
+
+
+class Qwen2ForCausalLM(LlamaForCausalLM):
+    pass
+
+
+class Qwen2ForSequenceClassification(LlamaForSequenceClassification):
+    pass
+
+
+class Qwen2ForTokenClassification(LlamaForTokenClassification):
+    pass
+
+
+class Qwen2ForQuestionAnswering(LlamaForQuestionAnswering):
+    pass

resampler_projector.py

@@ -0,0 +1,39 @@
+import torch
+import torch.nn as nn
+
+import math
+
+
+
+class ResamplerProjector(nn.Module):
+    def __init__(self, proj_input_size, hidden_size):
+        super().__init__()
+
+        self.pre_proj_layernorm = torch.nn.LayerNorm(proj_input_size)
+
+        self.mlp = nn.Sequential(
+            nn.Linear(proj_input_size, hidden_size, bias=False),
+            nn.GELU(),
+            nn.Linear(hidden_size, hidden_size, bias=False),
+        )
+        self.mlp.apply(init_weights)
+        self.pre_proj_layernorm.apply(init_weights)
+
+    def forward(self, x, *args, **kwargs):
+        x = x.reshape(x.shape[0], -1, x.shape[-1])
+        x = self.pre_proj_layernorm(x)
+        x = self.mlp(x)
+        # print(torch.distributed.get_rank(), {name: [param, param.grad] for name, param in self.pre_proj_layernorm.named_parameters()})
+        # print(torch.distributed.get_rank(), {name: [param, param.grad] for name, param in self.mlp.named_parameters()})
+        return x
+
+def init_weights(m):
+    if isinstance(m, nn.Linear):
+        torch.nn.init.normal_(m.weight, mean=0.0, std=0.02)
+        if m.bias is not None:
+            torch.nn.init.zeros_(m.bias)
+
+    if isinstance(m, nn.LayerNorm):
+        torch.nn.init.ones_(m.weight)
+        torch.nn.init.zeros_(m.bias)
+

rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e0c5997f3bf51022d1fc9d5bd40ee43afbde6e06c35b7c07ae6ed3e8a10114e
+size 14512

rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2ee512dc8bef76c04363f1c2ebc364608236b8ffdbdc44b00700849f3b33d38
+size 14512

special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf1020c1c9bf019e852c66c400521fd5c4199d958be84e3464677bb3800517b9
+size 14560566

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fcd17813e9a722a3efdb06a4e46335480a6f2c10561124b7ab986312165b730
+size 8376

zero_to_fp32.py

@@ -0,0 +1,674 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import json
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering Sharded Weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        state_dict_split = split_torch_state_dict_into_shards(state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard = {tensor: state_dict[tensor].contiguous() for tensor in tensors}
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard, output_path)
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)