add whisper llm

egs/speech_llm/ASR_LLM/README.md

@@ -0,0 +1,39 @@
+
+# Introduction
+
+This recipe includes scripts for training Zipformer model using multiple Chinese datasets.
+
+# Included Training Sets
+1. THCHS-30
+2. AiShell-{1,2,4}
+3. ST-CMDS
+4. Primewords
+5. MagicData
+6. Aidatatang_200zh
+7. AliMeeting
+8. WeNetSpeech
+9. KeSpeech-ASR
+
+|Datset| Number of hours| URL|
+|---|---:|---|
+|**TOTAL**|14,106|---|
+|THCHS-30|35|https://www.openslr.org/18/|
+|AiShell-1|170|https://www.openslr.org/33/|
+|AiShell-2|1,000|http://www.aishelltech.com/aishell_2|
+|AiShell-4|120|https://www.openslr.org/111/|
+|ST-CMDS|110|https://www.openslr.org/38/|
+|Primewords|99|https://www.openslr.org/47/|
+|aidatatang_200zh|200|https://www.openslr.org/62/|
+|MagicData|755|https://www.openslr.org/68/|
+|AliMeeting|100|https://openslr.org/119/|
+|WeNetSpeech|10,000|https://github.com/wenet-e2e/WenetSpeech|
+|KeSpeech|1,542|https://github.com/KeSpeech/KeSpeech|
+
+
+# Included Test Sets
+1. Aishell-{1,2,4}
+2. Aidatatang_200zh
+3. AliMeeting
+4. MagicData
+5. KeSpeech-ASR
+6. WeNetSpeech

egs/speech_llm/ASR_LLM/RESULTS.md

@@ -0,0 +1,116 @@
+## Results
+
+### Multi Chinese datasets (without datatang 200h) finetuning results on Whisper-large-v2
+#### Whisper
+[./whisper](./whisper)
+
+Character Error Rates (CERs) listed below are produced by the checkpoint of the second epoch using greedy search.
+
+| Datasets | alimeeting | alimeeting | aishell-1 | aishell-1 | aishell-2 | aishell-2 | aishell-4 | magicdata | magicdata | kespeech-asr | kespeech-asr | kespeech-asr | WenetSpeech  |
+|--------------------------------|-------------------|--------------|----------------|-------------|------------------|-------------|------------------|------------------|-------------|-----------------------|-----------------------|-------------|-------------------|
+|  Split |           eval| test | dev | test | dev| test | test      | dev| test | dev phase1 | dev phase2 | test | test meeting |
+| Greedy Search |  23.22 | 28.24 | 0.61 | 0.66 | 2.67 | 2.80 | 16.61 | 2.56 | 2.21 | 4.73 | 1.90 | 5.98 |                    8.13 |
+
+Command for training is:
+```bash
+pip install -r whisper/requirements.txt
+
+# We updated the label of wenetspeech to remove OCR deletion errors, see https://github.com/wenet-e2e/WenetSpeech/discussions/54
+
+torchrun --nproc-per-node 8 ./whisper/train.py \
+  --max-duration 200 \
+  --exp-dir whisper/exp_large_v2 \
+  --model-name large-v2 \
+  --deepspeed \
+  --deepspeed_config ./whisper/ds_config_zero1.json
+```
+
+Command for decoding using fine-tuned models:
+```bash
+git lfs install
+git clone https://huggingface.co/yuekai/icefall_asr_multi-hans-zh_whisper
+ln -s icefall_asr_multi-hans-zh_whisper/v1.1/epoch-3-avg-10.pt whisper/exp_large_v2/epoch-999.pt
+
+python3 ./whisper/decode.py \
+  --exp-dir whisper/exp_large_v2 \
+  --model-name large-v2 \
+  --epoch 999 --avg 1 \
+  --beam-size 10 --max-duration 50
+```
+
+Fine-tuned models, training logs, decoding logs, tensorboard and decoding results
+are available at
+<https://huggingface.co/yuekai/icefall_asr_multi-hans-zh_whisper>
+
+
+### Multi Chinese datasets char-based training results (Non-streaming) on zipformer model
+
+This is the [pull request #1238](https://github.com/k2-fsa/icefall/pull/1238) in icefall.
+
+#### Non-streaming (with CTC head)
+
+Best results (num of params : ~69M):
+
+The training command:
+
+```
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 20 \
+  --use-fp16 1 \
+  --max-duration 600 \
+  --num-workers 8 \
+  --use-ctc 1
+```
+
+The decoding command:
+
+```
+./zipformer/decode.py \
+  --epoch 20 \
+  --avg 1 \
+  --use-ctc 1
+```
+
+Character Error Rates (CERs) listed below are produced by the checkpoint of the 20th epoch using BPE model ( # tokens is 2000, byte fallback enabled).
+
+| 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 |
+| 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 |
+
+Pre-trained model can be found here : https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24/
+
+#### Non-streaming
+
+Best results (num of params : ~69M):
+
+The training command:
+
+```
+./zipformer/train.py \
+  --world-size 4 \
+  --num-epochs 20 \
+  --use-fp16 1 \
+  --max-duration 600 \
+  --num-workers 8
+```
+
+The decoding command:
+
+```
+./zipformer/decode.py \
+  --epoch 20 \
+  --avg 1
+```
+
+Character Error Rates (CERs) listed below are produced by the checkpoint of the 20th epoch using greedy search and BPE model ( # tokens is 2000, byte fallback enabled).
+
+| 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 |
+| 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 |
+
+
+Pre-trained model can be found here : https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-2023-9-2/

egs/speech_llm/ASR_LLM/run.sh

@@ -0,0 +1,15 @@
+
+export PYTHONPATH=$PYTHONPATH:/workspace/asr/icefall
+# pip install k2==1.24.3.dev20230524+cuda11.8.torch2.0.1 -f https://k2-fsa.github.io/k2/cuda.html
+# pip install -r whisper/requirements.txt
+
+method=mask_predict
+# method=cif_ar_distill_embedding
+torchrun --nproc_per_node 8 ./parawhisper/train.py \
+  --max-duration 200 \
+  --exp-dir parawhisper/exp_large_v2_${method} \
+  --model-name large-v2 \
+  --manifest-dir data/fbank \
+  --method $method \
+  --deepspeed \
+  --deepspeed_config ./whisper/ds_config_zero1.json

egs/speech_llm/ASR_LLM/whisper_llm_zh/asr_datamodule.py

@@ -0,0 +1 @@
+../../../multi_zh-hans/ASR/zipformer/asr_datamodule.py

egs/speech_llm/ASR_LLM/whisper_llm_zh/decode.py

@@ -0,0 +1,567 @@
+#!/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:
+git lfs install
+git clone https://huggingface.co/yuekai/icefall_asr_aishell_whisper
+ln -s icefall_asr_aishell_whisper/exp_large_v2/epoch-10-avg6.pt whisper/exp_large_v2/epoch-999.pt
+
+python3 ./whisper/decode.py \
+  --exp-dir whisper/exp_large_v2 \
+  --model-name large-v2 \
+  --epoch 999 --avg 1 \
+  --beam-size 10 --max-duration 50
+
+# Command for decoding using pretrained models (before fine-tuning):
+
+python3 ./whisper/decode.py \
+  --exp-dir whisper/exp_large_v2 \
+  --model-name large-v2 \
+  --epoch -1 --avg 1 \
+  --remove-whisper-encoder-input-length-restriction False \
+  --beam-size 10 --max-duration 50
+
+"""
+
+import argparse
+import logging
+import re
+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 whisper
+from asr_datamodule import AsrDataModule
+from lhotse.cut import Cut
+from multi_dataset import MultiDataset
+from tn.chinese.normalizer import Normalizer
+from whisper.normalizers import BasicTextNormalizer
+from whisper_decoder_forward_monkey_patch import replace_whisper_decoder_forward
+from whisper_encoder_forward_monkey_patch import replace_whisper_encoder_forward
+from zhconv import convert
+
+from icefall.checkpoint import average_checkpoints_with_averaged_model, 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 remove_punctuation(text: str or List[str]):
+    """Modified from https://github.com/yeyupiaoling/Whisper-Finetune/blob/master/utils/data_utils.py
+
+    Args:
+        text: It can be a string or a list of strings.
+    Returns:
+        Return a string or a list of strings without any punctuation.
+    """
+    punctuation = "!,.;:?、！，。；：？《》 "
+    if isinstance(text, str):
+        text = re.sub(r"[{}]+".format(punctuation), "", text).strip()
+        return text
+    elif isinstance(text, list):
+        result_text = []
+        for t in text:
+            t = re.sub(r"[{}]+".format(punctuation), "", t).strip()
+            result_text.append(t)
+        return result_text
+    else:
+        raise Exception(f"Not support type {type(text)}")
+
+
+def to_simple(text: str or List[str]):
+    """Convert traditional Chinese to simplified Chinese.
+    Args:
+        text: It can be a string or a list of strings.
+    Returns:
+        Return a string or a list of strings converted to simplified Chinese.
+    """
+    if isinstance(text, str):
+        text = convert(text, "zh-cn")
+        return text
+    elif isinstance(text, list):
+        result_text = []
+        for t in text:
+            t = convert(t, "zh-cn")
+            result_text.append(t)
+        return result_text
+    else:
+        raise Exception(f"Not support type{type(text)}")
+
+
+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(
+        "--model-name",
+        type=str,
+        default="large-v2",
+        choices=["large-v2", "large-v3", "medium", "base", "small", "tiny"],
+        help="""The model name to use.
+        """,
+    )
+
+    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(
+        "--use-distill-whisper",
+        type=str2bool,
+        default=False,
+        help="Whether to use architecture of distill whisper.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    params = AttributeDict(
+        {
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    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".
+    """
+    dtype = torch.float16
+    device = torch.device("cuda")
+
+    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)
+    results = model.decode(feature, params.decoding_options)
+    hyps = [result.text for result in results]
+
+    hyps = remove_punctuation(hyps)
+    hyps = to_simple(hyps)
+    hyps = [params.normalizer.normalize(hyp) for hyp in hyps]
+    print(hyps)
+    return {"beam-search": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+) -> 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".
+    """
+
+    def normalize_text_alimeeting(text: str, normalize: str = "m2met") -> str:
+        """
+        Text normalization similar to M2MeT challenge baseline.
+        See: https://github.com/yufan-aslp/AliMeeting/blob/main/asr/local/text_normalize.pl
+        """
+        if normalize == "none":
+            return text
+        elif normalize == "m2met":
+            import re
+
+            text = text.replace(" ", "")
+            text = text.replace("<sil>", "")
+            text = text.replace("<%>", "")
+            text = text.replace("<->", "")
+            text = text.replace("<$>", "")
+            text = text.replace("<#>", "")
+            text = text.replace("<_>", "")
+            text = text.replace("<space>", "")
+            text = text.replace("`", "")
+            text = text.replace("&", "")
+            text = text.replace(",", "")
+            if re.search("[a-zA-Z]", text):
+                text = text.upper()
+            text = text.replace("Ａ", "A")
+            text = text.replace("ａ", "A")
+            text = text.replace("ｂ", "B")
+            text = text.replace("ｃ", "C")
+            text = text.replace("ｋ", "K")
+            text = text.replace("ｔ", "T")
+            text = text.replace("，", "")
+            text = text.replace("丶", "")
+            text = text.replace("。", "")
+            text = text.replace("、", "")
+            text = text.replace("？", "")
+            return text
+
+    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"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            batch=batch,
+        )
+
+        for lm_scale, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_text = normalize_text_alimeeting(ref_text)
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            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[str, List[str], List[str]]]],
+):
+
+    enable_log = True
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.exp_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        if enable_log:
+            logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.exp_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        # we compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=enable_log
+            )
+            test_set_wers[key] = wer
+
+        if enable_log:
+            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.exp_dir / f"cer-summary-{test_set_name}-{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.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    setup_logger(
+        f"{params.exp_dir}/log-{params.method}-beam{params.beam_size}/log-decode-{params.suffix}"
+    )
+
+    options = whisper.DecodingOptions(
+        task="transcribe",
+        language="zh",
+        without_timestamps=True,
+        beam_size=params.beam_size,
+    )
+    params.decoding_options = options
+    params.cleaner = BasicTextNormalizer()
+    params.normalizer = Normalizer()
+
+    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()
+    if params.use_distill_whisper:
+        replace_whisper_decoder_forward()
+    model = whisper.load_model(params.model_name, "cpu")
+    if params.epoch > 0:
+        if params.avg > 1:
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            checkpoint = torch.load(
+                f"{params.exp_dir}/epoch-{params.epoch}.pt", map_location="cpu"
+            )
+            if "model" not in checkpoint:
+                # deepspeed converted checkpoint only contains model state_dict
+                filenames = [
+                    f"{params.exp_dir}/epoch-{epoch}.pt"
+                    for epoch in range(start, params.epoch + 1)
+                ]
+                model.load_state_dict(average_checkpoints(filenames))
+            else:
+                filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+                filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+                logging.info(
+                    f"Calculating the averaged model over epoch range from "
+                    f"{start} (excluded) to {params.epoch}"
+                )
+                model.to(device)
+                model.load_state_dict(
+                    average_checkpoints_with_averaged_model(
+                        filename_start=filename_start,
+                        filename_end=filename_end,
+                        device=device,
+                    )
+                )
+            # save checkpoints
+            filename = f"{params.exp_dir}/epoch-{params.epoch}-avg-{params.avg}.pt"
+            torch.save(model.state_dict(), filename)
+        else:
+            checkpoint = torch.load(
+                f"{params.exp_dir}/epoch-{params.epoch}.pt", map_location="cpu"
+            )
+            if "model" not in checkpoint:
+                model.load_state_dict(checkpoint, strict=True)
+            else:
+                load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    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
+
+    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,
+        )
+
+        save_results(params=params, test_set_name=test_set, results_dict=results_dict)
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()

egs/speech_llm/ASR_LLM/whisper_llm_zh/ds_config_zero1.json

@@ -0,0 +1 @@
+../../../aishell/ASR/whisper/ds_config_zero1.json

egs/speech_llm/ASR_LLM/whisper_llm_zh/label_smoothing.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/conformer_ctc/label_smoothing.py

egs/speech_llm/ASR_LLM/whisper_llm_zh/model.py

@@ -0,0 +1,145 @@
+from torch import nn
+import torch
+
+DEFAULT_SPEECH_TOKEN = -1997 # "<speech>"
+
+class EncoderProjector(nn.Module):
+
+    def __init__(self, encoder_dim, llm_dim):
+        super().__init__()
+        self.linear1 = nn.Linear(encoder_dim, llm_dim)
+        self.relu = nn.ReLU()
+        self.linear2 = nn.Linear(llm_dim, llm_dim)
+
+    def forward(self, x):    
+        x = self.linear1(x)
+        x = self.relu(x)
+        x = self.linear2(x)
+        return x
+
+class SPEECH_LLM(nn.Module):
+    # https://github.com/ddlBoJack/SLAM-LLM/blob/main/src/slam_llm/models/slam_model.py
+    def __init__(
+        self,
+        encoder: nn.Module,
+        llm: nn.Module,
+        encoder_projector: nn.Module,
+    ):
+        super().__init__()
+
+        self.encoder = encoder
+        self.encoder.eval()
+        self.llm = llm
+        self.llm.eval()
+        self.encoder_projector = encoder_projector
+        self.encoder_outputs_downsample_rate = 4
+
+    def _merge_input_ids_with_speech_features(self, speech_features, inputs_embeds, input_ids, attention_mask, labels):
+        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 == DEFAULT_SPEECH_TOKEN
+        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 != DEFAULT_SPEECH_TOKEN)
+
+        # 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), self., 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" "<speech>", "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)
+        # downsample encoder_outs by 4
+        encoder_outs = encoder_outs[:, ::self.encoder_outputs_downsample_rate]
+        speech_features = self.encoder_projector(encoder_outs)
+
+        inputs_embeds = self.llm.get_input_embeddings()(input_ids)
+        inputs_embeds, attention_mask, labels, position_ids = self._merge_input_ids_with_speech_features(
+            speech_features, inputs_embeds, input_ids, attention_mask, labels
+        )
+
+        outputs = self.language_model(
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        logits = outputs[0]
+
+        model_outputs = self.llm(inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels, position_ids=position_ids)
+
+        return model_outputs

egs/speech_llm/ASR_LLM/whisper_llm_zh/multi_dataset.py

@@ -0,0 +1,258 @@
+# Copyright      2023  Xiaomi Corp.        (authors: Zengrui Jin)
+#
+# 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.
+
+
+import glob
+import logging
+import re
+from pathlib import Path
+from typing import Dict, List
+
+import lhotse
+from lhotse import CutSet, load_manifest_lazy
+
+
+class MultiDataset:
+    def __init__(self, fbank_dir: str):
+        """
+        Args:
+          manifest_dir:
+            It is expected to contain the following files:
+            - aishell_cuts_train.jsonl.gz
+            - aishell2_cuts_train.jsonl.gz
+            - aishell4_cuts_train_L.jsonl.gz
+            - aishell4_cuts_train_M.jsonl.gz
+            - aishell4_cuts_train_S.jsonl.gz
+            - alimeeting-far_cuts_train.jsonl.gz
+            - magicdata_cuts_train.jsonl.gz
+            - primewords_cuts_train.jsonl.gz
+            - stcmds_cuts_train.jsonl.gz
+            - thchs_30_cuts_train.jsonl.gz
+            - kespeech/kespeech-asr_cuts_train_phase1.jsonl.gz
+            - kespeech/kespeech-asr_cuts_train_phase2.jsonl.gz
+            - wenetspeech/cuts_L_fixed.jsonl.gz
+        """
+        self.fbank_dir = Path(fbank_dir)
+
+    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),
+            ],
+        )
+
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get multidataset dev cuts")
+
+        # WeNetSpeech
+        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_cuts
+
+    def test_cuts(self) -> Dict[str, CutSet]:
+        logging.info("About to get multidataset test cuts")
+
+        # AISHELL
+        logging.info("Loading Aishell set in lazy mode")
+        aishell_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell_cuts_test.jsonl.gz"
+        )
+        aishell_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell_cuts_dev.jsonl.gz"
+        )
+
+        # AISHELL-2
+        logging.info("Loading Aishell-2 set in lazy mode")
+        aishell2_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell2_cuts_test.jsonl.gz"
+        )
+        aishell2_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell2_cuts_dev.jsonl.gz"
+        )
+
+        # AISHELL-4
+        logging.info("Loading Aishell-4 TEST set in lazy mode")
+        aishell4_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell4_cuts_test.jsonl.gz"
+        )
+
+        # Ali-Meeting
+        logging.info("Loading Ali-Meeting set in lazy mode")
+        alimeeting_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "alimeeting-far_cuts_test.jsonl.gz"
+        )
+        alimeeting_eval_cuts = load_manifest_lazy(
+            self.fbank_dir / "alimeeting-far_cuts_eval.jsonl.gz"
+        )
+
+        # MagicData
+        logging.info("Loading MagicData set in lazy mode")
+        magicdata_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "magicdata_cuts_test.jsonl.gz"
+        )
+        magicdata_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "magicdata_cuts_dev.jsonl.gz"
+        )
+
+        # KeSpeech
+        logging.info("Loading KeSpeech set in lazy mode")
+        kespeech_test_cuts = load_manifest_lazy(
+            self.fbank_dir / "kespeech" / "kespeech-asr_cuts_test.jsonl.gz"
+        )
+        kespeech_dev_phase1_cuts = load_manifest_lazy(
+            self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase1.jsonl.gz"
+        )
+        kespeech_dev_phase2_cuts = load_manifest_lazy(
+            self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase2.jsonl.gz"
+        )
+
+        # WeNetSpeech
+        logging.info("Loading WeNetSpeech set in lazy mode")
+        wenetspeech_test_meeting_cuts = load_manifest_lazy(
+            self.fbank_dir / "wenetspeech" / "cuts_TEST_MEETING.jsonl.gz"
+        )
+        wenetspeech_test_net_cuts = load_manifest_lazy(
+            self.fbank_dir / "wenetspeech" / "cuts_TEST_NET.jsonl.gz"
+        )
+        wenetspeech_dev_cuts = load_manifest_lazy(
+            self.fbank_dir / "wenetspeech" / "cuts_DEV_fixed.jsonl.gz"
+        )
+
+        return {
+            "wenetspeech-meeting_test": wenetspeech_test_meeting_cuts,
+            # "aishell_test": aishell_test_cuts,
+            # "aishell_dev": aishell_dev_cuts,
+            # "ali-meeting_test": alimeeting_test_cuts,
+            # "ali-meeting_eval": alimeeting_eval_cuts,
+            # "aishell-4_test": aishell4_test_cuts,
+            # "aishell-2_test": aishell2_test_cuts,
+            # "aishell-2_dev": aishell2_dev_cuts,
+            # "magicdata_test": magicdata_test_cuts,
+            # "magicdata_dev": magicdata_dev_cuts,
+            # "kespeech-asr_test": kespeech_test_cuts,
+            # "kespeech-asr_dev_phase1": kespeech_dev_phase1_cuts,
+            # "kespeech-asr_dev_phase2": kespeech_dev_phase2_cuts,
+            # "wenetspeech-net_test": wenetspeech_test_net_cuts,
+            # "wenetspeech_dev": wenetspeech_dev_cuts,
+        }
+
+    def aishell_train_cuts(self) -> CutSet:
+        logging.info("About to get multidataset train cuts")
+
+        # AISHELL-1
+        logging.info("Loading Aishell-1 in lazy mode")
+        aishell_cuts = load_manifest_lazy(
+            self.fbank_dir / "aishell_cuts_train.jsonl.gz"
+        )
+
+        return aishell_cuts

egs/speech_llm/ASR_LLM/whisper_llm_zh/train.py

@@ -0,0 +1,874 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#              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:
+
+#fine-tuning with deepspeed zero stage 1
+torchrun --nproc-per-node 8 ./whisper/train.py \
+  --max-duration 200 \
+  --exp-dir whisper/exp_large_v2 \
+  --model-name large-v2 \
+  --deepspeed \
+  --deepspeed_config ./whisper/ds_config_zero1.json
+"""
+
+import argparse
+import copy
+import logging
+import os
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import deepspeed
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+import whisper
+from asr_datamodule import AsrDataModule
+from model import SPEECH_LLM, EncoderProjector
+from deepspeed.utils.zero_to_fp32 import convert_zero_checkpoint_to_fp32_state_dict
+from label_smoothing import LabelSmoothingLoss
+from lhotse import CutSet, load_manifest
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from multi_dataset import MultiDataset
+# from optim import Eden, ScaledAdam
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.functional import pad as pad_tensor
+# from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from whisper_decoder_forward_monkey_patch import replace_whisper_decoder_forward
+from whisper_encoder_forward_monkey_patch import replace_whisper_encoder_forward
+
+from icefall import diagnostics
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import update_averaged_model
+from icefall.dist import cleanup_dist, get_rank, get_world_size, setup_dist
+from icefall.env import get_env_info
+from icefall.hooks import register_inf_check_hooks
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    filter_uneven_sized_batch,
+    setup_logger,
+    str2bool,
+)
+
+from transformers import AutoModelForCausalLM, AutoTokenizer
+LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]
+
+
+def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
+    if isinstance(model, DDP):
+        # get underlying nn.Module
+        model = model.module
+    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=4,
+        help="Downsample rate for the encoder projector.",
+    )
+
+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(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.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(
+        "--base-lr", type=float, default=1e-5, help="The base learning rate."
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--inf-check",
+        type=str2bool,
+        default=False,
+        help="Add hooks to check for infinite module outputs and gradients.",
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=200,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=True,
+        help="Whether to use half precision 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,
+            "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": 10000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+# def load_checkpoint_if_available(
+#     params: AttributeDict,
+#     model: nn.Module,
+#     model_avg: nn.Module = None,
+#     optimizer: Optional[torch.optim.Optimizer] = None,
+#     scheduler: Optional[LRSchedulerType] = None,
+# ) -> Optional[Dict[str, Any]]:
+#     """Load checkpoint from file.
+
+#     If params.start_batch is positive, it will load the checkpoint from
+#     `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+#     params.start_epoch is larger than 1, it will load the checkpoint from
+#     `params.start_epoch - 1`.
+
+#     Apart from loading state dict for `model` and `optimizer` it also updates
+#     `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+#     and `best_valid_loss` in `params`.
+
+#     Args:
+#       params:
+#         The return value of :func:`get_params`.
+#       model:
+#         The training model.
+#       model_avg:
+#         The stored model averaged from the start of training.
+#       optimizer:
+#         The optimizer that we are using.
+#       scheduler:
+#         The scheduler that we are using.
+#     Returns:
+#       Return a dict containing previously saved training info.
+#     """
+#     if params.start_batch > 0:
+#         filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+#     elif params.start_epoch > 1:
+#         filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+#     else:
+#         return None
+
+#     assert filename.is_file(), f"{filename} does not exist!"
+
+#     saved_params = load_checkpoint(
+#         filename,
+#         model=model,
+#         model_avg=model_avg,
+#         optimizer=optimizer,
+#         scheduler=scheduler,
+#     )
+
+#     keys = [
+#         "best_train_epoch",
+#         "best_valid_epoch",
+#         "batch_idx_train",
+#         "best_train_loss",
+#         "best_valid_loss",
+#     ]
+#     for k in keys:
+#         params[k] = saved_params[k]
+
+#     if params.start_batch > 0:
+#         if "cur_epoch" in saved_params:
+#             params["start_epoch"] = saved_params["cur_epoch"]
+
+#     return saved_params
+
+
+def compute_loss(
+    params: AttributeDict,
+    tokenizer: AutoTokenizer,
+    model: Union[nn.Module, DDP],
+    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.
+    """
+    # For the uneven-sized batch, the total duration after padding would possibly
+    # cause OOM. Hence, for each batch, which is sorted descendingly by length,
+    # we simply drop the last few shortest samples, so that the retained total frames
+    # (after padding) would not exceed `allowed_max_frames`:
+    # `allowed_max_frames = int(max_frames * (1.0 + allowed_excess_duration_ratio))`,
+    # where `max_frames = max_duration * 1000 // frame_shift_ms`.
+    # We set allowed_excess_duration_ratio=0.1.
+
+    def preprocess(
+        messages,
+        tokenizer: transformers.PreTrainedTokenizer,
+        max_len: int,
+    ) -> Dict:
+        """Preprocesses the data for supervised fine-tuning."""
+        texts = []
+        for i, msg in enumerate(messages):
+            texts.append(
+                tokenizer.apply_chat_template(
+                    msg,
+                    chat_template=TEMPLATE,
+                    tokenize=True,
+                    add_generation_prompt=False,
+                    padding="max_length",
+                    max_length=max_len,
+                    truncation=True,
+                )
+            )
+        # model_inputs = tokenizer([text], return_tensors="pt").to(device)
+        input_ids = torch.tensor(texts, dtype=torch.int)
+        target_ids = input_ids.clone()
+        target_ids[target_ids == tokenizer.pad_token_id] = IGNORE_TOKEN_ID
+        attention_mask = input_ids.ne(tokenizer.pad_token_id)
+
+        return input_ids, attention_mask, target_ids
+
+    def normalize_text_alimeeting(text: str, normalize: str = "m2met") -> str:
+        """
+        Text normalization similar to M2MeT challenge baseline.
+        See: https://github.com/yufan-aslp/AliMeeting/blob/main/asr/local/text_normalize.pl
+        """
+        if normalize == "none":
+            return text
+        elif normalize == "m2met":
+            import re
+
+            text = text.replace(" ", "")
+            text = text.replace("<sil>", "")
+            text = text.replace("<%>", "")
+            text = text.replace("<->", "")
+            text = text.replace("<$>", "")
+            text = text.replace("<#>", "")
+            text = text.replace("<_>", "")
+            text = text.replace("<space>", "")
+            text = text.replace("`", "")
+            text = text.replace("&", "")
+            text = text.replace(",", "")
+            if re.search("[a-zA-Z]", text):
+                text = text.upper()
+            text = text.replace("Ａ", "A")
+            text = text.replace("ａ", "A")
+            text = text.replace("ｂ", "B")
+            text = text.replace("ｃ", "C")
+            text = text.replace("ｋ", "K")
+            text = text.replace("ｔ", "T")
+            text = text.replace("，", "")
+            text = text.replace("丶", "")
+            text = text.replace("。", "")
+            text = text.replace("、", "")
+            text = text.replace("？", "")
+            return text
+
+    max_frames = params.max_duration * 1000 // params.frame_shift_ms
+    allowed_max_frames = int(max_frames * (1.0 + params.allowed_excess_duration_ratio))
+    batch = filter_uneven_sized_batch(batch, allowed_max_frames)
+
+    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
+    feature = batch["inputs"]
+
+    assert feature.ndim == 3
+    feature = feature.to(device)
+    feature = feature.transpose(1, 2)  # (N, C, T)
+
+    # feature_lens = supervisions["num_frames"].to(device)
+
+    batch_idx_train = params.batch_idx_train
+    supervisions = batch["supervisions"]
+    texts = batch["supervisions"]["text"]
+    # remove spaces in texts
+    texts = [normalize_text_alimeeting(text) for text in texts]
+
+    #     messages = [
+    #     {"role": "system", "content": "You are a helpful assistant."},
+    #     {"role": "user", "content": prompt}
+    # ]
+    input_ids, attention_mask, target_ids = preprocess(
+        texts, tokenizer, max_len=512
+    )
+
+    # decoder_criterion = LabelSmoothingLoss(
+    #     ignore_index=50256, label_smoothing=0.1, reduction="sum"
+    # )
+
+    # # ignore the first 3 tokens, which are always <|lang_id|>, <|transcibe|>, <|notimestampes|>
+    # ignore_prefix_size = 3
+    # with torch.set_grad_enabled(is_training):
+    #     encoder_out = model.encoder(feature)
+    #     text_logits = model.decoder(prev_outputs_tokens.to(device), encoder_out)
+    #     text_logits = text_logits[:, ignore_prefix_size:, :]
+    #     target_tokens = target_tokens[:, ignore_prefix_size:]
+    #     loss = decoder_criterion(text_logits, target_tokens.to(device))
+
+    # assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    tokenizer: whisper.tokenizer.Tokenizer,
+    model: Union[nn.Module, DDP],
+    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: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    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.encoder_projector.train()
+
+    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 and not params.print_diagnostics:
+            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()
+            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
+                )
+
+            model.save_checkpoint(
+                save_dir=params.exp_dir,
+                tag=f"epoch-{params.cur_epoch}-checkpoint-{batch_idx}",
+                client_state={"sampler": train_dl.sampler.state_dict()},
+            )
+            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}",
+                )
+                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")
+    
+    if 'whisper' in params.speech_encoder_path_or_name:
+        replace_whisper_encoder_forward()
+        # TODO: directly loading from whisper-ft checkpoint
+        # whisper-large-v2-multi-hans-zh-epoch-3-avg-10.pt
+        speech_encoder = whisper.load_model(params.model_name, "cpu").encoder
+        speech_encoder_dim = speech_encoder.dims.n_audio_ctx
+
+    llm = AutoModelForCausalLM.from_pretrained(
+        params.llm_path_or_name,
+        attn_implemented="flash_attention_2",
+        device_map="cpu"
+    )
+    tokenizer = AutoTokenizer.from_pretrained(params.llm_path_or_name)
+
+    encoder_projector = EncoderProjector(speech_encoder_dim, llm.config.hidden_size)
+
+    model = SPEECH_LLM(
+        speech_encoder,
+        llm,
+        encoder_projector,
+    )
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    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.start_epoch > 0, params.start_epoch
+    # checkpoints = load_checkpoint_if_available(
+    #     params=params, model=model, model_avg=model_avg
+    # )
+
+    assert params.deepspeed and world_size > 1
+    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
+
+    # train_cuts = multi_dataset.train_cuts()
+    train_cuts = multi_dataset.aishell_train_cuts()
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    # if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+    #     # We only load the sampler's state dict when it loads a checkpoint
+    #     # saved in the middle of an epoch
+    #     sampler_state_dict = checkpoints["sampler"]
+    # else:
+    #     sampler_state_dict = None
+    # TODO: load sampler state dict
+    train_dl = data_module.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    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
+
+    # if params.pretrained_model_path:
+    #     checkpoint = torch.load(params.pretrained_model_path, map_location="cpu")
+    #     if "model" not in checkpoint:
+    #         model.load_state_dict(checkpoint, strict=True)
+    #     else:
+    #         load_checkpoint(params.pretrained_model_path, model)
+
+    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"epoch-{params.cur_epoch}",
+            client_state={"sampler": train_dl.sampler.state_dict()},
+        )
+        if rank == 0:
+            convert_zero_checkpoint_to_fp32_state_dict(
+                params.exp_dir,
+                f"{params.exp_dir}/epoch-{params.cur_epoch}.pt",
+                tag=f"epoch-{params.cur_epoch}",
+            )
+            os.system(f"rm -rf {params.exp_dir}/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)
+    run(rank=rank, world_size=world_size, args=args)
+
+
+if __name__ == "__main__":
+    main()

egs/speech_llm/ASR_LLM/whisper_llm_zh/whisper_encoder_forward_monkey_patch.py

@@ -0,0 +1 @@
+../../../aishell/ASR/whisper/whisper_encoder_forward_monkey_patch.py