From dc07bba2365092e990ec821a5688d1b172fbb99e Mon Sep 17 00:00:00 2001 From: Yifan Yang Date: Wed, 30 Apr 2025 09:54:42 +0000 Subject: [PATCH 01/22] init fix --- egs/speech_llm/ASR_LLM/.gitignore | 1 + egs/speech_llm/ASR_LLM/whisper_llm_zh/requirements.txt | 7 +------ 2 files changed, 2 insertions(+), 6 deletions(-) create mode 100644 egs/speech_llm/ASR_LLM/.gitignore diff --git a/egs/speech_llm/ASR_LLM/.gitignore b/egs/speech_llm/ASR_LLM/.gitignore new file mode 100644 index 000000000..604f0f2cf --- /dev/null +++ b/egs/speech_llm/ASR_LLM/.gitignore @@ -0,0 +1 @@ +models diff --git a/egs/speech_llm/ASR_LLM/whisper_llm_zh/requirements.txt b/egs/speech_llm/ASR_LLM/whisper_llm_zh/requirements.txt index a07c7b157..20cc44269 100644 --- a/egs/speech_llm/ASR_LLM/whisper_llm_zh/requirements.txt +++ b/egs/speech_llm/ASR_LLM/whisper_llm_zh/requirements.txt @@ -1,11 +1,6 @@ -k2 -kaldialign -git+https://github.com/lhotse-speech/lhotse -sentencepiece -pypinyin -tensorboard librosa deepspeed transformers>=4.37.0 flash-attn peft +openai-whisper From 9c8c4314de05d2181fa3500f88b5f4a761e81e54 Mon Sep 17 00:00:00 2001 From: Yifan Yang Date: Wed, 7 May 2025 12:18:41 +0000 Subject: [PATCH 02/22] init zipformer_llm_zh --- .../zipformer_llm_zh/asr_datamodule.py | 1 + .../ASR_LLM/zipformer_llm_zh/decode.py | 608 +++++++++++++ .../zipformer_llm_zh/ds_config_zero1.json | 1 + .../ASR_LLM/zipformer_llm_zh/model.py | 285 ++++++ .../ASR_LLM/zipformer_llm_zh/multi_dataset.py | 1 + .../ASR_LLM/zipformer_llm_zh/train.py | 815 ++++++++++++++++++ 6 files changed, 1711 insertions(+) create mode 120000 egs/speech_llm/ASR_LLM/zipformer_llm_zh/asr_datamodule.py create mode 100755 egs/speech_llm/ASR_LLM/zipformer_llm_zh/decode.py create mode 120000 egs/speech_llm/ASR_LLM/zipformer_llm_zh/ds_config_zero1.json create mode 100644 egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py create mode 120000 egs/speech_llm/ASR_LLM/zipformer_llm_zh/multi_dataset.py create mode 100755 egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/asr_datamodule.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/asr_datamodule.py new file mode 120000 index 000000000..1da0242af --- /dev/null +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/asr_datamodule.py @@ -0,0 +1 @@ +../whisper_llm_zh/asr_datamodule.py \ No newline at end of file diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/decode.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/decode.py new file mode 100755 index 000000000..3036b471e --- /dev/null +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/decode.py @@ -0,0 +1,608 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, +# Fangjun Kuang, +# Wei Kang) +# 2024 Yuekai Zhang +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# 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. +""" +Usage: +# Command for decoding using fine-tuned models: + +pip install huggingface_hub['cli'] +mkdir -p models/whisper models/qwen models/checkpoint +huggingface-cli download --local-dir models/checkpoint yuekai/icefall_asr_aishell_whisper_qwen2_1.5B + +# For aishell fine-tuned whisper model +huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_aishell_whisper exp_large_v2/whisper-large-v2-aishell1-epoch-10-avg-6.pt +# For multi-hans fine-tuned whisper model +# huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_multi-hans-zh_whisper v1.1/whisper-large-v2-multi-hans-zh-epoch-3-avg-10.pt + +huggingface-clie download --local-dir models/qwen Qwen/Qwen2-7B-Instruct + +mkdir -p whisper_llm_zh/exp_aishell_whisper_qwen2_1.5B +ln -s models/checkpoint/epoch-10-avg-5.pt whisper_llm_zh/exp_aishell_whisper_qwen2_1.5B/epoch-999.pt + +python3 ./whisper_llm_zh/decode.py \ + --max-duration 80 \ + --exp-dir whisper_llm_zh/exp_aishell_whisper_qwen2_1.5B \ + --speech-encoder-path-or-name models/whisper/exp_large_v2/whisper-large-v2-aishell1-epoch-10-avg-6.pt \ + --llm-path-or-name models/qwen \ + --epoch 999 --avg 1 \ + --manifest-dir data/fbank \ + --use-flash-attn True \ + --use-lora True --dataset aishell +""" + +import argparse +import logging +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import torch +import torch.nn as nn +import transformers +import whisper +from asr_datamodule import AsrDataModule +from lhotse.cut import Cut +from model import SPEECH_LLM, EncoderProjector +from multi_dataset import MultiDataset +from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training +from train import DEFAULT_SPEECH_TOKEN +from transformers import AutoModelForCausalLM, AutoTokenizer +from whisper_encoder_forward_monkey_patch import replace_whisper_encoder_forward + +from icefall.checkpoint import load_checkpoint +from icefall.env import get_env_info +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def average_checkpoints( + filenames: List[Path], device: torch.device = torch.device("cpu") +) -> dict: + """Average a list of checkpoints. + The function is mainly used for deepspeed converted checkpoint averaging, which only include model state_dict. + + Args: + filenames: + Filenames of the checkpoints to be averaged. We assume all + checkpoints are saved by :func:`save_checkpoint`. + device: + Move checkpoints to this device before averaging. + Returns: + Return a dict (i.e., state_dict) which is the average of all + model state dicts contained in the checkpoints. + """ + n = len(filenames) + + if "model" in torch.load(filenames[0], map_location=device): + avg = torch.load(filenames[0], map_location=device)["model"] + else: + avg = torch.load(filenames[0], map_location=device) + + # Identify shared parameters. Two parameters are said to be shared + # if they have the same data_ptr + uniqued: Dict[int, str] = dict() + + for k, v in avg.items(): + v_data_ptr = v.data_ptr() + if v_data_ptr in uniqued: + continue + uniqued[v_data_ptr] = k + + uniqued_names = list(uniqued.values()) + + for i in range(1, n): + if "model" in torch.load(filenames[i], map_location=device): + state_dict = torch.load(filenames[i], map_location=device)["model"] + else: + state_dict = torch.load(filenames[i], map_location=device) + for k in uniqued_names: + avg[k] += state_dict[k] + + for k in uniqued_names: + if avg[k].is_floating_point(): + avg[k] /= n + else: + avg[k] //= n + + return avg + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--llm-path-or-name", + type=str, + default="/workspace/asr/Qwen1.5-0.5B-Chat", + help="Path or name of the large language model.", + ) + + parser.add_argument( + "--speech-encoder-path-or-name", + type=str, + default="whisper-large-v2", + help="Path or name of the speech encoder.", + ) + + parser.add_argument( + "--encoder-projector-ds-rate", + type=int, + default=8, + help="Downsample rate for the encoder projector.", + ) + + parser.add_argument( + "--use-flash-attn", + type=str2bool, + default=True, + help="Whether to use flash attention.", + ) + + parser.add_argument( + "--use-lora", + type=str2bool, + default=True, + help="Whether to use lora fine-tuned llm checkpoint.", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=-1, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + parser.add_argument( + "--avg", + type=int, + default=1, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--method", + type=str, + default="beam-search", + help="""Decoding method. + Supported values are: + - beam-search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=1, + help="beam size for beam search decoding", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="whisper/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--remove-whisper-encoder-input-length-restriction", + type=str2bool, + default=True, + help="replace whisper encoder forward method to remove input length restriction", + ) + + parser.add_argument( + "--dataset", + type=str, + default="aishell", + choices=["aishell", "speechio", "wenetspeech_test_meeting", "multi_hans_zh"], + help="The dataset to decode", + ) + + add_model_arguments(parser) + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + "env_info": get_env_info(), + } + ) + return params + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + tokenizer: AutoTokenizer, + batch: dict, +) -> Dict[str, List[List[int]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: "beam-search" + - value: A list of lists. Each sublist is a list of token IDs. + Args: + params: + It is returned by :func:`get_params`. + model: + The neural model. + batch: + It is returned by :meth:`torch.utils.data.DataLoader.__iter__`. + Returns: + Return a dict, whose key may be "beam-search". + """ + + def preprocess( + messages, + tokenizer: transformers.PreTrainedTokenizer, + max_len: int = 128, + ) -> Dict: + """Preprocesses the data for supervised fine-tuning.""" + texts = [] + TEMPLATE = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content']}}{% if loop.last %}{{''}}{% else %}{{ '<|im_end|>\n' }}{% endif %}{% endfor %}" + for i, msg in enumerate(messages): + texts.append( + tokenizer.apply_chat_template( + msg, + tokenize=True, + add_generation_prompt=False, + chat_template=TEMPLATE, + padding="longest", + max_length=max_len, + truncation=True, + ) + ) + max_len_texts = max([len(text) for text in texts]) + if tokenizer.padding_side == "right": + texts = [ + text + [tokenizer.pad_token_id] * (max_len_texts - len(text)) + for text in texts + ] + else: + texts = [ + [tokenizer.pad_token_id] * (max_len_texts - len(text)) + text + for text in texts + ] + + input_ids = torch.tensor(texts, dtype=torch.int) + + attention_mask = input_ids.ne(tokenizer.pad_token_id) + + return input_ids, attention_mask + + dtype = torch.float32 + device = model.llm.device + + feature = batch["inputs"] + assert feature.ndim == 3 + feature = feature.to(device, dtype=dtype).transpose(1, 2) + if not params.remove_whisper_encoder_input_length_restriction: + T = 3000 + if feature.shape[2] < T: + feature = torch.cat( + [ + feature, + torch.zeros( + feature.shape[0], feature.shape[1], T - feature.shape[2] + ).to(device, dtype=dtype), + ], + 2, + ) + + supervisions = batch["supervisions"] + feature_len = supervisions["num_frames"] + feature_len = feature_len.to(device, dtype=dtype) + + messages = [ + [ + {"role": "user", "content": f"{DEFAULT_SPEECH_TOKEN}请转写音频为文字"}, + {"role": "assistant", "content": ""}, + ] + ] * len(feature) + + input_ids, attention_mask = preprocess(messages, tokenizer, max_len=128) + + generated_ids = model.decode( + feature, input_ids.to(device, dtype=torch.long), attention_mask.to(device) + ) + hyps = tokenizer.batch_decode(generated_ids, skip_special_tokens=True) + + return {"beam-search": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + tokenizer: AutoTokenizer, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + The dataloader. + params: + It is returned by :func:`get_params`. + model: + The neural model. + Returns: + Return a dict, whose key may be "beam-search". + """ + results = [] + + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + texts = [list("".join(text.split())) for text in texts] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + batch=batch, + tokenizer=tokenizer, + ) + + for lm_scale, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_text, ref_text in zip(cut_ids, hyps, texts): + this_batch.append((cut_id, ref_text, hyp_text)) + + results[lm_scale].extend(this_batch) + + num_cuts += len(batch["supervisions"]["text"]) + + if batch_idx % 100 == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[int], List[int]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results, char_level=True) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out CERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, + f"{test_set_name}-{key}", + results, + enable_log=True, + compute_CER=True, + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tCER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, CER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + params.res_dir = params.exp_dir / f"{params.method}" + + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + setup_logger( + params.res_dir + / f"log-decode-{params.method}-beam{params.beam_size}-{params.suffix}" + ) + + logging.info("Decoding started") + logging.info(params) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda") + + logging.info(f"device: {device}") + + if params.remove_whisper_encoder_input_length_restriction: + replace_whisper_encoder_forward() + + whisper_model = whisper.load_model(params.speech_encoder_path_or_name, "cpu") + speech_encoder = whisper_model.encoder + speech_encoder_dim = whisper_model.dims.n_audio_state + tokenizer = AutoTokenizer.from_pretrained(params.llm_path_or_name) + + if params.use_flash_attn: + attn_implementation = "flash_attention_2" + # torch_dtype=torch.bfloat16 FIX ME + torch_dtype = torch.float16 + tokenizer.padding_side = "left" + + else: + attn_implementation = "eager" + torch_dtype = torch.float16 + tokenizer.padding_side = "right" + + llm = AutoModelForCausalLM.from_pretrained( + params.llm_path_or_name, + attn_implementation=attn_implementation, + torch_dtype=torch_dtype, + ) + if params.use_lora: + lora_config = LoraConfig( + r=64, + lora_alpha=16, + target_modules=[ + "q_proj", + "k_proj", + "v_proj", + "o_proj", + "up_proj", + "gate_proj", + "down_proj", + ], + task_type="CAUSAL_LM", + ) + llm = get_peft_model(llm, lora_config) + llm.print_trainable_parameters() + + special_tokens_dict = {"additional_special_tokens": [DEFAULT_SPEECH_TOKEN]} + tokenizer.add_special_tokens(special_tokens_dict) + llm.config.pad_token_id = tokenizer.convert_tokens_to_ids("<|endoftext|>") + llm.config.bos_token_id = tokenizer.convert_tokens_to_ids("<|im_start|>") + llm.config.eos_token_id = tokenizer.convert_tokens_to_ids("<|im_end|>") + + llm.config.default_speech_token_id = tokenizer.convert_tokens_to_ids( + DEFAULT_SPEECH_TOKEN + ) + + encoder_projector = EncoderProjector( + speech_encoder_dim, llm.config.hidden_size, params.encoder_projector_ds_rate + ) + + model = SPEECH_LLM( + speech_encoder, + llm, + encoder_projector, + ) + + if params.avg > 1: + start = params.epoch - params.avg + 1 + assert start >= 1, start + # deepspeed converted checkpoint only contains model state_dict + filenames = [ + f"{params.exp_dir}/epoch-{epoch}/pytorch_model.bin" + for epoch in range(start, params.epoch + 1) + ] + avg_checkpoint = average_checkpoints(filenames) + model.load_state_dict(avg_checkpoint, strict=False) + + # filename = f"{params.exp_dir}/epoch-{params.epoch}-avg-{params.avg}.pt" + # torch.save(avg_checkpoint, filename) + else: + checkpoint = torch.load( + f"{params.exp_dir}/epoch-{params.epoch}/pytorch_model.bin", + map_location="cpu", + ) + model.load_state_dict(checkpoint, strict=False) + + model.to(device) + model.eval() + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + + data_module = AsrDataModule(args) + multi_dataset = MultiDataset(args.manifest_dir) + + def remove_long_utt(c: Cut): + # Keep only utterances with duration in 30 seconds + # + if c.duration > 30.0: + logging.warning( + f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + ) + return False + return True + + if params.dataset == "aishell": + test_sets_cuts = multi_dataset.aishell_test_cuts() + elif params.dataset == "speechio": + test_sets_cuts = multi_dataset.speechio_test_cuts() + elif params.dataset == "wenetspeech_test_meeting": + test_sets_cuts = multi_dataset.wenetspeech_test_meeting_cuts() + else: + test_sets_cuts = multi_dataset.test_cuts() + + test_sets = test_sets_cuts.keys() + test_dls = [ + data_module.test_dataloaders(test_sets_cuts[cuts_name].filter(remove_long_utt)) + for cuts_name in test_sets + ] + + for test_set, test_dl in zip(test_sets, test_dls): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + tokenizer=tokenizer, + ) + + save_results(params=params, test_set_name=test_set, results_dict=results_dict) + + logging.info("Done!") + + +if __name__ == "__main__": + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + main() diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/ds_config_zero1.json b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/ds_config_zero1.json new file mode 120000 index 000000000..804a23d79 --- /dev/null +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/ds_config_zero1.json @@ -0,0 +1 @@ +../whisper_llm_zh/ds_config_zero1.json \ No newline at end of file diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py new file mode 100644 index 000000000..829ef4e2d --- /dev/null +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py @@ -0,0 +1,285 @@ +import torch +from torch import nn +from transformers.trainer_pt_utils import LabelSmoother + +IGNORE_TOKEN_ID = LabelSmoother.ignore_index + + +class EncoderProjector(nn.Module): + """ + The encoder projector module. It is used to project the encoder outputs to the same dimension as the language model. + Modified from https://github.com/X-LANCE/SLAM-LLM/blob/main/src/slam_llm/models/projector.py. + Args: + encoder_dim (:obj:`int`): The dimension of the encoder outputs. + llm_dim (:obj:`int`): The dimension of the language model. + downsample_rate (:obj:`int`, `optional`, defaults to 5): The downsample rate to use. + """ + + def __init__(self, encoder_dim, llm_dim, downsample_rate=5): + super().__init__() + self.downsample_rate = downsample_rate + self.linear1 = nn.Linear(encoder_dim * self.downsample_rate, llm_dim) + self.relu = nn.ReLU() + self.linear2 = nn.Linear(llm_dim, llm_dim) + + def forward(self, x): + + batch_size, seq_len, feat_dim = x.size() + num_frames_to_discard = seq_len % self.downsample_rate + if num_frames_to_discard > 0: + x = x[:, :-num_frames_to_discard, :] + seq_len = x.size(1) + + x = x.contiguous() + x = x.view( + batch_size, seq_len // self.downsample_rate, feat_dim * self.downsample_rate + ) + + x = self.linear1(x) + x = self.relu(x) + x = self.linear2(x) + return x + + +class SPEECH_LLM(nn.Module): + """ + The Speech-to-Text model. It consists of an encoder, a language model and an encoder projector. + The encoder is used to extract speech features from the input speech signal. + The encoder projector is used to project the encoder outputs to the same dimension as the language model. + The language model is used to generate the text from the speech features. + Args: + encoder (:obj:`nn.Module`): The encoder module. + llm (:obj:`nn.Module`): The language model module. + encoder_projector (:obj:`nn.Module`): The encoder projector module. + """ + + def __init__( + self, + encoder: nn.Module, + llm: nn.Module, + encoder_projector: nn.Module, + ): + super().__init__() + self.encoder = encoder + self.llm = llm + self.encoder_projector = encoder_projector + + def _merge_input_ids_with_speech_features( + self, speech_features, inputs_embeds, input_ids, attention_mask, labels=None + ): + """ + Merge the speech features with the input_ids and attention_mask. This is done by replacing the speech tokens + with the speech features and padding the input_ids to the maximum length of the speech features. + Modified from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava/modeling_llava.py#L277. + Args: + speech_features (:obj:`torch.Tensor`): The speech features to merge with the input_ids. + inputs_embeds (:obj:`torch.Tensor`): The embeddings of the input_ids. + input_ids (:obj:`torch.Tensor`): The input ids to merge. + attention_mask (:obj:`torch.Tensor`): The attention mask to merge. + labels (:obj:`torch.Tensor`, `optional`): The labels to merge. + Returns: + :obj:`Tuple(torch.Tensor)`: The merged embeddings, attention mask, labels and position ids. + """ + num_speechs, speech_len, embed_dim = speech_features.shape + batch_size, sequence_length = input_ids.shape + left_padding = not torch.sum( + input_ids[:, -1] == torch.tensor(self.llm.config.pad_token_id) + ) + # 1. Create a mask to know where special speech tokens are + special_speech_token_mask = input_ids == self.llm.config.default_speech_token_id + num_special_speech_tokens = torch.sum(special_speech_token_mask, dim=-1) + # Compute the maximum embed dimension + max_embed_dim = ( + num_special_speech_tokens.max() * (speech_len - 1) + ) + sequence_length + batch_indices, non_speech_indices = torch.where( + input_ids != self.llm.config.default_speech_token_id + ) + + # 2. Compute the positions where text should be written + # Calculate new positions for text tokens in merged speech-text sequence. + # `special_speech_token_mask` identifies speech tokens. Each speech token will be replaced by `nb_text_tokens_per_speechs - 1` text tokens. + # `torch.cumsum` computes how each speech token shifts subsequent text token positions. + # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one. + new_token_positions = ( + torch.cumsum((special_speech_token_mask * (speech_len - 1) + 1), -1) - 1 + ) + nb_speech_pad = max_embed_dim - 1 - new_token_positions[:, -1] + if left_padding: + new_token_positions += nb_speech_pad[:, None] # offset for left padding + text_to_overwrite = new_token_positions[batch_indices, non_speech_indices] + + # 3. Create the full embedding, already padded to the maximum position + final_embedding = torch.zeros( + batch_size, + max_embed_dim, + embed_dim, + dtype=inputs_embeds.dtype, + device=inputs_embeds.device, + ) + final_attention_mask = torch.zeros( + batch_size, + max_embed_dim, + dtype=attention_mask.dtype, + device=inputs_embeds.device, + ) + if labels is not None: + final_labels = torch.full( + (batch_size, max_embed_dim), + IGNORE_TOKEN_ID, + dtype=input_ids.dtype, + device=input_ids.device, + ) + # In case the Vision model or the Language model has been offloaded to CPU, we need to manually + # set the corresponding tensors into their correct target device. + target_device = inputs_embeds.device + batch_indices, non_speech_indices, text_to_overwrite = ( + batch_indices.to(target_device), + non_speech_indices.to(target_device), + text_to_overwrite.to(target_device), + ) + attention_mask = attention_mask.to(target_device) + + # 4. Fill the embeddings based on the mask. If we have ["hey" "", "how", "are"] + # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the speech features + final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[ + batch_indices, non_speech_indices + ] + final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[ + batch_indices, non_speech_indices + ] + if labels is not None: + final_labels[batch_indices, text_to_overwrite] = labels[ + batch_indices, non_speech_indices + ] + + # 5. Fill the embeddings corresponding to the speechs. Anything that is not `text_positions` needs filling (#29835) + speech_to_overwrite = torch.full( + (batch_size, max_embed_dim), + True, + dtype=torch.bool, + device=inputs_embeds.device, + ) + speech_to_overwrite[batch_indices, text_to_overwrite] = False + speech_to_overwrite &= speech_to_overwrite.cumsum(-1) - 1 >= nb_speech_pad[ + :, None + ].to(target_device) + + if speech_to_overwrite.sum() != speech_features.shape[:-1].numel(): + raise ValueError( + f"The input provided to the model are wrong. The number of speech tokens is {torch.sum(special_speech_token_mask)} while" + f" the number of speech given to the model is {num_speechs}. This prevents correct indexing and breaks batch generation." + ) + + final_embedding[speech_to_overwrite] = ( + speech_features.contiguous().reshape(-1, embed_dim).to(target_device) + ) + final_attention_mask |= speech_to_overwrite + position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_( + (final_attention_mask == 0), 1 + ) + + # 6. Mask out the embedding at padding positions, as we later use the past_key_value value to determine the non-attended tokens. + batch_indices, pad_indices = torch.where( + input_ids == self.llm.config.pad_token_id + ) + indices_to_mask = new_token_positions[batch_indices, pad_indices] + + final_embedding[batch_indices, indices_to_mask] = 0 + + if labels is None: + final_labels = None + + return final_embedding, final_attention_mask, final_labels, position_ids + + def forward( + self, + fbank: torch.Tensor = None, + input_ids: torch.LongTensor = None, + attention_mask: torch.Tensor = None, + labels: torch.LongTensor = None, + ): + encoder_outs = self.encoder(fbank) + + speech_features = self.encoder_projector(encoder_outs) + + inputs_embeds = self.llm.get_input_embeddings()(input_ids) + + ( + inputs_embeds, + attention_mask, + labels, + _, + ) = self._merge_input_ids_with_speech_features( + speech_features, inputs_embeds, input_ids, attention_mask, labels + ) + + model_outputs = self.llm( + inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels + ) + + with torch.no_grad(): + preds = torch.argmax(model_outputs.logits, -1) + acc = compute_accuracy( + preds.detach()[:, :-1], + labels.detach()[:, 1:], + ignore_label=IGNORE_TOKEN_ID, + ) + return model_outputs, acc + + def decode( + self, + fbank: torch.Tensor = None, + input_ids: torch.LongTensor = None, + attention_mask: torch.Tensor = None, + **kwargs, + ): + + encoder_outs = self.encoder(fbank) + speech_features = self.encoder_projector(encoder_outs) + speech_features = speech_features.to(torch.float16) + inputs_embeds = self.llm.get_input_embeddings()(input_ids) + ( + inputs_embeds, + attention_mask, + _, + position_ids, + ) = self._merge_input_ids_with_speech_features( + speech_features, inputs_embeds, input_ids, attention_mask + ) + generated_ids = self.llm.generate( + inputs_embeds=inputs_embeds, + max_new_tokens=kwargs.get("max_new_tokens", 200), + num_beams=kwargs.get("num_beams", 1), + do_sample=kwargs.get("do_sample", False), + min_length=kwargs.get("min_length", 1), + top_p=kwargs.get("top_p", 1.0), + repetition_penalty=kwargs.get("repetition_penalty", 1.0), + length_penalty=kwargs.get("length_penalty", 1.0), + temperature=kwargs.get("temperature", 1.0), + bos_token_id=self.llm.config.bos_token_id, + eos_token_id=self.llm.config.eos_token_id, + pad_token_id=self.llm.config.pad_token_id, + ) + + return generated_ids + + +def compute_accuracy(pad_outputs, pad_targets, ignore_label): + """Calculate accuracy. + Copied from https://github.com/X-LANCE/SLAM-LLM/blob/main/src/slam_llm/utils/metric.py + Args: + pad_outputs (LongTensor): Prediction tensors (B, Lmax). + pad_targets (LongTensor): Target label tensors (B, Lmax). + ignore_label (int): Ignore label id. + + Returns: + float: Accuracy value (0.0 - 1.0). + + """ + mask = pad_targets != ignore_label + numerator = torch.sum( + pad_outputs.masked_select(mask) == pad_targets.masked_select(mask) + ) + denominator = torch.sum(mask) + return numerator.float() / denominator.float() diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/multi_dataset.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/multi_dataset.py new file mode 120000 index 000000000..eb04c96d2 --- /dev/null +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/multi_dataset.py @@ -0,0 +1 @@ +../whisper_llm_zh/multi_dataset.py \ No newline at end of file diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py new file mode 100755 index 000000000..7947a60a5 --- /dev/null +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py @@ -0,0 +1,815 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Xiaoyu Yang) +# 2024 Yuekai Zhang +# 2025 Yifan Yang +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# 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. +""" +Usage: +# fine-tuning with whisper and Qwen2 +pip install huggingface_hub['cli'] +mkdir -p models/whisper models/qwen + +# For aishell fine-tuned whisper model +huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_aishell_whisper exp_large_v2/whisper-large-v2-aishell1-epoch-10-avg-6.pt +# For multi-hans fine-tuned whisper model +# huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_multi-hans-zh_whisper v1.1/whisper-large-v2-multi-hans-zh-epoch-3-avg-10.pt + +# huggingface-clie download --local-dir models/qwen Qwen/Qwen2-7B-Instruct +huggingface-clie download --local-dir models/qwen Qwen/Qwen2-1.5B-Instruct + +torchrun --nproc_per_node 8 ./whisper_llm_zh/train.py \ + --max-duration 200 \ + --exp-dir ./whisper_llm_zh/exp_test \ + --speech-encoder-path-or-name models/whisper/exp_large_v2/whisper-large-v2-aishell1-epoch-10-avg-6.pt \ + --llm-path-or-name Qwen/Qwen2-1.5B-Instruct \ + --manifest-dir data/fbank \ + --deepspeed \ + --deepspeed_config ./whisper_llm_zh/ds_config_zero1.json \ + --use-flash-attn True \ + --use-lora True --unfreeze-llm True +""" + +import argparse +import logging +import os +import warnings +from pathlib import Path +from typing import Dict, Optional, Tuple + +import deepspeed +import torch +import torch.nn as nn +import transformers +import whisper +from asr_datamodule import AsrDataModule +from deepspeed.utils.zero_to_fp32 import convert_zero_checkpoint_to_fp32_state_dict +from lhotse.cut import Cut +from lhotse.utils import fix_random_seed +from model import IGNORE_TOKEN_ID, SPEECH_LLM, EncoderProjector +from multi_dataset import MultiDataset +from peft import LoraConfig, get_peft_model +from torch import Tensor +from torch.utils.tensorboard import SummaryWriter +from transformers import AutoModelForCausalLM, AutoTokenizer +from whisper_encoder_forward_monkey_patch import replace_whisper_encoder_forward + +from icefall.dist import get_rank, get_world_size +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +DEFAULT_SPEECH_TOKEN = "" + + +def set_batch_count(model: nn.Module, batch_count: float) -> None: + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--llm-path-or-name", + type=str, + default="/workspace/asr/Qwen1.5-0.5B-Chat", + help="Path or name of the large language model.", + ) + + parser.add_argument( + "--speech-encoder-path-or-name", + type=str, + default="whisper-large-v2", + help="Path or name of the speech encoder.", + ) + + parser.add_argument( + "--encoder-projector-ds-rate", + type=int, + default=8, + help="Downsample rate for the encoder projector.", + ) + parser.add_argument( + "--use-flash-attn", + type=str2bool, + default=True, + help="Whether to use flash attention.", + ) + + parser.add_argument( + "--use-lora", + type=str2bool, + default=False, + help="Whether to use lora to fine-tune llm.", + ) + + parser.add_argument( + "--unfreeze-llm", + type=str2bool, + default=False, + help="Whether to unfreeze llm during training.", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=10, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="whisper_qwen/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--pretrained-model-path", + type=str, + default=None, + help="""The path to the pretrained model if it is not None. Training will + start from this model. e.g. ./wenetspeech/ASR/whisper/exp_large_v2/epoch-4-avg-3.pt + """, + ) + + parser.add_argument( + "--sampler-state-dict-path", + type=str, + default=None, + help="""The path to the sampler state dict if it is not None. Training will start from this sampler state dict. + """, + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=True, + help="Whether to use half precision training.", + ) + + parser.add_argument( + "--use-aishell", + type=str2bool, + default=True, + help="Whether to only use aishell1 dataset for training.", + ) + + parser = deepspeed.add_config_arguments(parser) + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - frame_shift_ms: The frame shift in milliseconds. + - allowed_excess_duration_ratio: The allowed excess duration ratio. + - best_train_loss: The best training loss so far. + - best_valid_loss: The best validation loss so far. + - best_train_epoch: The epoch where the best training loss is achieved. + - best_valid_epoch: The epoch where the best validation loss is achieved. + - batch_idx_train: The batch index of the current batch. + - log_interval: Log training stats every `log_interval` batches. + - reset_interval: Reset the stats every `reset_interval` batches. + - valid_interval: Run validation every `valid_interval` batches. + - env_info: The environment information. + """ + params = AttributeDict( + { + "allowed_excess_duration_ratio": 0.1, + "subsampling_factor": 2, + "frame_shift_ms": 10, + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 5000, + "env_info": get_env_info(), + } + ) + + return params + + +def compute_loss( + params: AttributeDict, + tokenizer: AutoTokenizer, + model: nn.Module, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute the loss for the given batch. + Args: + params: + It is returned by :func:`get_params`. + tokenizer: + The tokenizer used to encode the text. + model: + The model for training. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + Whether it is training. + Returns: + Return a tuple of two elements. The first element is the loss tensor. + """ + + def preprocess( + messages, + tokenizer: transformers.PreTrainedTokenizer, + max_len: int, + ) -> Dict: + """Preprocesses the data for supervised fine-tuning.""" + texts = [] + TEMPLATE = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content']}}{% if loop.last %}{{ '<|im_end|>'}}{% else %}{{ '<|im_end|>\n' }}{% endif %}{% endfor %}" + for i, msg in enumerate(messages): + texts.append( + tokenizer.apply_chat_template( + msg, + tokenize=True, + chat_template=TEMPLATE, + add_generation_prompt=False, + padding="longest", # FIX me change padding to longest + max_length=max_len, + truncation=True, + ) + ) + # padding texts to the same length, texts is a list of list, padding with tokenzier.pad_token_id + max_len_texts = max([len(text) for text in texts]) + if tokenizer.padding_side == "right": + texts = [ + text + [tokenizer.pad_token_id] * (max_len_texts - len(text)) + for text in texts + ] + else: + texts = [ + [tokenizer.pad_token_id] * (max_len_texts - len(text)) + text + for text in texts + ] + input_ids = torch.tensor(texts, dtype=torch.int) + # response = tokenizer.batch_decode(input_ids, skip_special_tokens=True)[0] + target_ids = input_ids.clone() + target_ids[target_ids == tokenizer.pad_token_id] = IGNORE_TOKEN_ID + # mask all tokens before token_id 151646 with IGNORE_TOKEN_ID + # first get the indices of the tokens + mask_prompt = True + if mask_prompt: + mask_indices = torch.where( + input_ids == tokenizer.convert_tokens_to_ids("assistant") + ) + for i in range(mask_indices[0].size(0)): + row = mask_indices[0][i] + col = mask_indices[1][i] + # + 2 to skip: 'assistant', '\n' + target_ids[row, : col + 2] = IGNORE_TOKEN_ID + + attention_mask = input_ids.ne(tokenizer.pad_token_id) + + return input_ids, attention_mask, target_ids + + device = next(model.parameters()).device + feature = batch["inputs"] + + assert feature.ndim == 3 + feature = feature.to(device) + feature = feature.transpose(1, 2) # (N, C, T) + + batch_idx_train = params.batch_idx_train + supervisions = batch["supervisions"] + texts = batch["supervisions"]["text"] + + messages = [] + for i, text in enumerate(texts): + text = text.replace(" ", "") + message = [ + {"role": "user", "content": f"{DEFAULT_SPEECH_TOKEN}请转写音频为文字"}, + {"role": "assistant", "content": text}, + ] + messages.append(message) + + input_ids, attention_mask, target_ids = preprocess(messages, tokenizer, max_len=128) + + target_ids = target_ids.type(torch.LongTensor) + input_ids = input_ids.type(torch.LongTensor) + + with torch.set_grad_enabled(is_training): + model_outputs, acc = model( + fbank=feature, + input_ids=input_ids.to(device), + attention_mask=attention_mask.to(device), + labels=target_ids.to(device), + ) + loss = model_outputs.loss + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + feature_lens = supervisions["num_frames"] + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["acc"] = ( + acc * info["frames"] + ) # WAR: to avoid normalization by the number of frames + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + tokenizer: whisper.tokenizer.Tokenizer, + model: nn.Module, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + tokenizer=tokenizer, + model=model, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + tokenizer: AutoTokenizer, + model: nn.Module, + optimizer: torch.optim.Optimizer, + scheduler: torch.optim.lr_scheduler, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + model.encoder.eval() + if not params.unfreeze_llm: + model.llm.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(train_dl): + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + if batch_idx % params.valid_interval == 0: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + tokenizer=tokenizer, + model=model, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + model.encoder.eval() + if not params.unfreeze_llm: + model.llm.eval() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + if batch_idx != 0: + model.save_checkpoint( + save_dir=params.exp_dir, + tag=f"epoch-{params.cur_epoch}-checkpoint-{batch_idx}", + client_state={}, + exclude_frozen_parameters=True, + ) + + if rank == 0: + convert_zero_checkpoint_to_fp32_state_dict( + params.exp_dir, + f"{params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}.pt", + tag=f"epoch-{params.cur_epoch}-checkpoint-{batch_idx}", + exclude_frozen_parameters=True, + ) + # save sampler state dict into checkpoint + sampler_state_dict = train_dl.sampler.state_dict() + torch.save( + sampler_state_dict, + f"{params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}-sampler.pt", + ) + os.system( + f"rm -rf {params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}" + ) + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + tokenizer=tokenizer, + model=model, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + + # deepspeed's backward() is different from torch's backward() + # in that it does not accept a loss tensor as input. + # It computes the loss internally. + model.backward(loss) + model.step() + + except: # noqa + display_and_save_batch(batch, params=params) + raise + + if batch_idx % params.log_interval == 0: + try: + cur_lr = scheduler.get_last_lr()[0] + except: # noqa + cur_lr = 0.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info(params) + + logging.info("About to create model") + + replace_whisper_encoder_forward() + whisper_model = whisper.load_model(params.speech_encoder_path_or_name, "cpu") + speech_encoder = whisper_model.encoder + speech_encoder_dim = whisper_model.dims.n_audio_state + for name, param in speech_encoder.named_parameters(): + param.requires_grad = False + + tokenizer = AutoTokenizer.from_pretrained(params.llm_path_or_name) + if params.use_flash_attn: + attn_implementation = "flash_attention_2" + # torch_dtype=torch.bfloat16 FIX ME + torch_dtype = torch.float16 + tokenizer.padding_side = "left" + + else: + attn_implementation = "eager" + torch_dtype = torch.float16 + tokenizer.padding_side = "right" + + llm = AutoModelForCausalLM.from_pretrained( + params.llm_path_or_name, + attn_implementation=attn_implementation, + torch_dtype=torch_dtype, + ) + + if not params.unfreeze_llm: + for name, param in llm.named_parameters(): + param.requires_grad = False + else: + if params.use_lora: + lora_config = LoraConfig( + r=64, + lora_alpha=16, + target_modules=[ + "q_proj", + "k_proj", + "v_proj", + "o_proj", + "up_proj", + "gate_proj", + "down_proj", + ], + lora_dropout=0.05, + task_type="CAUSAL_LM", + ) + llm = get_peft_model(llm, lora_config) + llm.print_trainable_parameters() + + special_tokens_dict = {"additional_special_tokens": [DEFAULT_SPEECH_TOKEN]} + tokenizer.add_special_tokens(special_tokens_dict) + llm.config.pad_token_id = tokenizer.pad_token_id + llm.config.default_speech_token_id = tokenizer.convert_tokens_to_ids( + DEFAULT_SPEECH_TOKEN + ) + + encoder_projector = EncoderProjector( + speech_encoder_dim, llm.config.hidden_size, params.encoder_projector_ds_rate + ) + + model = SPEECH_LLM( + speech_encoder, + llm, + encoder_projector, + ) + + if params.pretrained_model_path: + checkpoint = torch.load(params.pretrained_model_path, map_location="cpu") + missing_keys, unexpected_keys = model.load_state_dict(checkpoint, strict=False) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + logging.info("Trainable parameters (excluding model.eval modules):") + for name, param in model.named_parameters(): + if param.requires_grad: + logging.info(f"{name}: {param.shape}") + + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + else: + device = torch.device("cpu") + logging.info(f"Device: {device}") + model.to(device) + + assert params.deepspeed + logging.info("Using DeepSpeed") + model, optimizer, _, scheduler = deepspeed.initialize( + args=params, model=model, model_parameters=model.parameters() + ) + + data_module = AsrDataModule(args) + multi_dataset = MultiDataset(args.manifest_dir) + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 20.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + return True + + if params.use_aishell: + train_cuts = multi_dataset.aishell_train_cuts() + else: + train_cuts = multi_dataset.train_cuts() + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + sampler_state_dict = None + if params.sampler_state_dict_path: + sampler_state_dict = torch.load(params.sampler_state_dict_path) + sampler_state_dict["max_duration"] = params.max_duration + + train_dl = data_module.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + if params.use_aishell: + valid_cuts = multi_dataset.aishell_dev_cuts() + else: + valid_cuts = multi_dataset.dev_cuts() + valid_dl = data_module.valid_dataloaders(valid_cuts) + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + logging.info(f"start training from epoch {params.start_epoch}") + for epoch in range(params.start_epoch, params.num_epochs + 1): + + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + tokenizer=tokenizer, + model=model, + optimizer=optimizer, + scheduler=scheduler, + train_dl=train_dl, + valid_dl=valid_dl, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + model.save_checkpoint( + save_dir=params.exp_dir, + tag=f"zero-epoch-{params.cur_epoch}", + client_state={}, + exclude_frozen_parameters=True, + ) + if rank == 0: + convert_zero_checkpoint_to_fp32_state_dict( + params.exp_dir, + f"{params.exp_dir}/epoch-{params.cur_epoch}", + tag=f"zero-epoch-{params.cur_epoch}", + exclude_frozen_parameters=True, + ) + # save sampler state dict into checkpoint + sampler_state_dict = train_dl.sampler.state_dict() + torch.save( + sampler_state_dict, + f"{params.exp_dir}/epoch-{params.cur_epoch}-sampler.pt", + ) + + os.system(f"rm -rf {params.exp_dir}/zero-epoch-{params.cur_epoch}") + + logging.info("Done!") + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + +def main(): + parser = get_parser() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = get_world_size() + rank = get_rank() + + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + warnings.filterwarnings("ignore", category=FutureWarning) + run(rank=rank, world_size=world_size, args=args) + + +if __name__ == "__main__": + main() From 23b5a7ce3e02e9f93a05704e13ab55b78e12768b Mon Sep 17 00:00:00 2001 From: Yifan Yang Date: Wed, 7 May 2025 12:29:12 +0000 Subject: [PATCH 03/22] format multi_dataset.py --- egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py b/egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py index eae967500..f821fd29d 100644 --- a/egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py +++ b/egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py @@ -15,13 +15,10 @@ # limitations under the License. -import glob import logging -import re from pathlib import Path -from typing import Dict, List +from typing import Dict -import lhotse from lhotse import CutSet, load_manifest_lazy From 211c01bc1dcb3a403aac69838eca8b2b480b03e1 Mon Sep 17 00:00:00 2001 From: Yifan Yang Date: Wed, 7 May 2025 12:37:19 +0000 Subject: [PATCH 04/22] format train.py minor fix train.py --- egs/speech_llm/ASR_LLM/whisper_llm_zh/train.py | 18 +++--------------- 1 file changed, 3 insertions(+), 15 deletions(-) diff --git a/egs/speech_llm/ASR_LLM/whisper_llm_zh/train.py b/egs/speech_llm/ASR_LLM/whisper_llm_zh/train.py index 7947a60a5..6e43bf83f 100755 --- a/egs/speech_llm/ASR_LLM/whisper_llm_zh/train.py +++ b/egs/speech_llm/ASR_LLM/whisper_llm_zh/train.py @@ -18,18 +18,6 @@ # limitations under the License. """ Usage: -# fine-tuning with whisper and Qwen2 -pip install huggingface_hub['cli'] -mkdir -p models/whisper models/qwen - -# For aishell fine-tuned whisper model -huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_aishell_whisper exp_large_v2/whisper-large-v2-aishell1-epoch-10-avg-6.pt -# For multi-hans fine-tuned whisper model -# huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_multi-hans-zh_whisper v1.1/whisper-large-v2-multi-hans-zh-epoch-3-avg-10.pt - -# huggingface-clie download --local-dir models/qwen Qwen/Qwen2-7B-Instruct -huggingface-clie download --local-dir models/qwen Qwen/Qwen2-1.5B-Instruct - torchrun --nproc_per_node 8 ./whisper_llm_zh/train.py \ --max-duration 200 \ --exp-dir ./whisper_llm_zh/exp_test \ @@ -39,7 +27,8 @@ torchrun --nproc_per_node 8 ./whisper_llm_zh/train.py \ --deepspeed \ --deepspeed_config ./whisper_llm_zh/ds_config_zero1.json \ --use-flash-attn True \ - --use-lora True --unfreeze-llm True + --use-lora True \ + --unfreeze-llm True """ import argparse @@ -333,7 +322,6 @@ def compute_loss( feature = feature.to(device) feature = feature.transpose(1, 2) # (N, C, T) - batch_idx_train = params.batch_idx_train supervisions = batch["supervisions"] texts = batch["supervisions"]["text"] @@ -378,7 +366,7 @@ def compute_loss( def compute_validation_loss( params: AttributeDict, - tokenizer: whisper.tokenizer.Tokenizer, + tokenizer: AutoTokenizer, model: nn.Module, valid_dl: torch.utils.data.DataLoader, world_size: int = 1, From 489c42b45eef034c4a9c84cdb2310079c0dfe656 Mon Sep 17 00:00:00 2001 From: Yifan Yang Date: Thu, 8 May 2025 04:31:34 +0000 Subject: [PATCH 05/22] support zipformer encoder update update update update fix reformat support infer update --- .../ASR_LLM/whisper_llm_zh/model.py | 8 +- .../ASR_LLM/zipformer_llm_zh/decode.py | 133 +++------- .../zipformer_llm_zh/encoder_interface.py | 1 + .../ASR_LLM/zipformer_llm_zh/model.py | 61 ++++- .../ASR_LLM/zipformer_llm_zh/scaling.py | 1 + .../ASR_LLM/zipformer_llm_zh/subsampling.py | 1 + .../ASR_LLM/zipformer_llm_zh/train.py | 248 +++++++++++++++--- .../ASR_LLM/zipformer_llm_zh/zipformer.py | 1 + 8 files changed, 304 insertions(+), 150 deletions(-) create mode 120000 egs/speech_llm/ASR_LLM/zipformer_llm_zh/encoder_interface.py create mode 120000 egs/speech_llm/ASR_LLM/zipformer_llm_zh/scaling.py create mode 120000 egs/speech_llm/ASR_LLM/zipformer_llm_zh/subsampling.py create mode 120000 egs/speech_llm/ASR_LLM/zipformer_llm_zh/zipformer.py diff --git a/egs/speech_llm/ASR_LLM/whisper_llm_zh/model.py b/egs/speech_llm/ASR_LLM/whisper_llm_zh/model.py index 829ef4e2d..bc6a94613 100644 --- a/egs/speech_llm/ASR_LLM/whisper_llm_zh/model.py +++ b/egs/speech_llm/ASR_LLM/whisper_llm_zh/model.py @@ -194,10 +194,10 @@ class SPEECH_LLM(nn.Module): def forward( self, - fbank: torch.Tensor = None, - input_ids: torch.LongTensor = None, - attention_mask: torch.Tensor = None, - labels: torch.LongTensor = None, + fbank: torch.Tensor, + input_ids: torch.LongTensor, + attention_mask: torch.Tensor, + labels: torch.LongTensor, ): encoder_outs = self.encoder(fbank) diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/decode.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/decode.py index 3036b471e..2b757c2de 100755 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/decode.py +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/decode.py @@ -3,6 +3,7 @@ # Fangjun Kuang, # Wei Kang) # 2024 Yuekai Zhang +# 2025 Yifan Yang # # See ../../../../LICENSE for clarification regarding multiple authors # @@ -19,31 +20,17 @@ # limitations under the License. """ Usage: -# Command for decoding using fine-tuned models: - -pip install huggingface_hub['cli'] -mkdir -p models/whisper models/qwen models/checkpoint -huggingface-cli download --local-dir models/checkpoint yuekai/icefall_asr_aishell_whisper_qwen2_1.5B - -# For aishell fine-tuned whisper model -huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_aishell_whisper exp_large_v2/whisper-large-v2-aishell1-epoch-10-avg-6.pt -# For multi-hans fine-tuned whisper model -# huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_multi-hans-zh_whisper v1.1/whisper-large-v2-multi-hans-zh-epoch-3-avg-10.pt - -huggingface-clie download --local-dir models/qwen Qwen/Qwen2-7B-Instruct - -mkdir -p whisper_llm_zh/exp_aishell_whisper_qwen2_1.5B -ln -s models/checkpoint/epoch-10-avg-5.pt whisper_llm_zh/exp_aishell_whisper_qwen2_1.5B/epoch-999.pt - -python3 ./whisper_llm_zh/decode.py \ +python3 ./zipformer_llm_zh/decode.py \ --max-duration 80 \ - --exp-dir whisper_llm_zh/exp_aishell_whisper_qwen2_1.5B \ - --speech-encoder-path-or-name models/whisper/exp_large_v2/whisper-large-v2-aishell1-epoch-10-avg-6.pt \ + --exp-dir zipformer_llm_zh/exp \ + --speech-encoder-path-or-name models/zipformer/epoch-999.pt \ --llm-path-or-name models/qwen \ - --epoch 999 --avg 1 \ + --epoch 999 \ + --avg 1 \ --manifest-dir data/fbank \ --use-flash-attn True \ - --use-lora True --dataset aishell + --use-lora True \ + --dataset aishell """ import argparse @@ -56,15 +43,22 @@ import k2 import torch import torch.nn as nn import transformers -import whisper from asr_datamodule import AsrDataModule from lhotse.cut import Cut from model import SPEECH_LLM, EncoderProjector from multi_dataset import MultiDataset from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training -from train import DEFAULT_SPEECH_TOKEN +from train import ( + DEFAULT_SPEECH_TOKEN, + _to_int_tuple, + add_model_arguments, + get_encoder_embed, + get_encoder_model, + get_params, + load_model_params, +) from transformers import AutoModelForCausalLM, AutoTokenizer -from whisper_encoder_forward_monkey_patch import replace_whisper_encoder_forward +from zipformer import Zipformer2 from icefall.checkpoint import load_checkpoint from icefall.env import get_env_info @@ -129,43 +123,6 @@ def average_checkpoints( return avg -def add_model_arguments(parser: argparse.ArgumentParser): - parser.add_argument( - "--llm-path-or-name", - type=str, - default="/workspace/asr/Qwen1.5-0.5B-Chat", - help="Path or name of the large language model.", - ) - - parser.add_argument( - "--speech-encoder-path-or-name", - type=str, - default="whisper-large-v2", - help="Path or name of the speech encoder.", - ) - - parser.add_argument( - "--encoder-projector-ds-rate", - type=int, - default=8, - help="Downsample rate for the encoder projector.", - ) - - parser.add_argument( - "--use-flash-attn", - type=str2bool, - default=True, - help="Whether to use flash attention.", - ) - - parser.add_argument( - "--use-lora", - type=str2bool, - default=True, - help="Whether to use lora fine-tuned llm checkpoint.", - ) - - def get_parser(): parser = argparse.ArgumentParser( formatter_class=argparse.ArgumentDefaultsHelpFormatter @@ -207,17 +164,10 @@ def get_parser(): parser.add_argument( "--exp-dir", type=str, - default="whisper/exp", + default="zipformer/exp", help="The experiment dir", ) - parser.add_argument( - "--remove-whisper-encoder-input-length-restriction", - type=str2bool, - default=True, - help="replace whisper encoder forward method to remove input length restriction", - ) - parser.add_argument( "--dataset", type=str, @@ -230,15 +180,6 @@ def get_parser(): return parser -def get_params() -> AttributeDict: - params = AttributeDict( - { - "env_info": get_env_info(), - } - ) - return params - - def decode_one_batch( params: AttributeDict, model: nn.Module, @@ -299,28 +240,13 @@ def decode_one_batch( return input_ids, attention_mask - dtype = torch.float32 device = model.llm.device feature = batch["inputs"] assert feature.ndim == 3 - feature = feature.to(device, dtype=dtype).transpose(1, 2) - if not params.remove_whisper_encoder_input_length_restriction: - T = 3000 - if feature.shape[2] < T: - feature = torch.cat( - [ - feature, - torch.zeros( - feature.shape[0], feature.shape[1], T - feature.shape[2] - ).to(device, dtype=dtype), - ], - 2, - ) supervisions = batch["supervisions"] - feature_len = supervisions["num_frames"] - feature_len = feature_len.to(device, dtype=dtype) + feature_lens = supervisions["num_frames"] messages = [ [ @@ -332,7 +258,10 @@ def decode_one_batch( input_ids, attention_mask = preprocess(messages, tokenizer, max_len=128) generated_ids = model.decode( - feature, input_ids.to(device, dtype=torch.long), attention_mask.to(device) + feature.to(device), + feature_lens.to(device), + input_ids.to(device, dtype=torch.long), + attention_mask.to(device), ) hyps = tokenizer.batch_decode(generated_ids, skip_special_tokens=True) @@ -471,12 +400,15 @@ def main(): logging.info(f"device: {device}") - if params.remove_whisper_encoder_input_length_restriction: - replace_whisper_encoder_forward() + speech_encoder_embed = get_encoder_embed(params) + speech_encoder = get_encoder_model(params) + load_model_params( + params.speech_encoder_path_or_name, speech_encoder_embed, "encoder_embed" + ) + load_model_params(params.speech_encoder_path_or_name, speech_encoder, "encoder") + + speech_encoder_dim = max(_to_int_tuple(params.encoder_dim)) - whisper_model = whisper.load_model(params.speech_encoder_path_or_name, "cpu") - speech_encoder = whisper_model.encoder - speech_encoder_dim = whisper_model.dims.n_audio_state tokenizer = AutoTokenizer.from_pretrained(params.llm_path_or_name) if params.use_flash_attn: @@ -528,6 +460,7 @@ def main(): ) model = SPEECH_LLM( + speech_encoder_embed, speech_encoder, llm, encoder_projector, diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/encoder_interface.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/encoder_interface.py new file mode 120000 index 000000000..c2eaca671 --- /dev/null +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/encoder_interface.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/encoder_interface.py \ No newline at end of file diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py index 829ef4e2d..5f0d4b8e5 100644 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py @@ -1,7 +1,12 @@ +from typing import Tuple + import torch +from encoder_interface import EncoderInterface from torch import nn from transformers.trainer_pt_utils import LabelSmoother +from icefall.utils import make_pad_mask + IGNORE_TOKEN_ID = LabelSmoother.ignore_index @@ -55,11 +60,13 @@ class SPEECH_LLM(nn.Module): def __init__( self, - encoder: nn.Module, + encoder_embed: nn.Module, + encoder: EncoderInterface, llm: nn.Module, encoder_projector: nn.Module, ): super().__init__() + self.encoder_embed = encoder_embed self.encoder = encoder self.llm = llm self.encoder_projector = encoder_projector @@ -192,14 +199,46 @@ class SPEECH_LLM(nn.Module): return final_embedding, final_attention_mask, final_labels, position_ids + def forward_encoder( + self, x: torch.Tensor, x_lens: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Compute encoder outputs. + Args: + x: + A 3-D tensor of shape (N, T, C). + x_lens: + A 1-D tensor of shape (N,). It contains the number of frames in `x` + before padding. + + Returns: + encoder_out: + Encoder output, of shape (N, T, C). + encoder_out_lens: + Encoder output lengths, of shape (N,). + """ + # logging.info(f"Memory allocated at entry: {torch.cuda.memory_allocated() // 1000000}M") + x, x_lens = self.encoder_embed(x, x_lens) + # logging.info(f"Memory allocated after encoder_embed: {torch.cuda.memory_allocated() // 1000000}M") + + src_key_padding_mask = make_pad_mask(x_lens) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask) + + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + assert torch.all(encoder_out_lens > 0), (x_lens, encoder_out_lens) + + return encoder_out, encoder_out_lens + def forward( self, - fbank: torch.Tensor = None, - input_ids: torch.LongTensor = None, - attention_mask: torch.Tensor = None, - labels: torch.LongTensor = None, + fbank: torch.Tensor, + fbank_lens: torch.Tensor, + input_ids: torch.LongTensor, + attention_mask: torch.Tensor, + labels: torch.LongTensor, ): - encoder_outs = self.encoder(fbank) + encoder_outs, _ = self.forward_encoder(fbank, fbank_lens) speech_features = self.encoder_projector(encoder_outs) @@ -229,15 +268,17 @@ class SPEECH_LLM(nn.Module): def decode( self, - fbank: torch.Tensor = None, - input_ids: torch.LongTensor = None, - attention_mask: torch.Tensor = None, + fbank: torch.Tensor, + fbank_lens: torch.Tensor, + input_ids: torch.LongTensor, + attention_mask: torch.Tensor, **kwargs, ): + encoder_outs, _ = self.forward_encoder(fbank, fbank_lens) - encoder_outs = self.encoder(fbank) speech_features = self.encoder_projector(encoder_outs) speech_features = speech_features.to(torch.float16) + inputs_embeds = self.llm.get_input_embeddings()(input_ids) ( inputs_embeds, diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/scaling.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/scaling.py new file mode 120000 index 000000000..6f398f431 --- /dev/null +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/scaling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/subsampling.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/subsampling.py new file mode 120000 index 000000000..01ae9002c --- /dev/null +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/subsampling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/subsampling.py \ No newline at end of file diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py index 7947a60a5..77c6a9b95 100755 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py @@ -18,28 +18,17 @@ # limitations under the License. """ Usage: -# fine-tuning with whisper and Qwen2 -pip install huggingface_hub['cli'] -mkdir -p models/whisper models/qwen - -# For aishell fine-tuned whisper model -huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_aishell_whisper exp_large_v2/whisper-large-v2-aishell1-epoch-10-avg-6.pt -# For multi-hans fine-tuned whisper model -# huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_multi-hans-zh_whisper v1.1/whisper-large-v2-multi-hans-zh-epoch-3-avg-10.pt - -# huggingface-clie download --local-dir models/qwen Qwen/Qwen2-7B-Instruct -huggingface-clie download --local-dir models/qwen Qwen/Qwen2-1.5B-Instruct - -torchrun --nproc_per_node 8 ./whisper_llm_zh/train.py \ +torchrun --nproc_per_node 8 ./zipformer_llm_zh/train.py \ --max-duration 200 \ - --exp-dir ./whisper_llm_zh/exp_test \ - --speech-encoder-path-or-name models/whisper/exp_large_v2/whisper-large-v2-aishell1-epoch-10-avg-6.pt \ + --exp-dir ./zipformer_llm_zh/exp_test \ + --speech-encoder-path-or-name models/zipformer/exp/epoch-999.pt \ --llm-path-or-name Qwen/Qwen2-1.5B-Instruct \ --manifest-dir data/fbank \ --deepspeed \ - --deepspeed_config ./whisper_llm_zh/ds_config_zero1.json \ + --deepspeed_config ./zipformer_llm_zh/ds_config_zero1.json \ --use-flash-attn True \ - --use-lora True --unfreeze-llm True + --use-lora True \ + --unfreeze-llm True """ import argparse @@ -53,7 +42,6 @@ import deepspeed import torch import torch.nn as nn import transformers -import whisper from asr_datamodule import AsrDataModule from deepspeed.utils.zero_to_fp32 import convert_zero_checkpoint_to_fp32_state_dict from lhotse.cut import Cut @@ -61,10 +49,12 @@ from lhotse.utils import fix_random_seed from model import IGNORE_TOKEN_ID, SPEECH_LLM, EncoderProjector from multi_dataset import MultiDataset from peft import LoraConfig, get_peft_model +from scaling import ScheduledFloat +from subsampling import Conv2dSubsampling from torch import Tensor from torch.utils.tensorboard import SummaryWriter from transformers import AutoModelForCausalLM, AutoTokenizer -from whisper_encoder_forward_monkey_patch import replace_whisper_encoder_forward +from zipformer import Zipformer2 from icefall.dist import get_rank, get_world_size from icefall.env import get_env_info @@ -90,14 +80,14 @@ def add_model_arguments(parser: argparse.ArgumentParser): parser.add_argument( "--speech-encoder-path-or-name", type=str, - default="whisper-large-v2", + default="zipformer", help="Path or name of the speech encoder.", ) parser.add_argument( "--encoder-projector-ds-rate", type=int, - default=8, + default=4, help="Downsample rate for the encoder projector.", ) parser.add_argument( @@ -121,6 +111,185 @@ def add_model_arguments(parser: argparse.ArgumentParser): help="Whether to unfreeze llm during training.", ) + # Zipformer + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + parser.add_argument( + "--causal", + type=str2bool, + default=False, + help="If True, use causal version of model.", + ) + + parser.add_argument( + "--chunk-size", + type=str, + default="16,32,64,-1", + help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. " + " Must be just -1 if --causal=False", + ) + + parser.add_argument( + "--left-context-frames", + type=str, + default="64,128,256,-1", + help="Maximum left-contexts for causal training, measured in frames which will " + "be converted to a number of chunks. If splitting into chunks, " + "chunk left-context frames will be chosen randomly from this list; else not relevant.", + ) + + +def load_model_params(ckpt: str, model: nn.Module, module: str, strict: bool = True): + """Load model params from checkpoint + + Args: + ckpt (str): Path to the checkpoint + model (nn.Module): model to be loaded + module (str): Module to be initialized + + """ + logging.info(f"Loading checkpoint from {ckpt}") + checkpoint = torch.load(ckpt, map_location="cpu") + + src_state_dict = checkpoint["model"] + dst_state_dict = model.state_dict() + logging.info(f"Loading parameters starting with prefix {module}") + module_prefix = module.strip() + "." + src_keys = [ + k[len(module_prefix) :] + for k in src_state_dict.keys() + if k.startswith(module_prefix) + ] + dst_keys = [k for k in dst_state_dict.keys()] + assert set(src_keys) == set(dst_keys) # two sets should match exactly + for key in src_keys: + dst_state_dict[key] = src_state_dict.pop(module_prefix + key) + + model.load_state_dict(dst_state_dict, strict=strict) + + return None + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_embed(params: AttributeDict) -> nn.Module: + # encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7) // 2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7) // 2 + # (2) embedding: num_features -> encoder_dims + # In the normal configuration, we will downsample once more at the end + # by a factor of 2, and most of the encoder stacks will run at a lower + # sampling rate. + encoder_embed = Conv2dSubsampling( + in_channels=params.feature_dim, + out_channels=_to_int_tuple(params.encoder_dim)[0], + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + ) + return encoder_embed + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Zipformer2( + output_downsampling_factor=2, + downsampling_factor=_to_int_tuple(params.downsampling_factor), + num_encoder_layers=_to_int_tuple(params.num_encoder_layers), + encoder_dim=_to_int_tuple(params.encoder_dim), + encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim), + query_head_dim=_to_int_tuple(params.query_head_dim), + pos_head_dim=_to_int_tuple(params.pos_head_dim), + value_head_dim=_to_int_tuple(params.value_head_dim), + pos_dim=params.pos_dim, + num_heads=_to_int_tuple(params.num_heads), + feedforward_dim=_to_int_tuple(params.feedforward_dim), + cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=params.causal, + chunk_size=_to_int_tuple(params.chunk_size), + left_context_frames=_to_int_tuple(params.left_context_frames), + ) + return encoder + def get_parser(): parser = argparse.ArgumentParser( @@ -154,7 +323,7 @@ def get_parser(): parser.add_argument( "--exp-dir", type=str, - default="whisper_qwen/exp", + default="zipformer_llm_zh/exp", help="""The experiment dir. It specifies the directory where all training related files, e.g., checkpoints, log, etc, are saved @@ -166,7 +335,7 @@ def get_parser(): type=str, default=None, help="""The path to the pretrained model if it is not None. Training will - start from this model. e.g. ./wenetspeech/ASR/whisper/exp_large_v2/epoch-4-avg-3.pt + start from this model. e.g. ./wenetspeech/ASR/zipformer/exp/epoch-999.pt """, ) @@ -231,7 +400,6 @@ def get_params() -> AttributeDict: params = AttributeDict( { "allowed_excess_duration_ratio": 0.1, - "subsampling_factor": 2, "frame_shift_ms": 10, "best_train_loss": float("inf"), "best_valid_loss": float("inf"), @@ -241,6 +409,9 @@ def get_params() -> AttributeDict: "log_interval": 50, "reset_interval": 200, "valid_interval": 5000, + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. "env_info": get_env_info(), } ) @@ -327,14 +498,13 @@ def compute_loss( return input_ids, attention_mask, target_ids device = next(model.parameters()).device + feature = batch["inputs"] - assert feature.ndim == 3 - feature = feature.to(device) - feature = feature.transpose(1, 2) # (N, C, T) - batch_idx_train = params.batch_idx_train supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"] + texts = batch["supervisions"]["text"] messages = [] @@ -353,7 +523,8 @@ def compute_loss( with torch.set_grad_enabled(is_training): model_outputs, acc = model( - fbank=feature, + fbank=feature.to(device), + fbank_lens=feature_lens.to(device), input_ids=input_ids.to(device), attention_mask=attention_mask.to(device), labels=target_ids.to(device), @@ -364,7 +535,6 @@ def compute_loss( info = MetricsTracker() with warnings.catch_warnings(): warnings.simplefilter("ignore") - feature_lens = supervisions["num_frames"] info["frames"] = (feature_lens // params.subsampling_factor).sum().item() # Note: We use reduction=sum while computing the loss. @@ -378,7 +548,7 @@ def compute_loss( def compute_validation_loss( params: AttributeDict, - tokenizer: whisper.tokenizer.Tokenizer, + tokenizer: AutoTokenizer, model: nn.Module, valid_dl: torch.utils.data.DataLoader, world_size: int = 1, @@ -586,10 +756,16 @@ def run(rank, world_size, args): logging.info("About to create model") - replace_whisper_encoder_forward() - whisper_model = whisper.load_model(params.speech_encoder_path_or_name, "cpu") - speech_encoder = whisper_model.encoder - speech_encoder_dim = whisper_model.dims.n_audio_state + speech_encoder_embed = get_encoder_embed(params) + speech_encoder = get_encoder_model(params) + load_model_params( + params.speech_encoder_path_or_name, speech_encoder_embed, "encoder_embed" + ) + load_model_params(params.speech_encoder_path_or_name, speech_encoder, "encoder") + + speech_encoder_dim = max(_to_int_tuple(params.encoder_dim)) + for name, param in speech_encoder_embed.named_parameters(): + param.requires_grad = False for name, param in speech_encoder.named_parameters(): param.requires_grad = False @@ -646,6 +822,7 @@ def run(rank, world_size, args): ) model = SPEECH_LLM( + speech_encoder_embed, speech_encoder, llm, encoder_projector, @@ -790,7 +967,6 @@ def display_and_save_batch( logging.info(f"Saving batch to {filename}") torch.save(batch, filename) - supervisions = batch["supervisions"] features = batch["inputs"] logging.info(f"features shape: {features.shape}") diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/zipformer.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/zipformer.py new file mode 120000 index 000000000..23011dda7 --- /dev/null +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/zipformer.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/zipformer.py \ No newline at end of file From ec6c8f748d7e91f3477e081554bf3b39b55287ea Mon Sep 17 00:00:00 2001 From: yfyeung Date: Fri, 9 May 2025 17:18:22 +0000 Subject: [PATCH 06/22] fix data prepare update --- egs/speech_llm/ASR_LLM/prepare.sh | 10 ++++++++++ egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py | 2 +- 2 files changed, 11 insertions(+), 1 deletion(-) diff --git a/egs/speech_llm/ASR_LLM/prepare.sh b/egs/speech_llm/ASR_LLM/prepare.sh index 8ca3c1c36..d602ce194 100755 --- a/egs/speech_llm/ASR_LLM/prepare.sh +++ b/egs/speech_llm/ASR_LLM/prepare.sh @@ -37,6 +37,15 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then huggingface-cli download --repo-type dataset --local-dir data/fbank yuekai/wenetspeech_whisper_fbank_lhotse huggingface-cli download --repo-type dataset --local-dir data/fbank yuekai/multi_hans_zh_whisper_fbank_lhotse huggingface-cli download --repo-type dataset --local-dir data/fbank yuekai/alimeeting_aishell4_training_whisper_fbank_lhotse + mkdir data/fbank/wenetspeech + mv data/fbank/cuts_L_fixed.jsonl.gz data/fbank/wenetspeech/ + mv data/fbank/cuts_DEV_fixed.jsonl.gz data/fbank/wenetspeech/ + mv data/fbank/cuts_TEST_MEETING.jsonl.gz data/fbank/wenetspeech/ + mv data/fbank/cuts_TEST_NET.jsonl.gz data/fbank/wenetspeech/ + mv data/fbank/L_split_100 data/fbank/wenetspeech/ + mv data/fbank/feats_DEV.lca data/fbank/wenetspeech/ + mv data/fbank/feats_TEST_MEETING.lca data/fbank/wenetspeech/ + mv data/fbank/feats_TEST_NET.lca data/fbank/wenetspeech/ fi if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then @@ -46,4 +55,5 @@ if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then mkdir data_speechio huggingface-cli download --repo-type model --local-dir data_speechio yuekai/icefall_asr_speechio mv data_speechio/fbank/* data/fbank + rm -rf data_speechio fi diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py index 77c6a9b95..82ba1abb3 100755 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py @@ -364,7 +364,7 @@ def get_parser(): parser.add_argument( "--use-aishell", type=str2bool, - default=True, + default=False, help="Whether to only use aishell1 dataset for training.", ) From 2420d0c95ff467e33decea413417628e1494733b Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Sat, 10 May 2025 02:13:25 +0800 Subject: [PATCH 07/22] update multi_dataset.py --- .../ASR_LLM/whisper_llm_zh/multi_dataset.py | 43 ++++++++++++------- 1 file changed, 28 insertions(+), 15 deletions(-) diff --git a/egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py b/egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py index f821fd29d..d116857af 100644 --- a/egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py +++ b/egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py @@ -244,8 +244,7 @@ class MultiDataset: } def aishell_train_cuts(self) -> CutSet: - logging.info("About to get multidataset train cuts") - logging.info("Loading Aishell-1 in lazy mode") + logging.info("Loading Aishell-1 train set in lazy mode") aishell_cuts = load_manifest_lazy( self.fbank_dir / "aishell_cuts_train.jsonl.gz" ) @@ -253,8 +252,7 @@ class MultiDataset: return aishell_cuts def aishell_dev_cuts(self) -> CutSet: - logging.info("About to get multidataset dev cuts") - logging.info("Loading Aishell set in lazy mode") + logging.info("Loading Aishell-1 dev set in lazy mode") aishell_dev_cuts = load_manifest_lazy( self.fbank_dir / "aishell_cuts_dev.jsonl.gz" ) @@ -262,8 +260,7 @@ class MultiDataset: return aishell_dev_cuts def aishell_test_cuts(self) -> CutSet: - logging.info("About to get multidataset test cuts") - logging.info("Loading Aishell set in lazy mode") + logging.info("Loading Aishell-1 test set in lazy mode") aishell_test_cuts = load_manifest_lazy( self.fbank_dir / "aishell_cuts_test.jsonl.gz" ) @@ -273,8 +270,7 @@ class MultiDataset: } def aishell2_train_cuts(self) -> CutSet: - logging.info("About to get multidataset train cuts") - logging.info("Loading Aishell-2 in lazy mode") + logging.info("Loading Aishell-2 train set in lazy mode") aishell_2_cuts = load_manifest_lazy( self.fbank_dir / "aishell2_cuts_train.jsonl.gz" ) @@ -282,8 +278,7 @@ class MultiDataset: return aishell_2_cuts def aishell2_dev_cuts(self) -> CutSet: - logging.info("About to get multidataset dev cuts") - logging.info("Loading Aishell-2 set in lazy mode") + logging.info("Loading Aishell-2 dev set in lazy mode") aishell2_dev_cuts = load_manifest_lazy( self.fbank_dir / "aishell2_cuts_dev.jsonl.gz" ) @@ -291,8 +286,7 @@ class MultiDataset: return aishell2_dev_cuts def aishell2_test_cuts(self) -> CutSet: - logging.info("About to get multidataset test cuts") - logging.info("Loading Aishell-2 set in lazy mode") + logging.info("Loading Aishell-2 test set in lazy mode") aishell2_test_cuts = load_manifest_lazy( self.fbank_dir / "aishell2_cuts_test.jsonl.gz" ) @@ -301,9 +295,28 @@ class MultiDataset: "aishell2_test": aishell2_test_cuts, } + def wenetspeech_dev_cuts(self) -> CutSet: + logging.info("Loading WeNetSpeech DEV set in lazy mode") + wenetspeech_dev_cuts = load_manifest_lazy( + self.fbank_dir / "wenetspeech" / "cuts_DEV_fixed.jsonl.gz" + ) + + return { + "wenetspeech-dev": wenetspeech_dev_cuts, + } + + def wenetspeech_test_net_cuts(self) -> CutSet: + logging.info("Loading WeNetSpeech-net test set in lazy mode") + wenetspeech_test_net_cuts = load_manifest_lazy( + self.fbank_dir / "wenetspeech" / "cuts_TEST_NET.jsonl.gz" + ) + + return { + "wenetspeech-net_test": wenetspeech_test_net_cuts, + } + def wenetspeech_test_meeting_cuts(self) -> CutSet: - logging.info("About to get multidataset test cuts") - logging.info("Loading WeNetSpeech set in lazy mode") + logging.info("Loading WeNetSpeech-meeting test set in lazy mode") wenetspeech_test_meeting_cuts = load_manifest_lazy( self.fbank_dir / "wenetspeech" / "cuts_TEST_MEETING.jsonl.gz" ) @@ -313,7 +326,7 @@ class MultiDataset: } def speechio_test_cuts(self) -> Dict[str, CutSet]: - logging.info("About to get multidataset test cuts") + logging.info("Loading SpeechIO test set in lazy mode") start_index = 0 end_index = 26 dataset_parts = [] From c75767f600510c0f428c22cdd86ce12c97a38674 Mon Sep 17 00:00:00 2001 From: yfyeung Date: Sat, 10 May 2025 17:32:14 +0000 Subject: [PATCH 08/22] set world_size and rank explicitly update --- egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py index 82ba1abb3..edea3bdb9 100755 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py @@ -885,14 +885,21 @@ def run(rank, world_size, args): sampler_state_dict["max_duration"] = params.max_duration train_dl = data_module.train_dataloaders( - train_cuts, sampler_state_dict=sampler_state_dict + train_cuts, + sampler_state_dict=sampler_state_dict, + world_size=world_size, + rank=rank, ) if params.use_aishell: valid_cuts = multi_dataset.aishell_dev_cuts() else: valid_cuts = multi_dataset.dev_cuts() - valid_dl = data_module.valid_dataloaders(valid_cuts) + valid_dl = data_module.valid_dataloaders( + valid_cuts, + world_size=world_size, + rank=rank, + ) if args.tensorboard and rank == 0: tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") From cd3adad46dc0b43b3ba3b408ca3d1a23dc123f07 Mon Sep 17 00:00:00 2001 From: yfyeung Date: Sat, 10 May 2025 17:47:05 +0000 Subject: [PATCH 09/22] use quadratic-duration --- .../ASR/zipformer/asr_datamodule.py | 36 ++++++++++++++++++- 1 file changed, 35 insertions(+), 1 deletion(-) diff --git a/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py b/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py index 341579acb..c74d212d4 100644 --- a/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py +++ b/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py @@ -109,6 +109,25 @@ class AsrDataModule: help="The number of buckets for the DynamicBucketingSampler" "(you might want to increase it for larger datasets).", ) + group.add_argument( + "--num-cuts-for-bins-estimate", + type=int, + default=10000, + help="We will draw this many cuts to estimate the duration" + "bins for creating similar-duration buckets. Larger number" + "means a better estimate to the data distribution, possibly" + "at a longer init cost.", + ) + group.add_argument( + "--quadratic-duration", + type=float, + default=None, + help="When set, it adds an extra penalty that's quadratic" + "in size w.r.t. a cuts duration. This helps get a more" + "even GPU utilization across different input lengths when" + "models have quadratic input complexity.0 Set between 15" + "and 40 for transformers.", + ) group.add_argument( "--concatenate-cuts", type=str2bool, @@ -205,6 +224,8 @@ class AsrDataModule: self, cuts_train: CutSet, sampler_state_dict: Optional[Dict[str, Any]] = None, + world_size: Optional[int] = None, + rank: Optional[int] = None, ) -> DataLoader: """ Args: @@ -295,11 +316,15 @@ class AsrDataModule: train_sampler = DynamicBucketingSampler( cuts_train, max_duration=self.args.max_duration, + quadratic_duration=self.args.quadratic_duration, + num_cuts_for_bins_estimate=self.args.num_cuts_for_bins_estimate, shuffle=self.args.shuffle, num_buckets=self.args.num_buckets, buffer_size=self.args.num_buckets * 2000, shuffle_buffer_size=self.args.num_buckets * 5000, drop_last=self.args.drop_last, + world_size=world_size, + rank=rank, ) else: logging.info("Using SimpleCutSampler.") @@ -307,6 +332,8 @@ class AsrDataModule: cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, + world_size=world_size, + rank=rank, ) logging.info("About to create train dataloader") @@ -330,7 +357,12 @@ class AsrDataModule: return train_dl - def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader: + def valid_dataloaders( + self, + cuts_valid: CutSet, + world_size: Optional[int] = None, + rank: Optional[int] = None, + ) -> DataLoader: transforms = [] if self.args.concatenate_cuts: transforms = [ @@ -355,6 +387,8 @@ class AsrDataModule: cuts_valid, max_duration=self.args.max_duration, shuffle=False, + world_size=world_size, + rank=rank, ) logging.info("About to create dev dataloader") valid_dl = DataLoader( From 5fbeed9f96ff39e2f4551b8fd7d661327d94a70e Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Mon, 12 May 2025 00:48:42 +0800 Subject: [PATCH 10/22] fix SwooshR and SwooshL --- egs/librispeech/ASR/zipformer/scaling.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/egs/librispeech/ASR/zipformer/scaling.py b/egs/librispeech/ASR/zipformer/scaling.py index 6d6281903..11375385e 100644 --- a/egs/librispeech/ASR/zipformer/scaling.py +++ b/egs/librispeech/ASR/zipformer/scaling.py @@ -1403,9 +1403,9 @@ class SwooshL(torch.nn.Module): zero = torch.tensor(0.0, dtype=x.dtype, device=x.device) return logaddexp(zero, x - 4.0) - 0.08 * x - 0.035 if not x.requires_grad: - return k2.swoosh_l_forward(x) + return k2.swoosh_l_forward(x).to(x.dtype) else: - return k2.swoosh_l(x) + return k2.swoosh_l(x).to(x.dtype) # return SwooshLFunction.apply(x) @@ -1477,9 +1477,9 @@ class SwooshR(torch.nn.Module): zero = torch.tensor(0.0, dtype=x.dtype, device=x.device) return logaddexp(zero, x - 1.0) - 0.08 * x - 0.313261687 if not x.requires_grad: - return k2.swoosh_r_forward(x) + return k2.swoosh_r_forward(x).to(x.dtype) else: - return k2.swoosh_r(x) + return k2.swoosh_r(x).to(x.dtype) # return SwooshRFunction.apply(x) From 9939c2b72d079dc939633f6d038475e9bd3a5c4a Mon Sep 17 00:00:00 2001 From: yfyeung Date: Sun, 11 May 2025 17:03:44 +0000 Subject: [PATCH 11/22] remove duplicated torch autocast --- .../ASR_LLM/zipformer_llm_zh/train.py | 32 +++++++++---------- 1 file changed, 16 insertions(+), 16 deletions(-) diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py index edea3bdb9..5d47f128a 100755 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py @@ -501,6 +501,8 @@ def compute_loss( feature = batch["inputs"] assert feature.ndim == 3 + if params.use_fp16: + feature = feature.half() supervisions = batch["supervisions"] feature_lens = supervisions["num_frames"] @@ -559,14 +561,13 @@ def compute_validation_loss( tot_loss = MetricsTracker() for batch_idx, batch in enumerate(valid_dl): - with torch.cuda.amp.autocast(enabled=params.use_fp16): - loss, loss_info = compute_loss( - params=params, - tokenizer=tokenizer, - model=model, - batch=batch, - is_training=False, - ) + loss, loss_info = compute_loss( + params=params, + tokenizer=tokenizer, + model=model, + batch=batch, + is_training=False, + ) assert loss.requires_grad is False tot_loss = tot_loss + loss_info @@ -680,14 +681,13 @@ def train_one_epoch( f"rm -rf {params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}" ) try: - with torch.cuda.amp.autocast(enabled=params.use_fp16): - loss, loss_info = compute_loss( - params=params, - tokenizer=tokenizer, - model=model, - batch=batch, - is_training=True, - ) + loss, loss_info = compute_loss( + params=params, + tokenizer=tokenizer, + model=model, + batch=batch, + is_training=True, + ) # summary stats tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info From c078772e59e797754ec7c4e891f1f17aa2c82316 Mon Sep 17 00:00:00 2001 From: yfyeung Date: Sun, 11 May 2025 17:23:19 +0000 Subject: [PATCH 12/22] skip OOM --- .../ASR_LLM/zipformer_llm_zh/train.py | 33 +++++++++++++++---- 1 file changed, 27 insertions(+), 6 deletions(-) diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py index 5d47f128a..2ace5c532 100755 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py @@ -32,6 +32,7 @@ torchrun --nproc_per_node 8 ./zipformer_llm_zh/train.py \ """ import argparse +import gc import logging import os import warnings @@ -625,6 +626,12 @@ def train_one_epoch( The rank of the node in DDP training. If no DDP is used, it should be set to 0. """ + + def free_gpu_cache(): + gc.collect() + if torch.cuda.is_available(): + torch.cuda.empty_cache() + model.train() model.encoder.eval() if not params.unfreeze_llm: @@ -688,9 +695,6 @@ def train_one_epoch( batch=batch, is_training=True, ) - # summary stats - tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info - # NOTE: We use reduction==sum and loss is computed over utterances # in the batch and there is no normalization to it so far. @@ -700,9 +704,26 @@ def train_one_epoch( model.backward(loss) model.step() - except: # noqa - display_and_save_batch(batch, params=params) - raise + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + except Exception as e: + logging.warning(f"Caught exception: {e}") + if ( + "CUDA" not in str(e) + and "cuDNN error" not in str(e) + and "NCCL error" not in str(e) + ): + display_and_save_batch(batch, params=params) + raise e + + try: + loss = None + loss_info = None + except: + pass + + free_gpu_cache() if batch_idx % params.log_interval == 0: try: From 2793ccdf56bb881fafa0a397d47008d8a162452b Mon Sep 17 00:00:00 2001 From: yfyeung Date: Mon, 12 May 2025 06:36:20 +0000 Subject: [PATCH 13/22] remove checkpoint save after validation --- .../ASR_LLM/zipformer_llm_zh/train.py | 32 +++---------------- 1 file changed, 4 insertions(+), 28 deletions(-) diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py index 2ace5c532..eaae4a33e 100755 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py @@ -500,10 +500,10 @@ def compute_loss( device = next(model.parameters()).device - feature = batch["inputs"] - assert feature.ndim == 3 + features = batch["inputs"] + assert features.ndim == 3 if params.use_fp16: - feature = feature.half() + features = features.half() supervisions = batch["supervisions"] feature_lens = supervisions["num_frames"] @@ -526,7 +526,7 @@ def compute_loss( with torch.set_grad_enabled(is_training): model_outputs, acc = model( - fbank=feature.to(device), + fbank=features.to(device), fbank_lens=feature_lens.to(device), input_ids=input_ids.to(device), attention_mask=attention_mask.to(device), @@ -663,30 +663,6 @@ def train_one_epoch( valid_info.write_summary( tb_writer, "train/valid_", params.batch_idx_train ) - if batch_idx != 0: - model.save_checkpoint( - save_dir=params.exp_dir, - tag=f"epoch-{params.cur_epoch}-checkpoint-{batch_idx}", - client_state={}, - exclude_frozen_parameters=True, - ) - - if rank == 0: - convert_zero_checkpoint_to_fp32_state_dict( - params.exp_dir, - f"{params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}.pt", - tag=f"epoch-{params.cur_epoch}-checkpoint-{batch_idx}", - exclude_frozen_parameters=True, - ) - # save sampler state dict into checkpoint - sampler_state_dict = train_dl.sampler.state_dict() - torch.save( - sampler_state_dict, - f"{params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}-sampler.pt", - ) - os.system( - f"rm -rf {params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}" - ) try: loss, loss_info = compute_loss( params=params, From 06667e1f6d8eb45d30e883988e61106ae55e97ad Mon Sep 17 00:00:00 2001 From: yfyeung Date: Mon, 12 May 2025 16:49:42 +0000 Subject: [PATCH 14/22] add batch shave mechanism fix fix --- .../ASR_LLM/zipformer_llm_zh/train.py | 100 ++++++++++++------ 1 file changed, 67 insertions(+), 33 deletions(-) diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py index eaae4a33e..441b4f266 100755 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py @@ -362,6 +362,16 @@ def get_parser(): help="Whether to use half precision training.", ) + parser.add_argument( + "--shave-rate", + type=float, + default=0.1, + help="""The factor to reduce the batch when an OOM occurs. + If OOM persists for the same batch, this factor will be + progressively multiplied by 1.5. Set to 0 to disable. + """, + ) + parser.add_argument( "--use-aishell", type=str2bool, @@ -627,6 +637,17 @@ def train_one_epoch( be set to 0. """ + def shave_batch(batch: dict, factor: float): + n_utt = len(batch["supervisions"]["text"]) + skip_point = max(1, int(factor * n_utt)) + if n_utt - skip_point <= 0: + return False + for key in batch["supervisions"].keys(): + batch["supervisions"][key] = batch["supervisions"][key][skip_point:] + max_len = max(batch["supervisions"]["num_frames"]).item() + batch["inputs"] = batch["inputs"][skip_point:, :max_len] + return True + def free_gpu_cache(): gc.collect() if torch.cuda.is_available(): @@ -642,6 +663,7 @@ def train_one_epoch( for batch_idx, batch in enumerate(train_dl): params.batch_idx_train += 1 batch_size = len(batch["supervisions"]["text"]) + if batch_idx % params.valid_interval == 0: logging.info("Computing validation loss") valid_info = compute_validation_loss( @@ -663,44 +685,56 @@ def train_one_epoch( valid_info.write_summary( tb_writer, "train/valid_", params.batch_idx_train ) - try: - loss, loss_info = compute_loss( - params=params, - tokenizer=tokenizer, - model=model, - batch=batch, - is_training=True, - ) - # NOTE: We use reduction==sum and loss is computed over utterances - # in the batch and there is no normalization to it so far. - - # deepspeed's backward() is different from torch's backward() - # in that it does not accept a loss tensor as input. - # It computes the loss internally. - model.backward(loss) - model.step() - - # summary stats - tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info - - except Exception as e: - logging.warning(f"Caught exception: {e}") - if ( - "CUDA" not in str(e) - and "cuDNN error" not in str(e) - and "NCCL error" not in str(e) - ): - display_and_save_batch(batch, params=params) - raise e + shave_rate = params.shave_rate + while True: try: - loss = None - loss_info = None - except: - pass + loss, loss_info = compute_loss( + params=params, + tokenizer=tokenizer, + model=model, + batch=batch, + is_training=True, + ) + + # NOTE: we use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + # deepspeed's backward() is different from torch's backward() + model.backward(loss) + model.step() + + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # finish this step + break + except Exception as e: + logging.warning(f"Caught exception: {e}") + if shave_rate <= 0 or ( + "CUDA" not in str(e) + and "cuDNN error" not in str(e) + and "NCCL error" not in str(e) + ): + display_and_save_batch(batch, params=params) + raise e + + loss = None + loss_info = None free_gpu_cache() + if shave_batch(batch, shave_rate): + logging.warning( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}: {shave_rate * 100:.2f}% batch reduced", + ) + shave_rate = min(shave_rate * 1.5, 0.5) + else: + raise RuntimeError( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}: batch reduced to empty in retry" + ) + if batch_idx % params.log_interval == 0: try: cur_lr = scheduler.get_last_lr()[0] From 62dfe56cbef2b14ff76104d45756a8127071c42c Mon Sep 17 00:00:00 2001 From: yifanyeung Date: Tue, 13 May 2025 06:14:59 +0000 Subject: [PATCH 15/22] restore checkpoint save after validation --- .../ASR_LLM/zipformer_llm_zh/train.py | 24 +++++++++++++++++++ 1 file changed, 24 insertions(+) diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py index 441b4f266..c668e7f64 100755 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py @@ -685,6 +685,30 @@ def train_one_epoch( valid_info.write_summary( tb_writer, "train/valid_", params.batch_idx_train ) + if batch_idx != 0: + model.save_checkpoint( + save_dir=params.exp_dir, + tag=f"epoch-{params.cur_epoch}-checkpoint-{batch_idx}", + client_state={}, + exclude_frozen_parameters=True, + ) + + if rank == 0: + convert_zero_checkpoint_to_fp32_state_dict( + params.exp_dir, + f"{params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}.pt", + tag=f"epoch-{params.cur_epoch}-checkpoint-{batch_idx}", + exclude_frozen_parameters=True, + ) + # save sampler state dict into checkpoint + sampler_state_dict = train_dl.sampler.state_dict() + torch.save( + sampler_state_dict, + f"{params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}-sampler.pt", + ) + os.system( + f"rm -rf {params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}" + ) shave_rate = params.shave_rate while True: From 24b6f42340dfa29877d6230a2cbd2437ed28770c Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Tue, 13 May 2025 14:34:05 +0800 Subject: [PATCH 16/22] fix typos in docs fix typo in RESULTS.md Update RESULTS.md --- egs/multi_zh-hans/ASR/README.md | 4 ++-- egs/multi_zh-hans/ASR/RESULTS.md | 8 ++++---- 2 files changed, 6 insertions(+), 6 deletions(-) diff --git a/egs/multi_zh-hans/ASR/README.md b/egs/multi_zh-hans/ASR/README.md index 1e60c733c..8704889dc 100644 --- a/egs/multi_zh-hans/ASR/README.md +++ b/egs/multi_zh-hans/ASR/README.md @@ -3,7 +3,7 @@ This recipe includes scripts for training Zipformer model using multiple Chinese datasets. -# Included Training Sets +# Included Training Dataset 1. THCHS-30 2. AiShell-{1,2,4} 3. ST-CMDS @@ -14,7 +14,7 @@ This recipe includes scripts for training Zipformer model using multiple Chinese 8. WeNetSpeech 9. KeSpeech-ASR -|Datset| Number of hours| URL| +|Dataset| Number of hours| URL| |---|---:|---| |**TOTAL**|14,106|---| |THCHS-30|35|https://www.openslr.org/18/| diff --git a/egs/multi_zh-hans/ASR/RESULTS.md b/egs/multi_zh-hans/ASR/RESULTS.md index 622218d02..e689cee2c 100644 --- a/egs/multi_zh-hans/ASR/RESULTS.md +++ b/egs/multi_zh-hans/ASR/RESULTS.md @@ -99,7 +99,7 @@ Character Error Rates (CERs) listed below are produced by the checkpoint of the | Datasets | alimeeting | alimeeting | aishell-1 | aishell-1 | aishell-2 | aishell-2 | aishell-4 | magicdata | magicdata | kespeech-asr | kespeech-asr | kespeech-asr | WenetSpeech | WenetSpeech | WenetSpeech | |--------------------------------|-------------------|--------------|----------------|-------------|------------------|-------------|------------------|------------------|-------------|-----------------------|-----------------------|-------------|--------------------|-------------------------|---------------------| -| Zipformer CER (%) | eval | test | dev | test | dev | test | test | dev | test | dev phase1 | dev phase2 | test | dev | test meeting | test net | +| Split | eval | test | dev | test | dev | test | test | dev | test | dev phase1 | dev phase2 | test | dev | test meeting | test net | | Transducer Greedy Offline | 21.67 | 23.43 | 1.22 | 1.31 | 3.17 | 3.27 | 14.64 | 2.42 | 1.99 | 5.00 | 2.29 | 5.98 | 5.15 | 5.85 | 6.89 | Pre-trained model can be found here : https://huggingface.co/yuekai/icefall-asr-multi-zh-hans-zipformer-xl @@ -152,7 +152,7 @@ Character Error Rates (CERs) listed below are produced by the checkpoint of the | Datasets | alimeeting | alimeeting | aishell-1 | aishell-1 | aishell-2 | aishell-2 | aishell-4 | magicdata | magicdata | kespeech-asr | kespeech-asr | kespeech-asr | WenetSpeech | WenetSpeech | WenetSpeech | |--------------------------------|-------------------|--------------|----------------|-------------|------------------|-------------|------------------|------------------|-------------|-----------------------|-----------------------|-------------|--------------------|-------------------------|---------------------| -| Zipformer CER (%) | eval | test | dev | test | dev | test | test | dev | test | dev phase1 | dev phase2 | test | dev | test meeting | test net | +| Split | eval | test | dev | test | dev | test | test | dev | test | dev phase1 | dev phase2 | test | dev | test meeting | test net | | CTC Greedy Streaming | 26.50 | 28.10| 1.71 | 1.97| 3.89| 4.06 | 17.23 | 3.69 | 2.87 | 8.14 | 3.61 |9.51 | 6.11 | 8.13 | 10.62 | | CTC Greedy Offline | 23.47 | 25.02 | 1.39 | 1.50 | 3.15 | 3.41 | 15.14 | 3.07 | 2.37 | 6.06 | 2.90 | 7.13 | 5.40 | 6.52 | 9.64 | | Transducer Greedy Offline | 23.16 | 24.78 | 1.33 | 1.38 | 3.06 | 3.23 | 15.36 | 2.54 | 2.09 | 5.24 | 2.28 | 6.26 | 4.87 | 6.26 | 7.07 | @@ -193,7 +193,7 @@ Character Error Rates (CERs) listed below are produced by the checkpoint of the | Datasets | aidatatang _200zh | aidatatang _200zh | alimeeting | alimeeting | aishell-1 | aishell-1 | aishell-2 | aishell-2 | aishell-4 | magicdata | magicdata | kespeech-asr | kespeech-asr | kespeech-asr | WenetSpeech | WenetSpeech | WenetSpeech | |--------------------------------|------------------------------|-------------|-------------------|--------------|----------------|-------------|------------------|-------------|------------------|------------------|-------------|-----------------------|-----------------------|-------------|--------------------|-------------------------|---------------------| -| Zipformer CER (%) | dev | test | eval | test | dev | test | dev | test | test | dev | test | dev phase1 | dev phase2 | test | dev | test meeting | test net | +| Split | dev | test | eval | test | dev | test | dev | test | test | dev | test | dev phase1 | dev phase2 | test | dev | test meeting | test net | | CTC Decoding | 2.86 | 3.36 | 22.93 | 24.28 | 2.05 | 2.27 | 3.33 | 3.82 | 15.45 | 3.49 | 2.77 | 6.90 | 2.85 | 8.29 | 9.41 | 6.92 | 8.57 | | Greedy Search | 3.36 | 3.83 | 23.90 | 25.18 | 2.77 | 3.08 | 3.70 | 4.04 | 16.13 | 3.77 | 3.15 | 6.88 | 3.14 | 8.08 | 9.04 | 7.19 | 8.17 | @@ -226,7 +226,7 @@ Character Error Rates (CERs) listed below are produced by the checkpoint of the | Datasets | aidatatang _200zh | aidatatang _200zh | alimeeting | alimeeting | aishell-1 | aishell-1 | aishell-2 | aishell-2 | aishell-4 | magicdata | magicdata | kespeech-asr | kespeech-asr | kespeech-asr | WenetSpeech | WenetSpeech | WenetSpeech | |--------------------------------|------------------------------|-------------|-------------------|--------------|----------------|-------------|------------------|-------------|------------------|------------------|-------------|-----------------------|-----------------------|-------------|--------------------|-------------------------|---------------------| -| Zipformer CER (%) | dev | test | eval| test | dev | test | dev| test | test | dev| test | dev phase1 | dev phase2 | test | dev | test meeting | test net | +| Split | dev | test | eval| test | dev | test | dev| test | test | dev| test | dev phase1 | dev phase2 | test | dev | test meeting | test net | | Greedy Search | 3.2 | 3.67 | 23.15 | 24.78 | 2.91 | 3.04 | 3.59 | 4.03 | 15.68 | 3.68 | 3.12 | 6.69 | 3.19 | 8.01 | 9.32 | 7.05 | 8.78 | From 11ccaa3ab854f41f4c7f8ed382a7fced077be834 Mon Sep 17 00:00:00 2001 From: yfyeung Date: Mon, 26 May 2025 04:11:28 +0000 Subject: [PATCH 17/22] add requirements.txt --- egs/speech_llm/ASR_LLM/zipformer_llm_zh/requirements.txt | 5 +++++ 1 file changed, 5 insertions(+) create mode 100644 egs/speech_llm/ASR_LLM/zipformer_llm_zh/requirements.txt diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/requirements.txt b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/requirements.txt new file mode 100644 index 000000000..98f9be968 --- /dev/null +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/requirements.txt @@ -0,0 +1,5 @@ +librosa +deepspeed +transformers>=4.37.0 +flash-attn +peft From 7c30dd570b5c7161b9ee411ebfba9306e555e178 Mon Sep 17 00:00:00 2001 From: yfyeung Date: Wed, 28 May 2025 03:42:03 +0000 Subject: [PATCH 18/22] restrict deepspeed >=0.16.9 --- egs/speech_llm/ASR_LLM/zipformer_llm_zh/requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/requirements.txt b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/requirements.txt index 98f9be968..0c554f17b 100644 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/requirements.txt +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/requirements.txt @@ -1,5 +1,5 @@ librosa -deepspeed +deepspeed>=0.16.9 transformers>=4.37.0 flash-attn peft From 05e309442938545ac1d40c5f4651904d680356dd Mon Sep 17 00:00:00 2001 From: Zengwei Yao Date: Tue, 27 May 2025 12:09:59 +0800 Subject: [PATCH 19/22] refactor branch exchange in cr-ctc (#1954) --- egs/librispeech/ASR/zipformer/model.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/egs/librispeech/ASR/zipformer/model.py b/egs/librispeech/ASR/zipformer/model.py index c7dbe1e0a..f2791e51f 100644 --- a/egs/librispeech/ASR/zipformer/model.py +++ b/egs/librispeech/ASR/zipformer/model.py @@ -210,10 +210,10 @@ class AsrModel(nn.Module): ) # Compute consistency regularization loss - exchanged_targets = ctc_output.detach().chunk(2, dim=0) - exchanged_targets = torch.cat( - [exchanged_targets[1], exchanged_targets[0]], dim=0 - ) # exchange: [x1, x2] -> [x2, x1] + batch_size = ctc_output.shape[0] + assert batch_size % 2 == 0, batch_size + # exchange: [x1, x2] -> [x2, x1] + exchanged_targets = torch.roll(ctc_output.detach(), batch_size // 2, dims=0) cr_loss = nn.functional.kl_div( input=ctc_output, target=exchanged_targets, From 34639d52498266f4674e81b33bd56f4fc04ca2a7 Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Tue, 3 Jun 2025 21:45:47 +0800 Subject: [PATCH 20/22] use padding instead of trimming (suggested by @shylockasr) use ctc compress (suggested by @shylockasr) fix revert revert revert --- egs/speech_llm/ASR_LLM/.gitignore | 3 + .../ASR_LLM/whisper_llm_zh/multi_dataset.py | 92 +------------------ .../ASR_LLM/zipformer_llm_zh/model.py | 62 ++++++++++++- 3 files changed, 62 insertions(+), 95 deletions(-) diff --git a/egs/speech_llm/ASR_LLM/.gitignore b/egs/speech_llm/ASR_LLM/.gitignore index 604f0f2cf..72ea9549c 100644 --- a/egs/speech_llm/ASR_LLM/.gitignore +++ b/egs/speech_llm/ASR_LLM/.gitignore @@ -1 +1,4 @@ models +train*.sh +decode*.sh +sync*.sh diff --git a/egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py b/egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py index d116857af..3c960c716 100644 --- a/egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py +++ b/egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py @@ -47,103 +47,13 @@ class MultiDataset: def train_cuts(self) -> CutSet: logging.info("About to get multidataset train cuts") - # THCHS-30 - logging.info("Loading THCHS-30 in lazy mode") - thchs_30_cuts = load_manifest_lazy( - self.fbank_dir / "thchs_30_cuts_train.jsonl.gz" - ) - - # AISHELL-1 - logging.info("Loading Aishell-1 in lazy mode") - aishell_cuts = load_manifest_lazy( - self.fbank_dir / "aishell_cuts_train.jsonl.gz" - ) - - # AISHELL-2 - logging.info("Loading Aishell-2 in lazy mode") - aishell_2_cuts = load_manifest_lazy( - self.fbank_dir / "aishell2_cuts_train.jsonl.gz" - ) - - # AISHELL-4 - logging.info("Loading Aishell-4 in lazy mode") - aishell_4_L_cuts = load_manifest_lazy( - self.fbank_dir / "aishell4_cuts_train_L.jsonl.gz" - ) - aishell_4_M_cuts = load_manifest_lazy( - self.fbank_dir / "aishell4_cuts_train_M.jsonl.gz" - ) - aishell_4_S_cuts = load_manifest_lazy( - self.fbank_dir / "aishell4_cuts_train_S.jsonl.gz" - ) - - # ST-CMDS - logging.info("Loading ST-CMDS in lazy mode") - stcmds_cuts = load_manifest_lazy(self.fbank_dir / "stcmds_cuts_train.jsonl.gz") - - # Primewords - logging.info("Loading Primewords in lazy mode") - primewords_cuts = load_manifest_lazy( - self.fbank_dir / "primewords_cuts_train.jsonl.gz" - ) - - # MagicData - logging.info("Loading MagicData in lazy mode") - magicdata_cuts = load_manifest_lazy( - self.fbank_dir / "magicdata_cuts_train.jsonl.gz" - ) - - # Ali-Meeting - logging.info("Loading Ali-Meeting in lazy mode") - alimeeting_cuts = load_manifest_lazy( - self.fbank_dir / "alimeeting-far_cuts_train.jsonl.gz" - ) - # WeNetSpeech logging.info("Loading WeNetSpeech in lazy mode") wenetspeech_L_cuts = load_manifest_lazy( self.fbank_dir / "wenetspeech" / "cuts_L_fixed.jsonl.gz" ) - # KeSpeech - logging.info("Loading KeSpeech in lazy mode") - kespeech_1_cuts = load_manifest_lazy( - self.fbank_dir / "kespeech" / "kespeech-asr_cuts_train_phase1.jsonl.gz" - ) - kespeech_2_cuts = load_manifest_lazy( - self.fbank_dir / "kespeech" / "kespeech-asr_cuts_train_phase2.jsonl.gz" - ) - - return CutSet.mux( - thchs_30_cuts, - aishell_cuts, - aishell_2_cuts, - aishell_4_L_cuts, - aishell_4_M_cuts, - aishell_4_S_cuts, - alimeeting_cuts, - stcmds_cuts, - primewords_cuts, - magicdata_cuts, - wenetspeech_L_cuts, - kespeech_1_cuts, - kespeech_2_cuts, - weights=[ - len(thchs_30_cuts), - len(aishell_cuts), - len(aishell_2_cuts), - len(aishell_4_L_cuts), - len(aishell_4_M_cuts), - len(aishell_4_S_cuts), - len(alimeeting_cuts), - len(stcmds_cuts), - len(primewords_cuts), - len(magicdata_cuts), - len(wenetspeech_L_cuts), - len(kespeech_1_cuts), - len(kespeech_2_cuts), - ], - ) + return wenetspeech_L_cuts def dev_cuts(self) -> CutSet: logging.info("About to get multidataset dev cuts") diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py index 5f0d4b8e5..d585ec871 100644 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py @@ -30,9 +30,11 @@ class EncoderProjector(nn.Module): def forward(self, x): batch_size, seq_len, feat_dim = x.size() - num_frames_to_discard = seq_len % self.downsample_rate - if num_frames_to_discard > 0: - x = x[:, :-num_frames_to_discard, :] + num_padding_frames = ( + self.downsample_rate - seq_len % self.downsample_rate + ) % self.downsample_rate + if num_padding_frames > 0: + x = torch.nn.functional.pad(x, (0, 0, 0, num_padding_frames)) seq_len = x.size(1) x = x.contiguous() @@ -62,6 +64,7 @@ class SPEECH_LLM(nn.Module): self, encoder_embed: nn.Module, encoder: EncoderInterface, + ctc_output: nn.Module, llm: nn.Module, encoder_projector: nn.Module, ): @@ -230,6 +233,57 @@ class SPEECH_LLM(nn.Module): return encoder_out, encoder_out_lens + def ctc_compress( + self, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + blank_id: int = 0, + ) -> torch.Tensor: + """ + Remove frames from encoder_out where CTC argmax predicts blank. + Args: + encoder_out: Tensor of shape (N, T, C), encoder output. + encoder_out_lens: Tensor of shape (N,), lengths before padding. + blank_id: CTC blank token ID (default: 0). + + Returns: + Compressed CTC output of shape (N, T', C). + """ + # 1. Compute CTC argmax predictions + ctc_output = self.ctc_output(encoder_out) + ctc_preds = ctc_output.argmax(dim=-1) + + # 2. Create non-blank, non-pad mask + padding_mask = make_pad_mask(encoder_out_lens) + non_blank_mask = (ctc_preds != blank_id) & (~padding_mask) + + # 3. Compute lengths after compress + compressed_lens = non_blank_mask.sum(dim=1) + max_len = compressed_lens.max().item() + + # 4. Pre-pad output + pad_lens_list = ( + torch.full_like( + compressed_lens, + max_len, + device=ctc_output.device, + ) + - compressed_lens + ) + max_pad_len = int(pad_lens_list.max()) + padded_ctc_output = torch.nn.functional.pad(ctc_output, [0, 0, 0, max_pad_len]) + + # 5. Create final mask + padding_mask = ~make_pad_mask(pad_lens_list) + total_mask = torch.concat([non_blank_mask, padding_mask], dim=1) + + # 6. Apply mask and reshape + compressed_output = padded_ctc_output[total_mask].reshape( + ctc_output.shape[0], -1, ctc_output.shape[2] + ) + + return compressed_output + def forward( self, fbank: torch.Tensor, @@ -238,7 +292,7 @@ class SPEECH_LLM(nn.Module): attention_mask: torch.Tensor, labels: torch.LongTensor, ): - encoder_outs, _ = self.forward_encoder(fbank, fbank_lens) + encoder_outs, encoder_out_lens = self.forward_encoder(fbank, fbank_lens) speech_features = self.encoder_projector(encoder_outs) From 39d90356fe97f68d651996cd41a1e22bed0bdd0f Mon Sep 17 00:00:00 2001 From: yfyeung Date: Wed, 18 Jun 2025 04:44:10 +0000 Subject: [PATCH 21/22] fix deepspeed config fix --- egs/speech_llm/ASR_LLM/whisper_llm_zh/ds_config_zero1.json | 2 +- egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py | 1 - 2 files changed, 1 insertion(+), 2 deletions(-) diff --git a/egs/speech_llm/ASR_LLM/whisper_llm_zh/ds_config_zero1.json b/egs/speech_llm/ASR_LLM/whisper_llm_zh/ds_config_zero1.json index 730937a21..29e710e3c 100644 --- a/egs/speech_llm/ASR_LLM/whisper_llm_zh/ds_config_zero1.json +++ b/egs/speech_llm/ASR_LLM/whisper_llm_zh/ds_config_zero1.json @@ -5,7 +5,7 @@ "loss_scale_window": 100, "initial_scale_power": 16, "hysteresis": 2, - "min_loss_scale": 0.01 + "min_loss_scale": 1 }, "zero_optimization": { "stage": 1, diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py index d585ec871..b7ad888cd 100644 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/model.py @@ -64,7 +64,6 @@ class SPEECH_LLM(nn.Module): self, encoder_embed: nn.Module, encoder: EncoderInterface, - ctc_output: nn.Module, llm: nn.Module, encoder_projector: nn.Module, ): From 53111d0e4670fc51d22d42b184595ecf84940bea Mon Sep 17 00:00:00 2001 From: yfyeung Date: Wed, 18 Jun 2025 07:33:15 +0000 Subject: [PATCH 22/22] fix for multigpu --- egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py index c668e7f64..7565dd98b 100755 --- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py +++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py @@ -693,6 +693,9 @@ def train_one_epoch( exclude_frozen_parameters=True, ) + if world_size > 1: + torch.distributed.barrier() + if rank == 0: convert_zero_checkpoint_to_fp32_state_dict( params.exp_dir, @@ -710,6 +713,9 @@ def train_one_epoch( f"rm -rf {params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}" ) + if world_size > 1: + torch.distributed.barrier() + shave_rate = params.shave_rate while True: try: @@ -991,6 +997,10 @@ def run(rank, world_size, args): client_state={}, exclude_frozen_parameters=True, ) + + if world_size > 1: + torch.distributed.barrier() + if rank == 0: convert_zero_checkpoint_to_fp32_state_dict( params.exp_dir,