Upload 12 files

captioning.py

@@ -0,0 +1,88 @@
+# Copyright (2023) Tsinghua University, Bytedance Ltd. and/or its affiliates
+#
+# 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.
+
+import os
+import torch
+import argparse
+import json
+import pandas as pd
+import copy
+import numpy as np
+from tqdm import tqdm
+from model import SALMONN
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device", type=str, default="cuda")
+    parser.add_argument("--ckpt_path", type=str, default='./salomnn_7b.bin')
+    parser.add_argument("--whisper_path", type=str, default='whisper-large-v2')
+    parser.add_argument("--beats_path", type=str, default='BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt')
+    parser.add_argument("--vicuna_path", type=str, default='vicuna-7b-v1.5')
+    parser.add_argument("--audio_path", type=str, default='./Harmonixset/music_data')
+    parser.add_argument("--meta_path", type=str, default='./Harmonixset/metadata.csv')
+    parser.add_argument("--segment_path", type=str, default='./Harmonixset/segments')
+    parser.add_argument("--caption_path", type=str, default='./Harmonixset/captions')
+    parser.add_argument("--low_resource", action='store_true', default=False)
+    parser.add_argument("--debug", action="store_true", default=False)
+
+    args = parser.parse_args()
+
+    os.makedirs(args.caption_path, exist_ok=True)
+    meta = pd.read_csv(args.meta_path, header=0)[['File', 'BPM', 'Genre']]
+    samples = []
+    for i, row in meta.iterrows():
+        fname = row['File']
+        sample = row.to_dict()
+        sample['audio'] = f'{args.audio_path}/{fname}.wav'
+        sample['segment'] = f'{args.segment_path}/{fname}.txt'
+        if os.path.exists(sample['audio']) and os.path.exists(sample['segment']):
+            samples.append(sample)
+
+
+    model = SALMONN(
+        ckpt=args.ckpt_path,
+        whisper_path=args.whisper_path,
+        beats_path=args.beats_path,
+        vicuna_path=args.vicuna_path
+    ).to(torch.float16).cuda()
+    model.eval()
+
+    # prompt_tmp = 'Please describe functional music segments and their time boundaries. In each segment, describe the musical change of each segment and provide detailed analysis.'
+    # prompt_tmp = 'First describe the music in general in terms of mood, theme, tempo, melody, instruments and chord progression. Then provide a detailed music analysis by describing each functional segment and its time boundaries.'
+    prompt_tmp = 'This is a {genre} music of {bpm} beat-per-minute (BPM). First describe the music in general in terms of mood, theme, tempo, melody, instruments and chord progression. Then provide a detailed music analysis by describing each functional segment and its time boundaries. Please note that the music boundaries are {segments}.'
+
+    with torch.cuda.amp.autocast(dtype=torch.float16):
+        for sample in tqdm(samples):
+            fname = sample['File']
+            if os.path.exists(f'{args.caption_path}/{fname}.json'):
+                continue
+            # try:
+            wav_path = sample['audio']
+            ts, tag = zip(*[line.split(' ') for line in open(sample['segment']) if 'silence' not in line and line.strip()])
+            ts = np.asarray([float(t) for t in ts])
+            bdr = (ts[0], ts[-1])
+            ts = (ts - ts[0]) / (ts[-1] - ts[0])
+            ts = [np.round(t * 100) for t in ts]
+
+            prompt = prompt_tmp.format(genre=sample['Genre'], bpm=sample['BPM'], segments=ts)
+
+            save_sample = copy.deepcopy(sample)
+            captions = model.generate(wav_path, prompt=prompt, bdr=bdr, repetition_penalty=1.5, num_return_sequences=5, num_beams=10)
+            save_sample['tags'] = tag
+            save_sample['ts'] = ts
+            save_sample['captions'] = captions
+            json.dump(save_sample, open(f'{args.caption_path}/{fname}.json', 'w'))
+            # except Exception as e:
+            #     print(e)

cli_inference.py

@@ -0,0 +1,56 @@
+# Copyright (2023) Tsinghua University, Bytedance Ltd. and/or its affiliates
+#
+# 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.
+
+import torch
+import argparse
+from model import SALMONN
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device", type=str, default="cuda")
+    parser.add_argument("--ckpt_path", type=str, default='./salomnn_7b.bin')
+    parser.add_argument("--whisper_path", type=str, default='whisper-large-v2')
+    parser.add_argument("--beats_path", type=str, default='BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt')
+    parser.add_argument("--vicuna_path", type=str, default='vicuna-7b-v1.5')
+    parser.add_argument("--low_resource", action='store_true', default=False)
+    parser.add_argument("--debug", action="store_true", default=False)
+
+    args = parser.parse_args()
+
+    model = SALMONN(
+        ckpt=args.ckpt_path,
+        whisper_path=args.whisper_path,
+        beats_path=args.beats_path,
+        vicuna_path=args.vicuna_path
+    ).to(torch.float16).cuda()
+
+    prompt = 'First describe the music in general in terms of mood, theme, tempo, melody, instruments and chord progression. Then provide a detailed music analysis by describing each functional segment and its time boundaries.'
+    prompt_tmp = 'This is a Pop music of 69 beat-per-minute (BPM). First describe the music in general in terms of mood, theme, tempo, melody, instruments and chord progression. Then provide a detailed music analysis by describing each functional segment and its time boundaries. Please note that the music boundaries are [0, 41, 58, 83, 100].'
+    model.eval()
+    while True:
+        print("=====================================")
+        wav_path = input("Your Wav Path:\n")
+        prompt = input("Your Prompt:\n")
+        try:
+            print("Output:")
+            # for environment with cuda>=117
+            with torch.cuda.amp.autocast(dtype=torch.float16):
+                print(model.generate(wav_path, prompt=prompt, repetition_penalty=1.5, num_beams=10, top_p=.7, temperature=.2)[0])
+        except Exception as e:
+            print(e)
+            if args.debug:
+                import pdb
+
+                pdb.set_trace()

dataset.py

@@ -0,0 +1,124 @@
+import random
+import copy
+import json
+import torch
+import transformers
+import numpy as np
+import pickle as pkl
+from torch.utils.data import Dataset
+from dataclasses import dataclass, field
+from typing import Dict, Optional, Sequence, List
+
+IGNORE_INDEX = -100
+MAX_LENGTH = 2048
+
+
+@dataclass
+class DataArguments:
+    data_path: str = field(default='./MusicCaps', metadata={"help": "Path to the training data."})
+    feat_folder: Optional[str] = field(default='./MusicCaps/music_feat')
+
+def preprocess_v1(sources: str, tokenizer: transformers.PreTrainedTokenizer, metadata,
+                  prompt_pattern="USER: <Speech><SpeechHere></Speech> Describe the music in detail.\nASSISTANT:\n") -> Dict:
+    sources = sources.split('\n')
+    clips, duration, caption = metadata['clips'], metadata['duration'], []
+    length = 0
+    for l, c in zip(clips, sources):
+        caption.append(
+            f'From {int(length / duration * 100)} to {int((length + l) / duration * 100)},'
+            + ','.join(c.split(',')[1:])
+        )
+        length += l
+
+    targets = prompt_pattern + '\n'.join(caption)
+
+    targets_left, targets_right = targets.split('<SpeechHere>')
+    targets_right = tokenizer(targets_right, return_tensors="pt", add_special_tokens=False).input_ids[0]
+
+    sources_left, sources_right = prompt_pattern.split('<SpeechHere>')
+    sources_left = tokenizer(sources_left, return_tensors="pt", add_special_tokens=False).input_ids[0]
+    sources_right_length = tokenizer(sources_right, return_tensors="pt", add_special_tokens=False).input_ids.shape[-1]
+
+    sources_right = copy.deepcopy(targets_right)
+
+    targets_left = torch.LongTensor([IGNORE_INDEX] * len(sources_left))
+    targets_right[:sources_right_length] = IGNORE_INDEX
+
+    sources_right, targets_right = sources_right[:MAX_LENGTH], targets_right[:MAX_LENGTH]
+
+    return dict(input_ids=(sources_left, sources_right), labels=(targets_left, targets_right))
+
+
+def preprocess(sources: str, tokenizer: transformers.PreTrainedTokenizer, metadata,
+               prompt_pattern="USER: <Speech><SpeechHere></Speech> Describe the music in detail.\nASSISTANT:\n") -> Dict:
+    targets = prompt_pattern + sources
+
+    targets_left, targets_right = targets.split('<SpeechHere>')
+    targets_right = tokenizer(targets_right, return_tensors="pt", add_special_tokens=False).input_ids[0]
+
+    sources_left, sources_right = prompt_pattern.split('<SpeechHere>')
+    sources_left = tokenizer(sources_left, return_tensors="pt", add_special_tokens=False).input_ids[0]
+    sources_right_length = tokenizer(sources_right, return_tensors="pt", add_special_tokens=False).input_ids.shape[-1]
+
+    sources_right = copy.deepcopy(targets_right)
+
+    targets_left = torch.LongTensor([IGNORE_INDEX] * len(sources_left))
+    targets_right[:sources_right_length] = IGNORE_INDEX
+
+    sources_right, targets_right = sources_right[:MAX_LENGTH], targets_right[:MAX_LENGTH]
+
+    return dict(input_ids=(sources_left, sources_right), labels=(targets_left, targets_right))
+
+
+class LazySupervisedDataset(Dataset):
+    """Dataset for supervised fine-tuning."""
+
+    def __init__(self, data_path, tokenizer, data_args):
+        super(LazySupervisedDataset, self).__init__()
+
+        self.tokenizer = tokenizer
+        self.list_data_dict = json.load(open(data_path, "r"))
+        self.data_args = data_args
+
+    def __len__(self):
+        return len(self.list_data_dict)
+
+    def __getitem__(self, i):
+        source = copy.deepcopy(self.list_data_dict[i])
+
+        feature_path = '{}/{}.pkl'.format(self.data_args.feat_folder, source['id'])  # Added
+        music = pkl.load(open(feature_path, 'rb'))  # <N, 768> float16
+        speech = torch.from_numpy(music['speech'])
+        audio = torch.from_numpy(music['audio'])
+
+        captions = source['caption']
+        if not isinstance(captions, str):
+            weights = np.asarray([len(c) for c in captions])
+            weights = weights / weights.sum()
+            captions = random.choices(captions, weights, k=1)[0]
+
+        data_dict = preprocess(captions, self.tokenizer, source['meta'])
+        
+        data_dict['speeches'] = speech
+        data_dict['audios'] = audio
+        return data_dict
+
+
+@dataclass
+class DataCollatorForSupervisedDataset(object):
+    """Collate examples for supervised fine-tuning."""
+
+    tokenizer: transformers.PreTrainedTokenizer
+
+    def __call__(self, instances):
+        input_ids, labels, speeches, audios = tuple(
+            [instance[key] for instance in instances] for key in ("input_ids", "labels", "speeches", "audios"))
+        batch = dict(input_ids=input_ids, labels=labels, speeches=speeches, audios=audios)
+        return batch
+
+
+def make_supervised_data_module(tokenizer: transformers.PreTrainedTokenizer, data_args) -> Dict:
+    """Make dataset and collator for supervised fine-tuning."""
+    train_dataset = LazySupervisedDataset(tokenizer=tokenizer, data_path=data_args.data_path, data_args=data_args)
+    data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer)
+    return dict(train_dataset=train_dataset, eval_dataset=None, data_collator=data_collator)

evaluation.py

@@ -0,0 +1,76 @@
+# Copyright (2023) Tsinghua University, Bytedance Ltd. and/or its affiliates
+#
+# 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.
+
+import os
+import torch
+import argparse
+import json
+import pandas as pd
+import copy
+import numpy as np
+from tqdm import tqdm
+from model import SALMONN
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device", type=str, default="cuda")
+    parser.add_argument("--ckpt_path", type=str, default='./salomnn_7b.bin')
+    parser.add_argument("--whisper_path", type=str, default='whisper-large-v2')
+    parser.add_argument("--beats_path", type=str, default='BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt')
+    parser.add_argument("--vicuna_path", type=str, default='vicuna-7b-v1.5')
+    parser.add_argument("--audio_path", type=str, default='./Harmonixset/music_data')
+    parser.add_argument("--caption_path", type=str, default='./Harmonixset/captions')
+    parser.add_argument("--start", type=int, default=0)
+    parser.add_argument("--end", type=int, default=10000)
+    parser.add_argument("--low_resource", action='store_true', default=False)
+    parser.add_argument("--debug", action="store_true", default=False)
+
+    args = parser.parse_args()
+
+    os.makedirs(args.caption_path, exist_ok=True)
+
+    model = SALMONN(
+        ckpt=args.ckpt_path,
+        whisper_path=args.whisper_path,
+        beats_path=args.beats_path,
+        vicuna_path=args.vicuna_path
+    ).to(torch.float16).cuda()
+    model.eval()
+
+    prompt_tmp = 'First describe the music in general in terms of mood, theme, tempo, melody, instruments and chord progression. Then provide a detailed music analysis by describing each functional segment and its time boundaries.'
+
+    sample_list = os.listdir(args.audio_path)[args.start:args.end]
+    with torch.cuda.amp.autocast(dtype=torch.float16):
+        for sample in tqdm(sample_list):
+            if os.path.exists(f'{args.caption_path}/{sample}.json'):
+                continue
+            try:
+                wav_path = f'{args.audio_path}/{sample}'
+                prompt = prompt_tmp
+                save_sample = {'wav_path': sample}
+                captions = model.generate(
+                    wav_path,
+                    prompt=prompt,
+                    bdr=(0, 180),
+                    repetition_penalty=1.5,
+                    num_return_sequences=1,
+                    num_beams=5,
+                    top_p=0.95,
+                    top_k=50,
+                )
+                save_sample['captions'] = captions
+                json.dump(save_sample, open(f'{args.caption_path}/{sample}.json', 'w'))
+            except Exception as e:
+                print(e)

llama_flash_attn_monkey_patch.py

@@ -0,0 +1,124 @@
+from typing import List, Optional, Tuple
+import logging
+
+import torch
+from torch import nn
+
+import transformers
+from transformers.models.llama.modeling_llama import apply_rotary_pos_emb
+
+from einops import rearrange
+
+try:
+    from flash_attn.flash_attn_interface import flash_attn_unpadded_qkvpacked_func
+except ImportError:
+    from flash_attn.flash_attn_interface import flash_attn_varlen_qkvpacked_func as flash_attn_unpadded_qkvpacked_func
+from flash_attn.bert_padding import unpad_input, pad_input
+
+
+def forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.Tensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    """Input shape: Batch x Time x Channel
+
+    attention_mask: [bsz, q_len]
+    """
+    bsz, q_len, _ = hidden_states.size()
+
+    query_states = (
+        self.q_proj(hidden_states)
+        .view(bsz, q_len, self.num_heads, self.head_dim)
+        .transpose(1, 2)
+    )
+    key_states = (
+        self.k_proj(hidden_states)
+        .view(bsz, q_len, self.num_heads, self.head_dim)
+        .transpose(1, 2)
+    )
+    value_states = (
+        self.v_proj(hidden_states)
+        .view(bsz, q_len, self.num_heads, self.head_dim)
+        .transpose(1, 2)
+    )
+    # [bsz, q_len, nh, hd]
+    # [bsz, nh, q_len, hd]
+
+    kv_seq_len = key_states.shape[-2]
+    assert past_key_value is None, "past_key_value is not supported"
+
+    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+    query_states, key_states = apply_rotary_pos_emb(
+        query_states, key_states, cos, sin, position_ids
+    )
+    # [bsz, nh, t, hd]
+    assert not output_attentions, "output_attentions is not supported"
+    assert not use_cache, "use_cache is not supported"
+
+    # Flash attention codes from
+    # https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/flash_attention.py
+
+    # transform the data into the format required by flash attention
+    qkv = torch.stack(
+        [query_states, key_states, value_states], dim=2
+    )  # [bsz, nh, 3, q_len, hd]
+    qkv = qkv.transpose(1, 3)  # [bsz, q_len, 3, nh, hd]
+    # We have disabled _prepare_decoder_attention_mask in LlamaModel
+    # the attention_mask should be the same as the key_padding_mask
+    key_padding_mask = attention_mask
+
+    if key_padding_mask is None:
+        qkv = rearrange(qkv, "b s ... -> (b s) ...")
+        max_s = q_len
+        cu_q_lens = torch.arange(
+            0, (bsz + 1) * q_len, step=q_len, dtype=torch.int32, device=qkv.device
+        )
+        output = flash_attn_unpadded_qkvpacked_func(
+            qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
+        )
+        output = rearrange(output, "(b s) ... -> b s ...", b=bsz)
+    else:
+        nheads = qkv.shape[-2]
+        x = rearrange(qkv, "b s three h d -> b s (three h d)")
+        x_unpad, indices, cu_q_lens, max_s = unpad_input(x, key_padding_mask)
+        x_unpad = rearrange(
+            x_unpad, "nnz (three h d) -> nnz three h d", three=3, h=nheads
+        )
+        output_unpad = flash_attn_unpadded_qkvpacked_func(
+            x_unpad, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
+        )
+        output = rearrange(
+            pad_input(
+                rearrange(output_unpad, "nnz h d -> nnz (h d)"), indices, bsz, q_len
+            ),
+            "b s (h d) -> b s h d",
+            h=nheads,
+        )
+    return self.o_proj(rearrange(output, "b s h d -> b s (h d)")), None, None
+
+
+# Disable the transformation of the attention mask in LlamaModel as the flash attention
+# requires the attention mask to be the same as the key_padding_mask
+def _prepare_decoder_attention_mask(
+    self, attention_mask, input_shape, inputs_embeds, past_key_values_length
+):
+    # [bsz, seq_len]
+    return attention_mask
+
+
+def replace_llama_attn_with_flash_attn():
+    cuda_major, cuda_minor = torch.cuda.get_device_capability()
+    if cuda_major < 8:
+        logging.warning(
+            "Flash attention is only supported on A100 or H100 GPU during training due to head dim > 64 backward."
+            "ref: https://github.com/HazyResearch/flash-attention/issues/190#issuecomment-1523359593"
+        )
+    transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (
+        _prepare_decoder_attention_mask
+    )
+    transformers.models.llama.modeling_llama.LlamaAttention.forward = forward

model.py

@@ -0,0 +1,317 @@
+# Copyright (2023) Tsinghua University, Bytedance Ltd. and/or its affiliates
+#
+# 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.
+
+
+import torch
+import soundfile as sf
+import torch.nn as nn
+import torch.nn.functional as F
+from peft import LoraConfig, TaskType, get_peft_model
+from transformers import (
+    WhisperFeatureExtractor,
+    WhisperModel,
+    # LlamaForCausalLM,
+    LlamaTokenizer
+)
+from modeling_llama import LlamaForCausalLM
+import librosa
+from beats.BEATs import BEATsConfig, BEATs
+from qformer.Qformer import BertConfig, BertLMHeadModel
+from typing import List, Optional, Tuple, Union
+
+IGNORE_INDEX = -100
+
+
+class SALMONN(nn.Module):
+    def __init__(self, ckpt, whisper_path, beats_path, vicuna_path,
+                 speech_qformer_token_num=1, speech_qformer_layer=2,
+                 lora=True, lora_alpha=32, lora_rank=8, lora_dropout=0.1,
+                 second_per_frame=0.333333, second_stride=0.333333, compute_dtype=torch.float16):
+
+        super().__init__()
+
+        # feature_extractor
+        self.feature_extractor = WhisperFeatureExtractor.from_pretrained(whisper_path)
+
+        # whisper
+        self.speech_encoder = WhisperModel.from_pretrained(whisper_path).encoder
+        for name, param in self.speech_encoder.named_parameters():
+            param.requires_grad = False
+        self.ln_speech = nn.LayerNorm(self.speech_encoder.config.d_model)
+        print('Whisper model loaded ........')
+
+        # beats
+        self.beats_ckpt = beats_path
+        beats_checkpoint = torch.load(self.beats_ckpt, map_location='cpu')
+        beats = BEATs(BEATsConfig(beats_checkpoint['cfg']))
+        beats.load_state_dict(beats_checkpoint['model'])
+        self.beats = beats
+        for name, param in self.beats.named_parameters():
+            param.requires_grad = False
+        self.beats.eval()
+        self.ln_audio = nn.LayerNorm(self.beats.cfg.encoder_embed_dim)
+        print('Beats model loaded ........')
+
+        # init speech Qformer
+        self.speech_Qformer, self.speech_query_tokens = self.init_speech_Qformer(
+            speech_qformer_token_num,
+            self.speech_encoder.config.d_model + self.beats.cfg.encoder_embed_dim,
+            speech_qformer_layer,
+        )
+        self.second_per_frame = second_per_frame
+        self.second_stride = second_stride
+        print('Qformer model initialised ........')
+
+        # vicuna
+        self.llama_model = LlamaForCausalLM.from_pretrained(vicuna_path, torch_dtype=compute_dtype)
+        self.config = self.llama_model.config
+        print('Vicuna model loaded ........')
+
+        # lora
+        self.lora = lora
+        if lora:
+            target_modules = None
+            self.peft_config = LoraConfig(
+                task_type=TaskType.CAUSAL_LM, inference_mode=False, target_modules=target_modules,
+                r=lora_rank, lora_alpha=lora_alpha, lora_dropout=lora_dropout,
+            )
+            self.llama_model = get_peft_model(self.llama_model, self.peft_config)
+            print('Added LoRA ........')
+
+        # tokenizer
+        self.tokenizer = LlamaTokenizer.from_pretrained(vicuna_path, use_fast=False)
+        self.tokenizer.add_special_tokens({'pad_token': '[PAD]'})
+        self.tokenizer.padding_side = "right"
+
+        # proj
+        self.speech_llama_proj = nn.Linear(self.speech_Qformer.config.hidden_size, self.llama_model.config.hidden_size)
+
+        # load ckpt
+        print('Loading Parameters ........')
+        ckpt_dict = torch.load(ckpt, map_location='cpu')
+        if 'model' in ckpt_dict:
+            ckpt_dict = ckpt_dict['model']
+        for name, param in ckpt_dict.items():
+            if name in self.state_dict():
+                print('Loaded:', name)
+        self.load_state_dict(ckpt_dict, strict=False)
+
+    def forward(
+            self,
+            input_ids, labels, speeches, audios,
+            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,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ):
+        speech_embeds, sources, targets = [], [], []
+        for speech_embed, audio_embed, input_id, label in zip(speeches, audios, input_ids, labels):
+            speech_embed, audio_embed = speech_embed.to('cuda'), audio_embed.to('cuda')
+            # auditory embeds
+            speech_embed = self.ln_speech(speech_embed)
+            audio_embed = self.ln_audio(audio_embed)
+            audio_embed = F.pad(audio_embed, (0, 0, 0, speech_embed.size(1) - audio_embed.size(1)))
+            speech_embed = torch.cat([speech_embed, audio_embed], dim=-1)
+
+            # split frames
+            B, T, C = speech_embed.shape
+            kernel, stride = round(T * self.second_per_frame / 30.0), round(T * self.second_stride / 30.0)
+            kernel, stride = (1, kernel), (1, stride)
+            speech_embeds_tr = speech_embed.transpose(1, 2).unsqueeze(2)
+            speech_embeds_overlap = F.unfold(speech_embeds_tr, kernel_size=kernel, dilation=1, padding=0, stride=stride)
+            _, _, L = speech_embeds_overlap.shape
+            speech_embeds_overlap = speech_embeds_overlap.view(B, -1, kernel[1], L)
+            speech_embeds_overlap = torch.permute(speech_embeds_overlap, [0, 3, 2, 1])
+            speech_embed = speech_embeds_overlap.reshape(-1, kernel[1], C)
+            speech_atts = torch.ones(speech_embed.size()[:-1], dtype=torch.long, device=speech_embed.device)
+
+            # Qformer
+            query_tokens = self.speech_query_tokens.expand(speech_embed.shape[0], -1, -1)
+            query_output = self.speech_Qformer.bert(
+                query_embeds=query_tokens,
+                encoder_hidden_states=speech_embed,
+                encoder_attention_mask=speech_atts,
+                return_dict=True,
+                use_cache=False
+            )
+            speech_embed = self.speech_llama_proj(query_output.last_hidden_state)
+            speech_embed = speech_embed.view(B, -1, speech_embed.size(2)).contiguous()
+
+            sources.append(
+                torch.concat([
+                    torch.LongTensor([self.tokenizer.bos_token_id]).to(input_id[0].device),
+                    input_id[0],
+                    torch.LongTensor([self.tokenizer.bos_token_id] * speech_embed.shape[1]).to(input_id[0].device),
+                    input_id[1],
+                    torch.LongTensor([self.tokenizer.eos_token_id]).to(input_id[0].device),
+                ])
+            )
+            targets.append(
+                torch.concat([
+                    torch.LongTensor([IGNORE_INDEX]).to(label[0].device),
+                    label[0],
+                    torch.LongTensor([IGNORE_INDEX] * speech_embed.shape[1]).to(label[0].device),
+                    label[1],
+                    torch.LongTensor([self.tokenizer.eos_token_id]).to(label[0].device),
+                ])
+            )
+            speech_embeds.append(speech_embed)
+
+        start_length = len(input_ids[0][0]) + 1
+
+        # USER: <Speech>speech_embeds<Speech> prompt\nASSISTANT:
+        PADDING_TOKEN = 0
+        embed_tokens = self.llama_model.model.model.embed_tokens if self.lora else self.llama_model.model.embed_tokens
+
+        input_ids = torch.nn.utils.rnn.pad_sequence(sources, batch_first=True, padding_value=PADDING_TOKEN)
+        labels = torch.nn.utils.rnn.pad_sequence(targets, batch_first=True, padding_value=PADDING_TOKEN)
+        attention_mask = input_ids.ne(PADDING_TOKEN)
+
+        inputs_embeds = []
+        for input_id, speech_embed in zip(input_ids, speech_embeds):
+            left_embeds = embed_tokens(input_id[:start_length])
+            right_embeds = embed_tokens(input_id[start_length + speech_embed.shape[1]:])
+            concat_tensor = torch.concat([left_embeds, speech_embed[0], right_embeds], dim=0).contiguous()
+            inputs_embeds.append(concat_tensor)
+
+        inputs_embeds = torch.stack(inputs_embeds)
+
+        return self.llama_model.forward(
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            labels=labels,
+            use_cache=False,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict
+        )
+
+    def generate(
+            self,
+            wav_path, prompt, prompt_pattern="USER: <Speech><SpeechHere></Speech> {}\nASSISTANT:", device='cuda:0',
+            max_length=2048, num_beams=4, do_sample=True, min_length=1, top_p=0.9, top_k=50,
+            repetition_penalty=1.0, length_penalty=1.0, temperature=1.0, bdr=(0, 240), num_return_sequences=1
+    ):
+        # read wav
+        wav, sr = sf.read(wav_path)
+        if len(wav.shape) == 2:
+            wav = wav[:, 0]
+        wav = wav[int(bdr[0] * sr): int(bdr[1] * sr)]
+        if sr != 16000:
+            wav = librosa.resample(wav, orig_sr=sr, target_sr=16000, res_type="fft")
+
+        # whisper
+        spectrogram = self.feature_extractor(wav, return_tensors="pt", sampling_rate=16000).input_features.to(
+            device)  # [1, 80, 3000]
+        speech_embeds = self.speech_encoder(spectrogram, return_dict=True).last_hidden_state
+
+        # beats
+        raw_wav = torch.from_numpy(wav).to(device).unsqueeze(0)
+        audio_padding_mask = torch.zeros(raw_wav.shape, device=device).bool()
+        audio_embeds, _ = self.beats.extract_features(raw_wav, padding_mask=audio_padding_mask, feature_only=True)
+
+        # auditory embeds
+        speech_embeds = self.ln_speech(speech_embeds)
+        audio_embeds = self.ln_audio(audio_embeds)
+        audio_embeds = F.pad(audio_embeds, (0, 0, 0, speech_embeds.size(1) - audio_embeds.size(1)))
+        speech_embeds = torch.cat([speech_embeds, audio_embeds], dim=-1)
+
+        # split frames
+        B, T, C = speech_embeds.shape
+        kernel, stride = round(T * self.second_per_frame / 30.0), round(T * self.second_stride / 30.0)
+        kernel, stride = (1, kernel), (1, stride)
+        speech_embeds_tr = speech_embeds.transpose(1, 2).unsqueeze(2)
+        speech_embeds_overlap = F.unfold(speech_embeds_tr, kernel_size=kernel, dilation=1, padding=0, stride=stride)
+        _, _, L = speech_embeds_overlap.shape
+        speech_embeds_overlap = speech_embeds_overlap.view(B, -1, kernel[1], L)
+        speech_embeds_overlap = torch.permute(speech_embeds_overlap, [0, 3, 2, 1])
+        speech_embeds = speech_embeds_overlap.reshape(-1, kernel[1], C)
+        speech_atts = torch.ones(speech_embeds.size()[:-1], dtype=torch.long, device=speech_embeds.device)
+
+        # Qformer
+        query_tokens = self.speech_query_tokens.expand(speech_embeds.shape[0], -1, -1)
+        query_output = self.speech_Qformer.bert(
+            query_embeds=query_tokens,
+            encoder_hidden_states=speech_embeds,
+            encoder_attention_mask=speech_atts,
+            return_dict=True,
+        )
+        speech_embeds = self.speech_llama_proj(query_output.last_hidden_state)
+        speech_embeds = speech_embeds.view(B, -1, speech_embeds.size(2)).contiguous()
+
+        # USER: <Speech>speech_embeds<Speech> prompt\nASSISTANT:
+        embed_tokens = self.llama_model.model.model.embed_tokens if self.lora else self.llama_model.model.embed_tokens
+        prompt_left, prompts_right = prompt_pattern.format(prompt).split('<SpeechHere>')
+
+        prompt_left_ids = self.tokenizer(prompt_left, return_tensors="pt", add_special_tokens=False).to(
+            speech_embeds.device).input_ids
+        prompt_left_embeds = embed_tokens(prompt_left_ids)
+
+        prompt_right_ids = self.tokenizer(prompts_right, return_tensors="pt", add_special_tokens=False).to(
+            speech_embeds.device).input_ids
+        prompt_right_embeds = embed_tokens(prompt_right_ids)
+
+        bos_embeds = self.llama_model.model.embed_tokens(
+            torch.ones([1, 1], dtype=torch.long, device=device) * self.tokenizer.bos_token_id
+        ) if not self.lora else self.llama_model.model.model.embed_tokens(
+            torch.ones([1, 1], dtype=torch.long, device=device) * self.tokenizer.bos_token_id
+        )
+
+        embeds = torch.cat([bos_embeds, prompt_left_embeds, speech_embeds, prompt_right_embeds], dim=1)
+        atts = torch.ones(embeds.size()[:-1], dtype=torch.long).to(embeds.device)
+
+        # generate
+        output = self.llama_model.generate(
+            inputs_embeds=embeds,
+            max_length=max_length,
+            num_beams=num_beams,
+            do_sample=do_sample,
+            min_length=min_length,
+            top_p=top_p,
+            top_k=top_k,
+            repetition_penalty=repetition_penalty,
+            length_penalty=length_penalty,
+            temperature=temperature,
+            num_return_sequences=num_return_sequences,
+            attention_mask=atts,
+            bos_token_id=self.tokenizer.bos_token_id,
+            eos_token_id=self.tokenizer.eos_token_id,
+            pad_token_id=self.tokenizer.pad_token_id,
+            # use_cache=False
+        )
+
+        output_text = self.tokenizer.batch_decode(output, add_special_tokens=False, skip_special_tokens=True)
+
+        # output_text = self.tokenizer.batch_decode(output)
+        return output_text
+
+    def init_speech_Qformer(self, num_query_token, speech_width, num_hidden_layers=2):
+        encoder_config = BertConfig()
+        encoder_config.num_hidden_layers = num_hidden_layers
+        encoder_config.encoder_width = speech_width
+        encoder_config.add_cross_attention = True
+        encoder_config.cross_attention_freq = 1
+        encoder_config.query_length = num_query_token
+        Qformer = BertLMHeadModel(config=encoder_config)
+        query_tokens = nn.Parameter(
+            torch.zeros(1, num_query_token, encoder_config.hidden_size)
+        )
+        query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)
+        return Qformer, query_tokens

modeling_llama.py

@@ -0,0 +1,885 @@
+# coding=utf-8
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# 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.
+""" PyTorch LLaMA model."""
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from transformers.models.llama.configuration_llama import LlamaConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "LlamaConfig"
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(
+    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+class LlamaRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        LlamaRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+
+        # convert into half-precision if necessary
+        if self.weight.dtype in [torch.float16, torch.bfloat16]:
+            hidden_states = hidden_states.to(self.weight.dtype)
+
+        return self.weight * hidden_states
+
+
+class LlamaRotaryEmbedding(torch.nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
+        self.register_buffer("inv_freq", inv_freq)
+
+        # Build here to make `torch.jit.trace` work.
+        self.max_seq_len_cached = max_position_embeddings
+        t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
+        self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
+
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
+        if seq_len > self.max_seq_len_cached:
+            self.max_seq_len_cached = seq_len
+            t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
+            freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            # Different from paper, but it uses a different permutation in order to obtain the same calculation
+            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
+            self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
+            self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
+        return (
+            self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+            self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+        )
+
+
+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):
+    # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
+    cos = cos.squeeze(1).squeeze(0)  # [seq_len, dim]
+    sin = sin.squeeze(1).squeeze(0)  # [seq_len, dim]
+    cos = cos[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
+    sin = sin[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class LlamaMLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+    ):
+        super().__init__()
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.act_fn = ACT2FN[hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class LlamaAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: LlamaConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = config.max_position_embeddings
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self.rotary_emb = LlamaRotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+        # [bsz, nh, t, hd]
+
+        if past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+            attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class LlamaDecoderLayer(nn.Module):
+    def __init__(self, config: LlamaConfig):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = LlamaAttention(config=config)
+        self.mlp = LlamaMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+        )
+        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative 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.
+            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`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = 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,
+        )
+        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,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+LLAMA_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 ([`LlamaConfig`]):
+            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 LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class LlamaPreTrainedModel(PreTrainedModel):
+    config_class = LlamaConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LlamaDecoderLayer"]
+    _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
+
+    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_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, LlamaModel):
+            module.gradient_checkpointing = value
+
+
+LLAMA_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 `decoder_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 (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            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)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_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.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class LlamaModel(LlamaPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
+
+    Args:
+        config: LlamaConfig
+    """
+
+    def __init__(self, config: LlamaConfig):
+        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([LlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.gradient_checkpointing = False
+        # 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
+
+    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape,
+                inputs_embeds.dtype,
+                device=inputs_embeds.device,
+                past_key_values_length=past_key_values_length,
+            )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
+                inputs_embeds.device
+            )
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+
+        return combined_attention_mask
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: 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,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        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
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+
+        if past_key_values is not None:
+            past_key_values_length = past_key_values[0][0].shape[2]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            position_ids = position_ids.view(-1, seq_length).long()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        # embed positions
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
+            )
+        attention_mask = self._prepare_decoder_attention_mask(
+            attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
+        )
+
+        hidden_states = inputs_embeds
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, None)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    None,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            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,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class LlamaForCausalLM(LlamaPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = LlamaModel(config)
+
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, 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
+
+    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
+
+    @add_start_docstrings_to_model_forward(LLAMA_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,
+        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, 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]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, LlamaForCausalLM
+
+        >>> model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you consciours? 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 consciours? Can you talk to me?\nI'm not consciours, 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
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=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 = outputs[0]
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        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_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past
+
+
+@add_start_docstrings(
+    """
+    The LLaMa Model transformer with a sequence classification head on top (linear layer).
+
+    [`LlamaForSequenceClassification`] 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).
+    """,
+    LLAMA_START_DOCSTRING,
+)
+class LlamaForSequenceClassification(LlamaPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = LlamaModel(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(LLAMA_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: 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, 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:
+                sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).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:
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        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,
+        )

salmonn_trainer.py

@@ -0,0 +1,35 @@
+import os
+import torch
+
+from transformers import Trainer
+from typing import Optional
+
+LOG_INTERVAL = 5000
+
+def get_state(model):
+    trainable_state_dict = dict()
+    for name, param in model.state_dict().items():
+        try:
+            if model.get_parameter(name).requires_grad:
+                trainable_state_dict[name] = param
+        except:
+            trainable_state_dict[name] = param
+    return trainable_state_dict
+
+
+class SALMONNTrainer(Trainer):
+
+    def _save_checkpoint(self, model, trial, metrics=None):
+        from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
+        checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{(self.state.global_step // LOG_INTERVAL) * LOG_INTERVAL}"
+
+        run_dir = self._get_output_dir(trial=trial)
+        output_dir = os.path.join(run_dir, checkpoint_folder)
+        os.makedirs(output_dir, exist_ok=True)
+
+        # Only save Adapter
+        weight_to_save = get_state(self.model)
+        torch.save(weight_to_save, os.path.join(output_dir, f'salomnn_7b.bin'))
+
+    def _save(self, output_dir: Optional[str] = None, state_dict=None):
+        super(SALMONNTrainer, self)._save(output_dir, state_dict)

train.json

@@ -0,0 +1,27 @@
+{
+  "fp16": {
+    "enabled": "auto",
+    "loss_scale": 0,
+    "loss_scale_window": 1000,
+    "initial_scale_power": 16,
+    "hysteresis": 2,
+    "min_loss_scale": 1
+  },
+  "bf16": {
+    "enabled": "auto"
+  },
+  "train_micro_batch_size_per_gpu": "auto",
+  "train_batch_size": "auto",
+  "gradient_accumulation_steps": "auto",
+  "zero_optimization": {
+    "stage": 2,
+    "overlap_comm": true,
+    "contiguous_gradients": true,
+    "sub_group_size": 1e9,
+    "reduce_bucket_size": "auto"
+  },
+  "wandb": {
+    "enabled": true,
+    "project": "SALMONN"
+  }
+}

train_mem.py

@@ -0,0 +1,14 @@
+# Adopted from https://github.com/lm-sys/FastChat. Below is the original copyright:
+# Adopted from tatsu-lab@stanford_alpaca. Below is the original copyright:
+# Make it more memory efficient by monkey patching the LLaMA model with FlashAttn.
+
+# Need to call this before importing transformers.
+
+from llama_flash_attn_monkey_patch import replace_llama_attn_with_flash_attn
+
+replace_llama_attn_with_flash_attn()
+
+from trainer import train
+
+if __name__ == "__main__":
+    train()

trainer.py

@@ -0,0 +1,81 @@
+# Adopted from https://github.com/lm-sys/FastChat. Below is the original copyright:
+# Adopted from tatsu-lab@stanford_alpaca. Below is the original copyright:
+#    Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li
+#
+#    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.
+
+import os
+from dataclasses import dataclass, field
+import logging
+import pathlib
+from typing import Dict, Optional, Sequence, List
+import torch
+import transformers
+import sys
+
+from salmonn_trainer import SALMONNTrainer, get_state
+from dataset import make_supervised_data_module, DataArguments
+from model import SALMONN
+from utils import print_trainable_parameters
+
+import wandb
+
+
+@dataclass
+class ModelArguments:
+    ckpt_path: Optional[str] = field(default='./salmonn_7b_v0.pth')
+    whisper_path: Optional[str] = field(default='./whisper-large-v2')
+    beats_path: Optional[str] = field(default='./BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt')
+    vicuna_path: Optional[str] = field(default='./vicuna-7b-v1.5')
+    version: Optional[str] = field(default="v0")
+    device: Optional[str] = field(default='cuda')
+
+
+@dataclass
+class TrainingArguments(transformers.TrainingArguments):
+    output_dir: Optional[str] = field(default='./checkpoints/')
+    optim: str = field(default="adamw_torch")
+    bf16: bool = True
+    fp16: bool = False
+    lora_alpha: int = 32
+    model_max_length: int = 2048
+    use_cache: bool = False
+    gradient_checkpointing: bool = False
+
+
+def train():
+    parser = transformers.HfArgumentParser((ModelArguments, DataArguments, TrainingArguments))
+    model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+    compute_dtype = (torch.float16 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32))
+    wandb.init(project='SALMONN', name=training_args.run_name)
+
+    model = SALMONN(
+        model_args.ckpt_path, model_args.whisper_path, model_args.beats_path, model_args.vicuna_path,
+        lora_alpha=training_args.lora_alpha, compute_dtype=compute_dtype
+    ).cuda()
+    print_trainable_parameters(model, vb=0)
+
+    data_module = make_supervised_data_module(tokenizer=model.tokenizer, data_args=data_args)
+    trainer = SALMONNTrainer(model=model, tokenizer=model.tokenizer, args=training_args, **data_module)
+
+    trainer.train()
+
+    # Only save Adapter
+    weight_to_save = get_state(model.named_parameters())
+    torch.save(weight_to_save, os.path.join(training_args.output_dir, f'salomnn_7b.bin'))
+
+    trainer.save_state()
+
+
+if __name__ == "__main__":
+    train()

utils.py

@@ -0,0 +1,12 @@
+def print_trainable_parameters(model, vb=0):
+    trainable_params = 0
+    all_param = 0
+    for _, param in model.named_parameters():
+        all_param += param.numel()
+        if param.requires_grad:
+            trainable_params += param.numel()
+            if vb > 0:
+                print(_, param.requires_grad, param.numel())
+    print(
+        f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param:.2f}"
+    )