diff --git a/ckpts/qwen3-1.7b-whisper-260126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/ckpts/global_step5000/mp_rank_00_model_states.pt b/ckpts/qwen3-1.7b-whisper-260126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/ckpts/global_step5000/mp_rank_00_model_states.pt new file mode 100644 index 0000000000000000000000000000000000000000..e9fed876e342c23634654391f7036cef356cec9c --- /dev/null +++ b/ckpts/qwen3-1.7b-whisper-260126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/ckpts/global_step5000/mp_rank_00_model_states.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:add229c9677d024a7cfe5c813eca04b92142d74e89a2a68bb3b7575092543bad +size 10042805638 diff --git a/ckpts/qwen3-1.7b-whisper-260126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/ckpts/latest b/ckpts/qwen3-1.7b-whisper-260126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/ckpts/latest new file mode 100644 index 0000000000000000000000000000000000000000..f805186fa43374540c3fa51dfd3cca9ac06e56a5 --- /dev/null +++ b/ckpts/qwen3-1.7b-whisper-260126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/ckpts/latest @@ -0,0 +1 @@ +global_step5000 \ No newline at end of file diff --git a/ckpts/qwen3-1.7b-whisper-260126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/ckpts/zero_to_fp32.py b/ckpts/qwen3-1.7b-whisper-260126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/ckpts/zero_to_fp32.py new file mode 100644 index 0000000000000000000000000000000000000000..24cc342e78d1a006c782b3a4cd68d9ce786d8fd8 --- /dev/null +++ b/ckpts/qwen3-1.7b-whisper-260126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/ckpts/zero_to_fp32.py @@ -0,0 +1,604 @@ +#!/usr/bin/env python + +# Copyright (c) Microsoft Corporation. +# SPDX-License-Identifier: Apache-2.0 + +# DeepSpeed Team + +# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets +# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in +# the future. Once extracted, the weights don't require DeepSpeed and can be used in any +# application. +# +# example: python zero_to_fp32.py . pytorch_model.bin + +import argparse +import torch +import glob +import math +import os +import re +from collections import OrderedDict +from dataclasses import dataclass + +# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with +# DeepSpeed data structures it has to be available in the current python environment. +from deepspeed.utils import logger +from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS, + FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES, + FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS) + + +@dataclass +class zero_model_state: + buffers: dict() + param_shapes: dict() + shared_params: list + ds_version: int + frozen_param_shapes: dict() + frozen_param_fragments: dict() + + +debug = 0 + +# load to cpu +device = torch.device('cpu') + + +def atoi(text): + return int(text) if text.isdigit() else text + + +def natural_keys(text): + ''' + alist.sort(key=natural_keys) sorts in human order + http://nedbatchelder.com/blog/200712/human_sorting.html + (See Toothy's implementation in the comments) + ''' + return [atoi(c) for c in re.split(r'(\d+)', text)] + + +def get_model_state_file(checkpoint_dir, zero_stage): + if not os.path.isdir(checkpoint_dir): + raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist") + + # there should be only one file + if zero_stage <= 2: + file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt") + elif zero_stage == 3: + file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt") + + if not os.path.exists(file): + raise FileNotFoundError(f"can't find model states file at '{file}'") + + return file + + +def get_checkpoint_files(checkpoint_dir, glob_pattern): + # XXX: need to test that this simple glob rule works for multi-node setup too + ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys) + + if len(ckpt_files) == 0: + raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'") + + return ckpt_files + + +def get_optim_files(checkpoint_dir): + return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt") + + +def get_model_state_files(checkpoint_dir): + return get_checkpoint_files(checkpoint_dir, "*_model_states.pt") + + +def parse_model_states(files): + zero_model_states = [] + for file in files: + state_dict = torch.load(file, map_location=device) + + if BUFFER_NAMES not in state_dict: + raise ValueError(f"{file} is not a model state checkpoint") + buffer_names = state_dict[BUFFER_NAMES] + if debug: + print("Found buffers:", buffer_names) + + # recover just the buffers while restoring them to fp32 if they were saved in fp16 + buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names} + param_shapes = state_dict[PARAM_SHAPES] + + # collect parameters that are included in param_shapes + param_names = [] + for s in param_shapes: + for name in s.keys(): + param_names.append(name) + + # update with frozen parameters + frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None) + if frozen_param_shapes is not None: + if debug: + print(f"Found frozen_param_shapes: {frozen_param_shapes}") + param_names += list(frozen_param_shapes.keys()) + + # handle shared params + shared_params = [[k, v] for k, v in state_dict["shared_params"].items()] + + ds_version = state_dict.get(DS_VERSION, None) + + frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None) + + z_model_state = zero_model_state(buffers=buffers, + param_shapes=param_shapes, + shared_params=shared_params, + ds_version=ds_version, + frozen_param_shapes=frozen_param_shapes, + frozen_param_fragments=frozen_param_fragments) + zero_model_states.append(z_model_state) + + return zero_model_states + + +def parse_optim_states(files, ds_checkpoint_dir): + + total_files = len(files) + state_dicts = [] + for f in files: + state_dict = torch.load(f, map_location=device) + # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights + # and also handle the case where it was already removed by another helper script + state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None) + state_dicts.append(state_dict) + + if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]: + raise ValueError(f"{files[0]} is not a zero checkpoint") + zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE] + world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT] + + # For ZeRO-2 each param group can have different partition_count as data parallelism for expert + # parameters can be different from data parallelism for non-expert parameters. So we can just + # use the max of the partition_count to get the dp world_size. + + if type(world_size) is list: + world_size = max(world_size) + + if world_size != total_files: + raise ValueError( + f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. " + "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes." + ) + + # the groups are named differently in each stage + if zero_stage <= 2: + fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS + elif zero_stage == 3: + fp32_groups_key = FP32_FLAT_GROUPS + else: + raise ValueError(f"unknown zero stage {zero_stage}") + + if zero_stage <= 2: + fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))] + elif zero_stage == 3: + # if there is more than one param group, there will be multiple flattened tensors - one + # flattened tensor per group - for simplicity merge them into a single tensor + # + # XXX: could make the script more memory efficient for when there are multiple groups - it + # will require matching the sub-lists of param_shapes for each param group flattened tensor + + fp32_flat_groups = [ + torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts)) + ] + + return zero_stage, world_size, fp32_flat_groups + + +def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters): + """ + Returns fp32 state_dict reconstructed from ds checkpoint + + Args: + - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are) + + """ + print(f"Processing zero checkpoint '{ds_checkpoint_dir}'") + + optim_files = get_optim_files(ds_checkpoint_dir) + zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir) + print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}") + + model_files = get_model_state_files(ds_checkpoint_dir) + + zero_model_states = parse_model_states(model_files) + print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}') + + if zero_stage <= 2: + return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states, + exclude_frozen_parameters) + elif zero_stage == 3: + return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states, + exclude_frozen_parameters) + + +def _zero2_merge_frozen_params(state_dict, zero_model_states): + if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0: + return + + frozen_param_shapes = zero_model_states[0].frozen_param_shapes + frozen_param_fragments = zero_model_states[0].frozen_param_fragments + + if debug: + num_elem = sum(s.numel() for s in frozen_param_shapes.values()) + print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}') + + wanted_params = len(frozen_param_shapes) + wanted_numel = sum(s.numel() for s in frozen_param_shapes.values()) + avail_numel = sum([p.numel() for p in frozen_param_fragments.values()]) + print(f'Frozen params: Have {avail_numel} numels to process.') + print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params') + + total_params = 0 + total_numel = 0 + for name, shape in frozen_param_shapes.items(): + total_params += 1 + unpartitioned_numel = shape.numel() + total_numel += unpartitioned_numel + + state_dict[name] = frozen_param_fragments[name] + + if debug: + print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ") + + print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements") + + +def _has_callable(obj, fn): + attr = getattr(obj, fn, None) + return callable(attr) + + +def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states): + param_shapes = zero_model_states[0].param_shapes + + # Reconstruction protocol: + # + # XXX: document this + + if debug: + for i in range(world_size): + for j in range(len(fp32_flat_groups[0])): + print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}") + + # XXX: memory usage doubles here (zero2) + num_param_groups = len(fp32_flat_groups[0]) + merged_single_partition_of_fp32_groups = [] + for i in range(num_param_groups): + merged_partitions = [sd[i] for sd in fp32_flat_groups] + full_single_fp32_vector = torch.cat(merged_partitions, 0) + merged_single_partition_of_fp32_groups.append(full_single_fp32_vector) + avail_numel = sum( + [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups]) + + if debug: + wanted_params = sum([len(shapes) for shapes in param_shapes]) + wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes]) + # not asserting if there is a mismatch due to possible padding + print(f"Have {avail_numel} numels to process.") + print(f"Need {wanted_numel} numels in {wanted_params} params.") + + # params + # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support + # out-of-core computing solution + total_numel = 0 + total_params = 0 + for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups): + offset = 0 + avail_numel = full_single_fp32_vector.numel() + for name, shape in shapes.items(): + + unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape) + total_numel += unpartitioned_numel + total_params += 1 + + if debug: + print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ") + state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape) + offset += unpartitioned_numel + + # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and + # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex + # paddings performed in the code it's almost impossible to predict the exact numbers w/o the + # live optimizer object, so we are checking that the numbers are within the right range + align_to = 2 * world_size + + def zero2_align(x): + return align_to * math.ceil(x / align_to) + + if debug: + print(f"original offset={offset}, avail_numel={avail_numel}") + + offset = zero2_align(offset) + avail_numel = zero2_align(avail_numel) + + if debug: + print(f"aligned offset={offset}, avail_numel={avail_numel}") + + # Sanity check + if offset != avail_numel: + raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong") + + print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements") + + +def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states, + exclude_frozen_parameters): + state_dict = OrderedDict() + + # buffers + buffers = zero_model_states[0].buffers + state_dict.update(buffers) + if debug: + print(f"added {len(buffers)} buffers") + + if not exclude_frozen_parameters: + _zero2_merge_frozen_params(state_dict, zero_model_states) + + _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states) + + # recover shared parameters + for pair in zero_model_states[0].shared_params: + if pair[1] in state_dict: + state_dict[pair[0]] = state_dict[pair[1]] + + return state_dict + + +def zero3_partitioned_param_info(unpartitioned_numel, world_size): + remainder = unpartitioned_numel % world_size + padding_numel = (world_size - remainder) if remainder else 0 + partitioned_numel = math.ceil(unpartitioned_numel / world_size) + return partitioned_numel, padding_numel + + +def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states): + if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0: + return + + if debug: + for i in range(world_size): + num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values()) + print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}') + + frozen_param_shapes = zero_model_states[0].frozen_param_shapes + wanted_params = len(frozen_param_shapes) + wanted_numel = sum(s.numel() for s in frozen_param_shapes.values()) + avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size + print(f'Frozen params: Have {avail_numel} numels to process.') + print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params') + + total_params = 0 + total_numel = 0 + for name, shape in zero_model_states[0].frozen_param_shapes.items(): + total_params += 1 + unpartitioned_numel = shape.numel() + total_numel += unpartitioned_numel + + param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states) + state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape) + + partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size) + + if debug: + print( + f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}" + ) + + print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements") + + +def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states): + param_shapes = zero_model_states[0].param_shapes + avail_numel = fp32_flat_groups[0].numel() * world_size + # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each + # param, re-consolidating each param, while dealing with padding if any + + # merge list of dicts, preserving order + param_shapes = {k: v for d in param_shapes for k, v in d.items()} + + if debug: + for i in range(world_size): + print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}") + + wanted_params = len(param_shapes) + wanted_numel = sum(shape.numel() for shape in param_shapes.values()) + # not asserting if there is a mismatch due to possible padding + avail_numel = fp32_flat_groups[0].numel() * world_size + print(f"Trainable params: Have {avail_numel} numels to process.") + print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.") + + # params + # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support + # out-of-core computing solution + offset = 0 + total_numel = 0 + total_params = 0 + for name, shape in param_shapes.items(): + + unpartitioned_numel = shape.numel() + total_numel += unpartitioned_numel + total_params += 1 + + partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size) + + if debug: + print( + f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}" + ) + + # XXX: memory usage doubles here + state_dict[name] = torch.cat( + tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)), + 0).narrow(0, 0, unpartitioned_numel).view(shape) + offset += partitioned_numel + + offset *= world_size + + # Sanity check + if offset != avail_numel: + raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong") + + print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements") + + +def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states, + exclude_frozen_parameters): + state_dict = OrderedDict() + + # buffers + buffers = zero_model_states[0].buffers + state_dict.update(buffers) + if debug: + print(f"added {len(buffers)} buffers") + + if not exclude_frozen_parameters: + _zero3_merge_frozen_params(state_dict, world_size, zero_model_states) + + _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states) + + # recover shared parameters + for pair in zero_model_states[0].shared_params: + if pair[1] in state_dict: + state_dict[pair[0]] = state_dict[pair[1]] + + return state_dict + + +def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False): + """ + Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with + ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example + via a model hub. + + Args: + - ``checkpoint_dir``: path to the desired checkpoint folder + - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14`` + - ``exclude_frozen_parameters``: exclude frozen parameters + + Returns: + - pytorch ``state_dict`` + + Note: this approach may not work if your application doesn't have sufficient free CPU memory and + you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with + the checkpoint. + + A typical usage might be :: + + from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint + # do the training and checkpoint saving + state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu + model = model.cpu() # move to cpu + model.load_state_dict(state_dict) + # submit to model hub or save the model to share with others + + In this example the ``model`` will no longer be usable in the deepspeed context of the same + application. i.e. you will need to re-initialize the deepspeed engine, since + ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it. + + If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead. + + """ + if tag is None: + latest_path = os.path.join(checkpoint_dir, 'latest') + if os.path.isfile(latest_path): + with open(latest_path, 'r') as fd: + tag = fd.read().strip() + else: + raise ValueError(f"Unable to find 'latest' file at {latest_path}") + + ds_checkpoint_dir = os.path.join(checkpoint_dir, tag) + + if not os.path.isdir(ds_checkpoint_dir): + raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist") + + return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters) + + +def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None, exclude_frozen_parameters=False): + """ + Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be + loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed. + + Args: + - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``) + - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin) + - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14`` + - ``exclude_frozen_parameters``: exclude frozen parameters + """ + + state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters) + print(f"Saving fp32 state dict to {output_file}") + torch.save(state_dict, output_file) + + +def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None): + """ + 1. Put the provided model to cpu + 2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` + 3. Load it into the provided model + + Args: + - ``model``: the model object to update + - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``) + - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14`` + + Returns: + - ``model`: modified model + + Make sure you have plenty of CPU memory available before you call this function. If you don't + have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it + conveniently placed for you in the checkpoint folder. + + A typical usage might be :: + + from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint + model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir) + # submit to model hub or save the model to share with others + + Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context + of the same application. i.e. you will need to re-initialize the deepspeed engine, since + ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it. + + """ + logger.info(f"Extracting fp32 weights") + state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag) + + logger.info(f"Overwriting model with fp32 weights") + model = model.cpu() + model.load_state_dict(state_dict, strict=False) + + return model + + +if __name__ == "__main__": + + parser = argparse.ArgumentParser() + parser.add_argument("checkpoint_dir", + type=str, + help="path to the desired checkpoint folder, e.g., path/checkpoint-12") + parser.add_argument( + "output_file", + type=str, + help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)") + parser.add_argument("-t", + "--tag", + type=str, + default=None, + help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1") + parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters") + parser.add_argument("-d", "--debug", action='store_true', help="enable debug") + args = parser.parse_args() + + debug = args.debug + + convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, + args.output_file, + tag=args.tag, + exclude_frozen_parameters=args.exclude_frozen_parameters) diff --git a/ckpts/qwen3-1.7b-whisper-260126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/log.txt b/ckpts/qwen3-1.7b-whisper-260126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/log.txt new file mode 100644 index 0000000000000000000000000000000000000000..dac4bc5539784f593827a5cb81096eef78629d65 --- /dev/null +++ b/ckpts/qwen3-1.7b-whisper-260126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/log.txt @@ -0,0 +1,27513 @@ +[2026-01-26 13:46:13,851] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect) +[2026-01-26 13:46:15,346] [WARNING] [runner.py:202:fetch_hostfile] Unable to find hostfile, will proceed with training with local resources only. +[2026-01-26 13:46:15,346] [INFO] [runner.py:568:main] cmd = /usr/bin/python -u -m deepspeed.launcher.launch --world_info=eyJsb2NhbGhvc3QiOiBbMCwgMSwgMiwgM119 --master_addr=127.0.0.1 --master_port=29500 --enable_each_rank_log=None pretrain_demo_qwenaudio.py --num_worker 16 --save_path /fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B --save_steps 1000 --logging_steps 1 --eval_steps 4000001 --train_batch_size 56 --micro_train_batch_size 14 --micro_eval_batch_size 8 --model_path /fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-1221_4x3000h_chunked_flylora_none_fixed/ckpts/global_step24000_hf_merged_lora --tokenizer_path /fs/nlp/common/plms/qwen-audio/qwen3-1.7b-chat-audio-whisper-v3-convproj --max_epochs 3 --max_len 1024 --zero_stage 2 --max_ckpt_num 5 --learning_rate 2e-5 --flash_attn --use_custom_qwen3 --dataset_config /fs/nlp/common_intern/meiyuxiang/uniscale_multimodal/scripts/myx/data_config/train_config_12x1000h_mls_lite1h.yaml --tensorboard-dir /fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/tb_logs --n_mels 128 --audio_subsampling_scale 4 --packing_samples --seed 421 --freeze_llm --freeze_audio --lora_rank 32 --lora_alpha 64 --specaug_policy SM --post_audio_ratio 1 --gradient_checkpointing --load_checkpoint --bf16 --randomize_chunk_window --use_zipper_lora --use_lid_router --zipper_lora_num_languages 12 --zipper_lora_r 32 --trainable_A --zipper_lora_scope audio --use_soft_routing --init_B +[2026-01-26 13:46:17,584] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect) +[2026-01-26 13:46:18,879] [INFO] [launch.py:138:main] 0 NCCL_VERSION=2.19.3 +[2026-01-26 13:46:18,879] [INFO] [launch.py:138:main] 0 NCCL_SOCKET_IFNAME=eth1 +[2026-01-26 13:46:18,879] [INFO] [launch.py:138:main] 0 NCCL_CUMEM_HOST_ENABLE=0 +[2026-01-26 13:46:18,879] [INFO] [launch.py:138:main] 0 NCCL_IB_HCA=^=mlx5_4 +[2026-01-26 13:46:18,879] [INFO] [launch.py:145:main] WORLD INFO DICT: {'localhost': [0, 1, 2, 3]} +[2026-01-26 13:46:18,879] [INFO] [launch.py:151:main] nnodes=1, num_local_procs=4, node_rank=0 +[2026-01-26 13:46:18,879] [INFO] [launch.py:162:main] global_rank_mapping=defaultdict(, {'localhost': [0, 1, 2, 3]}) +[2026-01-26 13:46:18,879] [INFO] [launch.py:163:main] dist_world_size=4 +[2026-01-26 13:46:18,879] [INFO] [launch.py:165:main] Setting CUDA_VISIBLE_DEVICES=0,1,2,3 +[2026-01-26 13:46:18,880] [INFO] [launch.py:253:main] process 413693 spawned with command: ['/usr/bin/python', '-u', 'pretrain_demo_qwenaudio.py', '--local_rank=0', '--num_worker', '16', '--save_path', '/fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B', '--save_steps', '1000', '--logging_steps', '1', '--eval_steps', '4000001', '--train_batch_size', '56', '--micro_train_batch_size', '14', '--micro_eval_batch_size', '8', '--model_path', '/fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-1221_4x3000h_chunked_flylora_none_fixed/ckpts/global_step24000_hf_merged_lora', '--tokenizer_path', '/fs/nlp/common/plms/qwen-audio/qwen3-1.7b-chat-audio-whisper-v3-convproj', '--max_epochs', '3', '--max_len', '1024', '--zero_stage', '2', '--max_ckpt_num', '5', '--learning_rate', '2e-5', '--flash_attn', '--use_custom_qwen3', '--dataset_config', '/fs/nlp/common_intern/meiyuxiang/uniscale_multimodal/scripts/myx/data_config/train_config_12x1000h_mls_lite1h.yaml', '--tensorboard-dir', '/fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/tb_logs', '--n_mels', '128', '--audio_subsampling_scale', '4', '--packing_samples', '--seed', '421', '--freeze_llm', '--freeze_audio', '--lora_rank', '32', '--lora_alpha', '64', '--specaug_policy', 'SM', '--post_audio_ratio', '1', '--gradient_checkpointing', '--load_checkpoint', '--bf16', '--randomize_chunk_window', '--use_zipper_lora', '--use_lid_router', '--zipper_lora_num_languages', '12', '--zipper_lora_r', '32', '--trainable_A', '--zipper_lora_scope', 'audio', '--use_soft_routing', '--init_B'] +[2026-01-26 13:46:18,882] [INFO] [launch.py:253:main] process 413694 spawned with command: ['/usr/bin/python', '-u', 'pretrain_demo_qwenaudio.py', '--local_rank=1', '--num_worker', '16', '--save_path', '/fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B', '--save_steps', '1000', '--logging_steps', '1', '--eval_steps', '4000001', '--train_batch_size', '56', '--micro_train_batch_size', '14', '--micro_eval_batch_size', '8', '--model_path', '/fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-1221_4x3000h_chunked_flylora_none_fixed/ckpts/global_step24000_hf_merged_lora', '--tokenizer_path', '/fs/nlp/common/plms/qwen-audio/qwen3-1.7b-chat-audio-whisper-v3-convproj', '--max_epochs', '3', '--max_len', '1024', '--zero_stage', '2', '--max_ckpt_num', '5', '--learning_rate', '2e-5', '--flash_attn', '--use_custom_qwen3', '--dataset_config', '/fs/nlp/common_intern/meiyuxiang/uniscale_multimodal/scripts/myx/data_config/train_config_12x1000h_mls_lite1h.yaml', '--tensorboard-dir', '/fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/tb_logs', '--n_mels', '128', '--audio_subsampling_scale', '4', '--packing_samples', '--seed', '421', '--freeze_llm', '--freeze_audio', '--lora_rank', '32', '--lora_alpha', '64', '--specaug_policy', 'SM', '--post_audio_ratio', '1', '--gradient_checkpointing', '--load_checkpoint', '--bf16', '--randomize_chunk_window', '--use_zipper_lora', '--use_lid_router', '--zipper_lora_num_languages', '12', '--zipper_lora_r', '32', '--trainable_A', '--zipper_lora_scope', 'audio', '--use_soft_routing', '--init_B'] +[2026-01-26 13:46:18,882] [INFO] [launch.py:253:main] process 413695 spawned with command: ['/usr/bin/python', '-u', 'pretrain_demo_qwenaudio.py', '--local_rank=2', '--num_worker', '16', '--save_path', '/fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B', '--save_steps', '1000', '--logging_steps', '1', '--eval_steps', '4000001', '--train_batch_size', '56', '--micro_train_batch_size', '14', '--micro_eval_batch_size', '8', '--model_path', '/fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-1221_4x3000h_chunked_flylora_none_fixed/ckpts/global_step24000_hf_merged_lora', '--tokenizer_path', '/fs/nlp/common/plms/qwen-audio/qwen3-1.7b-chat-audio-whisper-v3-convproj', '--max_epochs', '3', '--max_len', '1024', '--zero_stage', '2', '--max_ckpt_num', '5', '--learning_rate', '2e-5', '--flash_attn', '--use_custom_qwen3', '--dataset_config', '/fs/nlp/common_intern/meiyuxiang/uniscale_multimodal/scripts/myx/data_config/train_config_12x1000h_mls_lite1h.yaml', '--tensorboard-dir', '/fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/tb_logs', '--n_mels', '128', '--audio_subsampling_scale', '4', '--packing_samples', '--seed', '421', '--freeze_llm', '--freeze_audio', '--lora_rank', '32', '--lora_alpha', '64', '--specaug_policy', 'SM', '--post_audio_ratio', '1', '--gradient_checkpointing', '--load_checkpoint', '--bf16', '--randomize_chunk_window', '--use_zipper_lora', '--use_lid_router', '--zipper_lora_num_languages', '12', '--zipper_lora_r', '32', '--trainable_A', '--zipper_lora_scope', 'audio', '--use_soft_routing', '--init_B'] +[2026-01-26 13:46:18,883] [INFO] [launch.py:253:main] process 413696 spawned with command: ['/usr/bin/python', '-u', 'pretrain_demo_qwenaudio.py', '--local_rank=3', '--num_worker', '16', '--save_path', '/fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B', '--save_steps', '1000', '--logging_steps', '1', '--eval_steps', '4000001', '--train_batch_size', '56', '--micro_train_batch_size', '14', '--micro_eval_batch_size', '8', '--model_path', '/fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-1221_4x3000h_chunked_flylora_none_fixed/ckpts/global_step24000_hf_merged_lora', '--tokenizer_path', '/fs/nlp/common/plms/qwen-audio/qwen3-1.7b-chat-audio-whisper-v3-convproj', '--max_epochs', '3', '--max_len', '1024', '--zero_stage', '2', '--max_ckpt_num', '5', '--learning_rate', '2e-5', '--flash_attn', '--use_custom_qwen3', '--dataset_config', '/fs/nlp/common_intern/meiyuxiang/uniscale_multimodal/scripts/myx/data_config/train_config_12x1000h_mls_lite1h.yaml', '--tensorboard-dir', '/fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/tb_logs', '--n_mels', '128', '--audio_subsampling_scale', '4', '--packing_samples', '--seed', '421', '--freeze_llm', '--freeze_audio', '--lora_rank', '32', '--lora_alpha', '64', '--specaug_policy', 'SM', '--post_audio_ratio', '1', '--gradient_checkpointing', '--load_checkpoint', '--bf16', '--randomize_chunk_window', '--use_zipper_lora', '--use_lid_router', '--zipper_lora_num_languages', '12', '--zipper_lora_r', '32', '--trainable_A', '--zipper_lora_scope', 'audio', '--use_soft_routing', '--init_B'] +[2026-01-26 13:46:23,367] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect) +[2026-01-26 13:46:23,368] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect) +[2026-01-26 13:46:23,370] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect) +[2026-01-26 13:46:23,383] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect) +[2026-01-26 13:46:23,793] [INFO] [comm.py:637:init_distributed] cdb=None +[2026-01-26 13:46:23,793] [INFO] [comm.py:668:init_distributed] Initializing TorchBackend in DeepSpeed with backend nccl +[2026-01-26 13:46:24,022] [INFO] [comm.py:637:init_distributed] cdb=None +[2026-01-26 13:46:24,024] [INFO] [comm.py:637:init_distributed] cdb=None +[W CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator()) +[2026-01-26 13:46:24,024] [INFO] [comm.py:637:init_distributed] cdb=None +[W CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator()) +[W CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator()) +[W CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator()) +2026-01-26 13:46:24.546 | INFO | model.unigpt_audio_models.tokenization_qwen_audio:__init__:147 - vocab_size: 155165, audio_start_id: 155163, audio_end_id: 155164, audio_pad_id: 151851. +2026-01-26 13:46:24.552 | INFO | model.unigpt_audio_models.tokenization_qwen_audio:__init__:147 - vocab_size: 155165, audio_start_id: 155163, audio_end_id: 155164, audio_pad_id: 151851. +2026-01-26 13:46:24.552 | INFO | model.unigpt_audio_models.tokenization_qwen_audio:__init__:147 - vocab_size: 155165, audio_start_id: 155163, audio_end_id: 155164, audio_pad_id: 151851. +2026-01-26 13:46:24.558 | INFO | model.unigpt_audio_models.tokenization_qwen_audio:__init__:147 - vocab_size: 155165, audio_start_id: 155163, audio_end_id: 155164, audio_pad_id: 151851. +loading dataset com_voice_ar with +loading dataset com_voice_ar with +loading dataset com_voice_ar with +loading dataset com_voice_ar with + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 0it [00:00, ?it/s] 0it [00:00, ?it/s] 3619it [00:00, 36187.36it/s] 3516it [00:00, 35151.23it/s] 3661it [00:00, 36596.25it/s] 3634it [00:00, 36333.36it/s] 6363it [00:00, 37982.38it/s] +2026-01-26 13:46:24.759 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 41 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ar_pack_new.jsonl + 0it [00:00, ?it/s] 6363it [00:00, 37706.66it/s] 6363it [00:00, 37618.16it/s] + +2026-01-26 13:46:24.760 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 41 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ar_pack_new.jsonl +2026-01-26 13:46:24.760 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 41 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ar_pack_new.jsonl + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 6363it [00:00, 37182.17it/s] +2026-01-26 13:46:24.762 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 41 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ar_pack_new.jsonl + 0it [00:00, ?it/s] 4563it [00:00, 45623.84it/s] 4567it [00:00, 45660.57it/s] 4567it [00:00, 45666.34it/s] 4619it [00:00, 46181.02it/s] 6363it [00:00, 46501.23it/s] +2026-01-26 13:46:24.897 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 6363 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ar_pack_new.jsonl +loading dataset mls_de with + 0it [00:00, ?it/s] 6363it [00:00, 46327.76it/s] + 6363it [00:00, 46288.71it/s] +2026-01-26 13:46:24.899 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 6363 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ar_pack_new.jsonl + 6363it [00:00, 46908.42it/s] +2026-01-26 13:46:24.899 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 6363 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ar_pack_new.jsonl +loading dataset mls_de with +loading dataset mls_de with + 0it [00:00, ?it/s] 0it [00:00, ?it/s]2026-01-26 13:46:24.899 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 6363 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ar_pack_new.jsonl +loading dataset mls_de with + 0it [00:00, ?it/s] 11886it [00:00, 118853.31it/s] 11102it [00:00, 111006.61it/s] 11110it [00:00, 111081.57it/s] 12098it [00:00, 120974.92it/s] 24594it [00:00, 123684.17it/s] 23606it [00:00, 119256.69it/s] 23621it [00:00, 119321.71it/s] 24862it [00:00, 124889.60it/s] 37794it [00:00, 127476.23it/s] 36632it [00:00, 124277.19it/s] 36652it [00:00, 124333.53it/s] 37926it [00:00, 127508.24it/s] 50542it [00:00, 126950.47it/s] 49060it [00:00, 124182.04it/s] 49086it [00:00, 124182.32it/s] 50677it [00:00, 127454.64it/s] 63327it [00:00, 127271.62it/s] 61479it [00:00, 124029.66it/s] 63423it [00:00, 127385.67it/s] 61505it [00:00, 124041.17it/s] 63529it [00:00, 126321.15it/s] + 63529it [00:00, 126733.60it/s] + 63529it [00:00, 123240.98it/s] 63529it [00:00, 123256.09it/s] + +2026-01-26 13:46:25.424 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 32304 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_de_pack_new.jsonl +2026-01-26 13:46:25.424 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 32304 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_de_pack_new.jsonl +loading dataset mls_es with +loading dataset mls_es with + 0it [00:00, ?it/s] 0it [00:00, ?it/s]2026-01-26 13:46:25.438 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 32304 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_de_pack_new.jsonl +2026-01-26 13:46:25.438 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 32304 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_de_pack_new.jsonl +loading dataset mls_es with +loading dataset mls_es with + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 11520it [00:00, 115186.65it/s] 11474it [00:00, 114723.70it/s] 12201it [00:00, 122000.81it/s] 12291it [00:00, 122903.08it/s] 24570it [00:00, 124183.17it/s] 24388it [00:00, 123196.88it/s] 24846it [00:00, 124611.68it/s] 24933it [00:00, 124963.29it/s] 30551it [00:00, 121501.81it/s] + 30551it [00:00, 121269.99it/s] +2026-01-26 13:46:25.688 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 33300 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_es_pack_new.jsonl + 0it [00:00, ?it/s]2026-01-26 13:46:25.689 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 33300 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_es_pack_new.jsonl + 0it [00:00, ?it/s] 30551it [00:00, 122200.27it/s] 30551it [00:00, 122533.19it/s] + +2026-01-26 13:46:25.700 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 33300 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_es_pack_new.jsonl +2026-01-26 13:46:25.700 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 33300 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_es_pack_new.jsonl + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 11204it [00:00, 112024.88it/s] 11329it [00:00, 113283.36it/s] 11029it [00:00, 110279.59it/s] 11043it [00:00, 110406.68it/s] 22741it [00:00, 113978.65it/s] 22855it [00:00, 114442.27it/s] 22323it [00:00, 111834.50it/s] 22334it [00:00, 111871.00it/s] 30551it [00:00, 114767.06it/s] 30551it [00:00, 114325.74it/s] + +2026-01-26 13:46:25.957 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 30551 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_es_pack_new.jsonl +2026-01-26 13:46:25.957 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 30551 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_es_pack_new.jsonl +loading dataset mls_fr with +loading dataset mls_fr with + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 30551it [00:00, 112788.04it/s] 30551it [00:00, 112829.55it/s] + +2026-01-26 13:46:25.973 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 30551 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_es_pack_new.jsonl +2026-01-26 13:46:25.973 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 30551 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_es_pack_new.jsonl +loading dataset mls_fr with +loading dataset mls_fr with + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 8478it [00:00, 84774.83it/s] 8366it [00:00, 83644.13it/s] 8967it [00:00, 89660.47it/s] 8969it [00:00, 89684.95it/s] 17073it [00:00, 85437.14it/s] 16731it [00:00, 83539.50it/s] 17934it [00:00, 87380.10it/s] 17938it [00:00, 87390.84it/s] 25971it [00:00, 87051.47it/s] 25678it [00:00, 86240.63it/s] 26730it [00:00, 87638.58it/s] 26732it [00:00, 87634.71it/s] 34959it [00:00, 88149.60it/s] 34777it [00:00, 88096.19it/s] 36216it [00:00, 86845.98it/s] + 35508it [00:00, 87692.17it/s] 35531it [00:00, 87767.33it/s] 36216it [00:00, 86113.09it/s] + 36216it [00:00, 87720.38it/s] + 36216it [00:00, 87680.03it/s] +2026-01-26 13:46:26.397 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 33637 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_fr_pack_new.jsonl +loading dataset mls_it with + 0it [00:00, ?it/s]2026-01-26 13:46:26.401 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 33637 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_fr_pack_new.jsonl +loading dataset mls_it with + 0it [00:00, ?it/s]2026-01-26 13:46:26.409 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 33637 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_fr_pack_new.jsonl +2026-01-26 13:46:26.409 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 33637 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_fr_pack_new.jsonl +loading dataset mls_it with +loading dataset mls_it with + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 8282it [00:00, 127937.57it/s] + 8282it [00:00, 120765.91it/s] +2026-01-26 13:46:26.468 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 16564 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_it_pack_new.jsonl +loading dataset com_voice_ja with +2026-01-26 13:46:26.468 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 16564 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_it_pack_new.jsonl +loading dataset com_voice_ja with + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 8282it [00:00, 134666.51it/s] + 8282it [00:00, 133684.41it/s] +2026-01-26 13:46:26.472 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 16564 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_it_pack_new.jsonl +loading dataset com_voice_ja with +2026-01-26 13:46:26.472 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 16564 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/mls_it_pack_new.jsonl + 0it [00:00, ?it/s]loading dataset com_voice_ja with + 0it [00:00, ?it/s] 3970it [00:00, 39695.31it/s] 3917it [00:00, 39161.63it/s] 4074it [00:00, 40737.90it/s] 4048it [00:00, 40470.10it/s] 8315it [00:00, 41899.93it/s] 8239it [00:00, 41538.88it/s] 8394it [00:00, 42179.33it/s] 8313it [00:00, 41749.24it/s] 12565it [00:00, 42169.25it/s] 12447it [00:00, 41780.94it/s] 12639it [00:00, 42301.68it/s] 12637it [00:00, 42429.08it/s] 16796it [00:00, 42221.13it/s] 16691it [00:00, 42039.32it/s] 16870it [00:00, 42032.81it/s] 16880it [00:00, 41982.72it/s] 17319it [00:00, 41811.13it/s] +2026-01-26 13:46:26.883 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 36 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ja_pack_new.jsonl + 17319it [00:00, 41607.66it/s] +2026-01-26 13:46:26.885 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 36 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ja_pack_new.jsonl +loading dataset com_voice_ko with + 0it [00:00, ?it/s] 17319it [00:00, 41840.25it/s] +2026-01-26 13:46:26.887 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 36 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ja_pack_new.jsonl + 17319it [00:00, 41793.57it/s] +loading dataset com_voice_ko with +2026-01-26 13:46:26.888 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 36 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ja_pack_new.jsonl + 0it [00:00, ?it/s]loading dataset com_voice_ko with + 0it [00:00, ?it/s]loading dataset com_voice_ko with + 0it [00:00, ?it/s] 235it [00:00, 34661.23it/s] + 235it [00:00, 27836.47it/s] 235it [00:00, 46469.35it/s] + + 235it [00:00, 55077.19it/s] +2026-01-26 13:46:26.895 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 36 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ko_pack_new.jsonl +2026-01-26 13:46:26.895 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 36 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ko_pack_new.jsonl +loading dataset com_voice_pt with 2026-01-26 13:46:26.895 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 36 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ko_pack_new.jsonl + +2026-01-26 13:46:26.895 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 36 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_ko_pack_new.jsonl +loading dataset com_voice_pt with +loading dataset com_voice_pt with +loading dataset com_voice_pt with + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 0it [00:00, ?it/s] 0it [00:00, ?it/s] 65it [00:00, 20424.76it/s] + 65it [00:00, 20449.28it/s] 65it [00:00, 20218.76it/s] + 65it [00:00, 20257.82it/s] + +2026-01-26 13:46:26.899 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 22 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_pt_pack_new.jsonl +2026-01-26 13:46:26.899 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 22 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_pt_pack_new.jsonl +loading dataset msr86k_ru with 2026-01-26 13:46:26.899 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 22 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_pt_pack_new.jsonl + +2026-01-26 13:46:26.899 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 22 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_pt_pack_new.jsonl +loading dataset msr86k_ru with +loading dataset msr86k_ru with +loading dataset msr86k_ru with + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 0it [00:00, ?it/s] 0it [00:00, ?it/s] 3975it [00:00, 39742.74it/s] 4049it [00:00, 40488.11it/s] 3955it [00:00, 39535.99it/s] 3990it [00:00, 39889.01it/s] 8176it [00:00, 41070.82it/s] 8234it [00:00, 41445.33it/s] 8337it [00:00, 41887.46it/s] 8177it [00:00, 41048.26it/s] 12360it [00:00, 41420.08it/s] 12531it [00:00, 42138.05it/s] 12585it [00:00, 42155.36it/s] 12359it [00:00, 41399.04it/s] 16625it [00:00, 41903.04it/s] 16625it [00:00, 41894.65it/s] 16923it [00:00, 42838.28it/s] 16926it [00:00, 42648.15it/s] 20815it [00:00, 40471.68it/s] 20816it [00:00, 40470.57it/s] 21191it [00:00, 40897.13it/s] 21207it [00:00, 41039.73it/s] 24871it [00:00, 38980.50it/s] 24872it [00:00, 38986.43it/s] 25293it [00:00, 39479.25it/s] 25324it [00:00, 39576.63it/s] 28782it [00:00, 38122.52it/s] 28784it [00:00, 38124.21it/s] 29255it [00:00, 38652.69it/s] 29297it [00:00, 38683.05it/s] 32604it [00:00, 37372.98it/s] 32606it [00:00, 37372.34it/s] 33130it [00:00, 37824.30it/s] 33176it [00:00, 37925.90it/s] 36348it [00:00, 36884.84it/s] 36350it [00:00, 36884.63it/s] 36976it [00:00, 37401.79it/s] 36920it [00:00, 37194.52it/s] 40041it [00:01, 36758.34it/s] 40043it [00:01, 36761.64it/s] 40721it [00:01, 37241.62it/s] 40644it [00:01, 36851.29it/s] 43719it [00:01, 36688.30it/s] 43722it [00:01, 36687.62it/s] 44448it [00:01, 37210.43it/s] 44332it [00:01, 36686.82it/s] 47390it [00:01, 36601.40it/s] 47393it [00:01, 36597.29it/s] 48171it [00:01, 37097.69it/s] 48018it [00:01, 36735.24it/s] 51051it [00:01, 36519.04it/s] 51054it [00:01, 36519.63it/s] 51882it [00:01, 37003.01it/s] 51693it [00:01, 36692.89it/s] 54704it [00:01, 36500.16it/s] 54707it [00:01, 36496.46it/s] 55583it [00:01, 36909.58it/s] 55363it [00:01, 36427.29it/s] 58355it [00:01, 36304.58it/s] 58357it [00:01, 36306.40it/s] 59275it [00:01, 36711.02it/s] 59007it [00:01, 36286.63it/s] 61986it [00:01, 36271.23it/s] 61988it [00:01, 36273.37it/s] 62947it [00:01, 36620.07it/s] 62636it [00:01, 36263.88it/s] 65647it [00:01, 36371.78it/s] 65650it [00:01, 36372.68it/s] 66651it [00:01, 36743.30it/s] 66338it [00:01, 36487.54it/s] 69342it [00:01, 36543.36it/s] 69346it [00:01, 36545.13it/s] 70392it [00:01, 36939.32it/s] 70097it [00:01, 36814.47it/s] 73025it [00:01, 36627.56it/s] 73028it [00:01, 36621.47it/s] 74134it [00:01, 37081.72it/s] 73834it [00:01, 36978.82it/s] 76705it [00:02, 36678.72it/s] 76711it [00:02, 36680.68it/s] 77875it [00:02, 37177.33it/s] 77590it [00:02, 37150.49it/s] 80374it [00:02, 36644.35it/s] 80380it [00:02, 36630.43it/s] 81593it [00:02, 37021.97it/s] 81306it [00:02, 37010.32it/s] 84039it [00:02, 36522.79it/s] 84044it [00:02, 36538.20it/s] 85296it [00:02, 36944.80it/s] 85008it [00:02, 36880.50it/s] 85473it [00:02, 37853.04it/s] + 85473it [00:02, 37680.31it/s] + 85473it [00:02, 37381.14it/s] 85473it [00:02, 37380.90it/s] + +2026-01-26 13:46:29.186 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 35898 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/msr86k_ru_pack_new.jsonl +loading dataset com_voice_th with +2026-01-26 13:46:29.196 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 35898 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/msr86k_ru_pack_new.jsonl +loading dataset com_voice_th with +2026-01-26 13:46:29.214 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 35898 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/msr86k_ru_pack_new.jsonl +2026-01-26 13:46:29.215 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 35898 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/msr86k_ru_pack_new.jsonl +loading dataset com_voice_th with +loading dataset com_voice_th with + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 0it [00:00, ?it/s] 0it [00:00, ?it/s] 2876it [00:00, 28752.97it/s] 3068it [00:00, 30672.94it/s] 3014it [00:00, 30132.56it/s] 3023it [00:00, 30216.70it/s] 6267it [00:00, 31780.61it/s] 6570it [00:00, 33228.88it/s] 6400it [00:00, 32320.93it/s] 6419it [00:00, 32414.26it/s] 9805it [00:00, 33421.24it/s] 10112it [00:00, 34224.12it/s] 9872it [00:00, 33413.22it/s] 9890it [00:00, 33457.83it/s] 13559it [00:00, 35044.76it/s] 13902it [00:00, 35670.81it/s] 13564it [00:00, 34795.49it/s] 13585it [00:00, 34833.02it/s] 17398it [00:00, 35013.66it/s] 17398it [00:00, 35702.35it/s] + +2026-01-26 13:46:29.724 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 34796 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_th_pack_new.jsonl +2026-01-26 13:46:29.724 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 34796 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_th_pack_new.jsonl +loading dataset msr86k_vi with +loading dataset msr86k_vi with + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 17398it [00:00, 34921.10it/s] 17398it [00:00, 34957.02it/s] + +2026-01-26 13:46:29.753 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 34796 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_th_pack_new.jsonl +2026-01-26 13:46:29.753 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 34796 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/com_voice_th_pack_new.jsonl +loading dataset msr86k_vi with +loading dataset msr86k_vi with + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 3282it [00:00, 32816.51it/s] 3276it [00:00, 32756.44it/s] 3457it [00:00, 34559.16it/s] 3433it [00:00, 34319.72it/s] 6673it [00:00, 33465.54it/s] 6681it [00:00, 33496.97it/s] 6865it [00:00, 33677.40it/s] 6913it [00:00, 33631.59it/s] 10060it [00:00, 33649.31it/s] 10079it [00:00, 33712.88it/s] 10234it [00:00, 33642.08it/s] 10278it [00:00, 33505.28it/s] 13568it [00:00, 34211.84it/s] 13570it [00:00, 34182.23it/s] 13702it [00:00, 34048.21it/s] 13700it [00:00, 33783.40it/s] 17085it [00:00, 34553.33it/s] 17088it [00:00, 34538.15it/s] 17141it [00:00, 34166.48it/s] 17118it [00:00, 33921.41it/s] 20607it [00:00, 34776.15it/s] 20674it [00:00, 34986.03it/s] 20659it [00:00, 34505.60it/s] 20646it [00:00, 34377.51it/s] 24181it [00:00, 35090.17it/s] 24260it [00:00, 35268.14it/s] 24186it [00:00, 34750.75it/s] 24170it [00:00, 34657.48it/s] 27691it [00:00, 34248.08it/s] 27787it [00:00, 34336.57it/s] 27637it [00:00, 34064.12it/s] 27662it [00:00, 33852.26it/s] 31120it [00:00, 33179.70it/s] 31226it [00:00, 33285.75it/s] 31046it [00:00, 32889.66it/s] 31053it [00:00, 32763.73it/s] 34447it [00:01, 32591.98it/s] 34564it [00:01, 32700.92it/s] 34339it [00:01, 31748.87it/s] 34344it [00:01, 31840.35it/s] 37713it [00:01, 32215.01it/s] 37841it [00:01, 32233.74it/s] 37524it [00:01, 31551.97it/s] 37539it [00:01, 31568.89it/s] 40939it [00:01, 31795.44it/s] 41069it [00:01, 31839.55it/s] 40686it [00:01, 31293.65it/s] 40703it [00:01, 31345.33it/s] 44122it [00:01, 31657.65it/s] 44256it [00:01, 31696.76it/s] 43820it [00:01, 31275.87it/s] 43842it [00:01, 31305.44it/s] 47290it [00:01, 31569.80it/s] 47428it [00:01, 31450.53it/s] 46976it [00:01, 31116.88it/s] 46951it [00:01, 31011.82it/s] 50448it [00:01, 31339.04it/s] 50574it [00:01, 31330.71it/s] 50081it [00:01, 31093.39it/s] 50090it [00:01, 31096.35it/s] 53583it [00:01, 31162.02it/s] 53708it [00:01, 31199.33it/s] 53192it [00:01, 30951.73it/s] 53201it [00:01, 30948.28it/s] 56756it [00:01, 31327.72it/s] 56872it [00:01, 31327.94it/s] 56322it [00:01, 31052.38it/s] 56331it [00:01, 31049.47it/s] 59890it [00:01, 31175.68it/s] 60006it [00:01, 31185.68it/s] 59429it [00:01, 30961.33it/s] 59437it [00:01, 30932.89it/s] 63019it [00:01, 31208.89it/s] 63125it [00:01, 31159.72it/s] 62526it [00:01, 30940.17it/s] 62534it [00:01, 30942.73it/s] 66150it [00:02, 31232.99it/s] 66242it [00:02, 31096.75it/s] 65629it [00:02, 30920.66it/s] 65621it [00:02, 30807.67it/s] 69274it [00:02, 31063.34it/s] 69377it [00:02, 31169.74it/s] 68724it [00:02, 30926.37it/s] 68754it [00:02, 30960.16it/s] 72387it [00:02, 31082.00it/s] 72495it [00:02, 31150.97it/s] 71817it [00:02, 30854.49it/s] 71851it [00:02, 30869.00it/s] 75545it [00:02, 31227.85it/s] 75628it [00:02, 31200.83it/s] 74920it [00:02, 30904.94it/s] 74957it [00:02, 30924.85it/s] 78669it [00:02, 31178.67it/s] 78749it [00:02, 31178.83it/s] 78035it [00:02, 30974.63it/s] 78068it [00:02, 30978.23it/s] 81794it [00:02, 31199.08it/s] 81871it [00:02, 31189.78it/s] 81144it [00:02, 31008.73it/s] 81177it [00:02, 31008.45it/s] 84993it [00:02, 31197.55it/s] 84915it [00:02, 31142.47it/s] 84245it [00:02, 30953.77it/s] 84278it [00:02, 30984.15it/s] 88030it [00:02, 31139.27it/s] 88113it [00:02, 31163.47it/s] 87353it [00:02, 30988.37it/s] 87377it [00:02, 30837.75it/s] 91246it [00:02, 31211.37it/s] 91167it [00:02, 31205.52it/s] 90469it [00:02, 31038.85it/s] 90516it [00:02, 31002.05it/s] 94430it [00:02, 31396.46it/s] 94360it [00:02, 31419.13it/s] 93578it [00:02, 31053.52it/s] 93640it [00:02, 31071.67it/s] 97629it [00:03, 31572.13it/s] 97566it [00:03, 31607.47it/s] 96684it [00:03, 31033.96it/s] 96751it [00:03, 31080.16it/s] 100794it [00:03, 31592.07it/s] 100727it [00:03, 31521.34it/s] 99828it [00:03, 31155.11it/s] 99895it [00:03, 31184.97it/s] 103954it [00:03, 31589.85it/s] 103897it [00:03, 31574.15it/s] 102944it [00:03, 31100.29it/s] 103014it [00:03, 31143.88it/s] 107124it [00:03, 31620.58it/s] 107081it [00:03, 31652.89it/s] 106055it [00:03, 31072.30it/s] 106129it [00:03, 31084.76it/s] 110287it [00:03, 31561.38it/s] 110261it [00:03, 31696.50it/s] 109182it [00:03, 31131.04it/s] 109238it [00:03, 31047.16it/s] 113488it [00:03, 31694.15it/s] 113483it [00:03, 31852.04it/s] 112303it [00:03, 31154.26it/s] 112394it [00:03, 31200.06it/s] 116671it [00:03, 31733.48it/s] 116669it [00:03, 31805.43it/s] 115419it [00:03, 30912.59it/s] 115515it [00:03, 30984.78it/s] 119845it [00:03, 31657.24it/s] 119850it [00:03, 31708.41it/s] 118573it [00:03, 31098.84it/s] 118649it [00:03, 31086.69it/s] 123020it [00:03, 31684.56it/s] 123047it [00:03, 31782.54it/s] 121684it [00:03, 31037.47it/s] 121758it [00:03, 31041.40it/s] 126195it [00:03, 31702.63it/s] 126227it [00:03, 31786.48it/s] 124789it [00:03, 31018.03it/s] 124863it [00:03, 31034.96it/s] 129366it [00:04, 31686.15it/s] 129406it [00:04, 31760.69it/s] 127933it [00:04, 31143.65it/s] 127967it [00:04, 31010.83it/s] 132535it [00:04, 31529.78it/s] 132583it [00:04, 31577.08it/s] 131068it [00:04, 31203.02it/s] 131119it [00:04, 31160.87it/s] 135699it [00:04, 31559.36it/s] 135742it [00:04, 31580.03it/s] 134225it [00:04, 31312.54it/s] 134299it [00:04, 31349.67it/s] 138856it [00:04, 31545.86it/s] 138901it [00:04, 31553.81it/s] 137357it [00:04, 31216.03it/s] 137435it [00:04, 31255.71it/s] 141611it [00:04, 31924.16it/s] 141611it [00:04, 31924.10it/s] + +2026-01-26 13:46:34.194 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 36252 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/msr86k_vi_pack_new.jsonl +2026-01-26 13:46:34.194 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 36252 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/msr86k_vi_pack_new.jsonl + 140558it [00:04, 31452.50it/s]loading dataset librispeech_en with +loading dataset librispeech_en with + 140640it [00:04, 31489.50it/s] 141611it [00:04, 31561.07it/s] + 141611it [00:04, 31561.10it/s] +2026-01-26 13:46:34.274 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 36252 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/msr86k_vi_pack_new.jsonl +2026-01-26 13:46:34.274 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 36252 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/msr86k_vi_pack_new.jsonl + 0it [00:00, ?it/s] 0it [00:00, ?it/s]loading dataset librispeech_en with +loading dataset librispeech_en with + 0it [00:00, ?it/s] 0it [00:00, ?it/s] 6675it [00:00, 66741.79it/s] 6642it [00:00, 66413.25it/s] 7505it [00:00, 75037.91it/s] 7246it [00:00, 72449.53it/s] 21331it [00:00, 113682.75it/s] 21440it [00:00, 114387.90it/s] 30983it [00:00, 118837.81it/s] 30983it [00:00, 119148.45it/s] + +2026-01-26 13:46:34.540 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 30983 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/librispeech_en_pack_new.jsonl +2026-01-26 13:46:34.540 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 30983 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/librispeech_en_pack_new.jsonl + 21967it [00:00, 115960.21it/s] 21970it [00:00, 116435.49it/s] 30983it [00:00, 122554.49it/s] 30983it [00:00, 122544.55it/s] + +2026-01-26 13:46:34.614 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 30983 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/librispeech_en_pack_new.jsonl +2026-01-26 13:46:34.614 | INFO | data.qwen_audio_pretrain_dataset_pack:load_meta_dicts:76 - loaded dataset of 30983 samples from /fs/nlp/common_intern/meiyuxiang/data_prepare/step3_pack_2048_scale_4_fixed_1212/librispeech_en_pack_new.jsonl +AudioEncoderConfig(add_audio_bos_eos_token=True, audio_attn_implementation='flash_attention_2', audio_start_id=155163, avg_pool=True, is_causal=False, n_ctx=1500, n_head=20, n_layer=32, n_mels=128, n_state=1280, output_dim=2048, post_subsampling_scale=4, use_conformer_encoder=False, use_wav2vec2_encoder=False, wav2vec2_model_dim=None, wav2vec2_num_heads=None, wav2vec2_num_layers=None, wav2vec2_ffn_dim=None, use_conv_subsampling=True, use_chunk_encoder=True, chunk_n_window=400, conv_chunksize=500, chunk_n_window_candidates=[50, 100, 200, 400], randomize_chunk_window=True, chunk_level_router=False, num_moe_experts=8) +AudioEncoderConfig(add_audio_bos_eos_token=True, audio_attn_implementation='flash_attention_2', audio_start_id=155163, avg_pool=True, is_causal=False, n_ctx=1500, n_head=20, n_layer=32, n_mels=128, n_state=1280, output_dim=2048, post_subsampling_scale=4, use_conformer_encoder=False, use_wav2vec2_encoder=False, wav2vec2_model_dim=None, wav2vec2_num_heads=None, wav2vec2_num_layers=None, wav2vec2_ffn_dim=None, use_conv_subsampling=True, use_chunk_encoder=True, chunk_n_window=400, conv_chunksize=500, chunk_n_window_candidates=[50, 100, 200, 400], randomize_chunk_window=True, chunk_level_router=False, num_moe_experts=8) +AudioEncoderConfig(add_audio_bos_eos_token=True, audio_attn_implementation='flash_attention_2', audio_start_id=155163, avg_pool=True, is_causal=False, n_ctx=1500, n_head=20, n_layer=32, n_mels=128, n_state=1280, output_dim=2048, post_subsampling_scale=4, use_conformer_encoder=False, use_wav2vec2_encoder=False, wav2vec2_model_dim=None, wav2vec2_num_heads=None, wav2vec2_num_layers=None, wav2vec2_ffn_dim=None, use_conv_subsampling=True, use_chunk_encoder=True, chunk_n_window=400, conv_chunksize=500, chunk_n_window_candidates=[50, 100, 200, 400], randomize_chunk_window=True, chunk_level_router=False, num_moe_experts=8) +AudioEncoderConfig(add_audio_bos_eos_token=True, audio_attn_implementation='flash_attention_2', audio_start_id=155163, avg_pool=True, is_causal=False, n_ctx=1500, n_head=20, n_layer=32, n_mels=128, n_state=1280, output_dim=2048, post_subsampling_scale=4, use_conformer_encoder=False, use_wav2vec2_encoder=False, wav2vec2_model_dim=None, wav2vec2_num_heads=None, wav2vec2_num_layers=None, wav2vec2_ffn_dim=None, use_conv_subsampling=True, use_chunk_encoder=True, chunk_n_window=400, conv_chunksize=500, chunk_n_window_candidates=[50, 100, 200, 400], randomize_chunk_window=True, chunk_level_router=False, num_moe_experts=8) +2026-01-26 13:46:35.701 | INFO | model.unigpt_audio_models.audio:__init__:639 - initializing AudioEncoder with audio_attn_implementation=flash_attention_2, is_causal=False +2026-01-26 13:46:35.735 | INFO | model.unigpt_audio_models.audio:__init__:639 - initializing AudioEncoder with audio_attn_implementation=flash_attention_2, is_causal=False +2026-01-26 13:46:35.741 | INFO | model.unigpt_audio_models.audio:__init__:639 - initializing AudioEncoder with audio_attn_implementation=flash_attention_2, is_causal=False +2026-01-26 13:46:35.751 | INFO | model.unigpt_audio_models.audio:__init__:639 - initializing AudioEncoder with audio_attn_implementation=flash_attention_2, is_causal=False + Loading checkpoint shards: 0%| | 0/2 [00:00 +[2026-01-26 13:47:46,129] [INFO] [logging.py:96:log_dist] [Rank 0] Creating torch.bfloat16 ZeRO stage 2 optimizer +[2026-01-26 13:47:46,129] [INFO] [stage_1_and_2.py:149:__init__] Reduce bucket size 500,000,000 +[2026-01-26 13:47:46,129] [INFO] [stage_1_and_2.py:150:__init__] Allgather bucket size 500,000,000 +[2026-01-26 13:47:46,129] [INFO] [stage_1_and_2.py:151:__init__] CPU Offload: False +[2026-01-26 13:47:46,129] [INFO] [stage_1_and_2.py:152:__init__] Round robin gradient partitioning: False +[2026-01-26 13:48:02,678] [INFO] [utils.py:800:see_memory_usage] Before initializing optimizer states +[2026-01-26 13:48:02,679] [INFO] [utils.py:801:see_memory_usage] MA 5.23 GB Max_MA 5.23 GB CA 5.66 GB Max_CA 6 GB +[2026-01-26 13:48:02,686] [INFO] [utils.py:808:see_memory_usage] CPU Virtual Memory: used = 0.0 GB, percent = 0.0% +[2026-01-26 13:48:03,008] [INFO] [utils.py:800:see_memory_usage] After initializing optimizer states +[2026-01-26 13:48:03,009] [INFO] [utils.py:801:see_memory_usage] MA 5.23 GB Max_MA 5.51 GB CA 5.94 GB Max_CA 6 GB +[2026-01-26 13:48:03,009] [INFO] [utils.py:808:see_memory_usage] CPU Virtual Memory: used = 0.0 GB, percent = 0.0% +[2026-01-26 13:48:03,010] [INFO] [stage_1_and_2.py:539:__init__] optimizer state initialized +[2026-01-26 13:48:03,196] [INFO] [utils.py:800:see_memory_usage] After initializing ZeRO optimizer +[2026-01-26 13:48:03,197] [INFO] [utils.py:801:see_memory_usage] MA 5.23 GB Max_MA 5.23 GB CA 5.94 GB Max_CA 6 GB +[2026-01-26 13:48:03,197] [INFO] [utils.py:808:see_memory_usage] CPU Virtual Memory: used = 0.0 GB, percent = 0.0% +[2026-01-26 13:48:03,216] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed Final Optimizer = FusedAdam +[2026-01-26 13:48:03,216] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed using client LR scheduler +[2026-01-26 13:48:03,216] [INFO] [logging.py:96:log_dist] [Rank 0] DeepSpeed LR Scheduler = +[2026-01-26 13:48:03,216] [INFO] [logging.py:96:log_dist] [Rank 0] step=0, skipped=0, lr=[0.0, 0.0], mom=[(0.9, 0.95), (0.9, 0.95)] +[2026-01-26 13:48:03,224] [INFO] [config.py:987:print] DeepSpeedEngine configuration: +[2026-01-26 13:48:03,224] [INFO] [config.py:991:print] activation_checkpointing_config { + "partition_activations": false, + "contiguous_memory_optimization": false, + "cpu_checkpointing": false, + "number_checkpoints": null, + "synchronize_checkpoint_boundary": false, + "profile": false +} +[2026-01-26 13:48:03,224] [INFO] [config.py:991:print] aio_config ................... {'block_size': 1048576, 'queue_depth': 8, 'thread_count': 1, 'single_submit': False, 'overlap_events': True} +[2026-01-26 13:48:03,224] [INFO] [config.py:991:print] amp_enabled .................. False +[2026-01-26 13:48:03,224] [INFO] [config.py:991:print] amp_params ................... False +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] autotuning_config ............ { + "enabled": false, + "start_step": null, + "end_step": null, + "metric_path": null, + "arg_mappings": null, + "metric": "throughput", + "model_info": null, + "results_dir": "autotuning_results", + "exps_dir": "autotuning_exps", + "overwrite": true, + "fast": true, + "start_profile_step": 3, + "end_profile_step": 5, + "tuner_type": "gridsearch", + "tuner_early_stopping": 5, + "tuner_num_trials": 50, + "model_info_path": null, + "mp_size": 1, + "max_train_batch_size": null, + "min_train_batch_size": 1, + "max_train_micro_batch_size_per_gpu": 1.024000e+03, + "min_train_micro_batch_size_per_gpu": 1, + "num_tuning_micro_batch_sizes": 3 +} +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] bfloat16_enabled ............. True +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] checkpoint_parallel_write_pipeline False +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] checkpoint_tag_validation_enabled True +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] checkpoint_tag_validation_fail False +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] comms_config ................. +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] communication_data_type ...... None +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] compile_config ............... enabled=False backend='inductor' kwargs={} +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] compression_config ........... {'weight_quantization': {'shared_parameters': {'enabled': False, 'quantizer_kernel': False, 'schedule_offset': 0, 'quantize_groups': 1, 'quantize_verbose': False, 'quantization_type': 'symmetric', 'quantize_weight_in_forward': False, 'rounding': 'nearest', 'fp16_mixed_quantize': False, 'quantize_change_ratio': 0.001}, 'different_groups': {}}, 'activation_quantization': {'shared_parameters': {'enabled': False, 'quantization_type': 'symmetric', 'range_calibration': 'dynamic', 'schedule_offset': 1000}, 'different_groups': {}}, 'sparse_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'row_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'head_pruning': {'shared_parameters': {'enabled': False, 'method': 'topk', 'schedule_offset': 1000}, 'different_groups': {}}, 'channel_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'layer_reduction': {'enabled': False}} +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] curriculum_enabled_legacy .... False +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] curriculum_params_legacy ..... False +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] data_efficiency_config ....... {'enabled': False, 'seed': 1234, 'data_sampling': {'enabled': False, 'num_epochs': 1000, 'num_workers': 0, 'curriculum_learning': {'enabled': False}}, 'data_routing': {'enabled': False, 'random_ltd': {'enabled': False, 'layer_token_lr_schedule': {'enabled': False}}}} +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] data_efficiency_enabled ...... False +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] dataloader_drop_last ......... False +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] disable_allgather ............ False +[2026-01-26 13:48:03,225] [INFO] [config.py:991:print] dump_state ................... False +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] dynamic_loss_scale_args ...... None +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] eigenvalue_enabled ........... False +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] eigenvalue_gas_boundary_resolution 1 +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] eigenvalue_layer_name ........ bert.encoder.layer +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] eigenvalue_layer_num ......... 0 +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] eigenvalue_max_iter .......... 100 +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] eigenvalue_stability ......... 1e-06 +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] eigenvalue_tol ............... 0.01 +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] eigenvalue_verbose ........... False +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] elasticity_enabled ........... False +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] flops_profiler_config ........ { + "enabled": false, + "recompute_fwd_factor": 0.0, + "profile_step": 1, + "module_depth": -1, + "top_modules": 1, + "detailed": true, + "output_file": null +} +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] fp16_auto_cast ............... None +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] fp16_enabled ................. False +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] fp16_master_weights_and_gradients False +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] global_rank .................. 0 +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] grad_accum_dtype ............. fp32 +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] gradient_accumulation_steps .. 1 +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] gradient_clipping ............ 1.0 +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] gradient_predivide_factor .... 1.0 +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] graph_harvesting ............. False +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] hybrid_engine ................ enabled=False max_out_tokens=512 inference_tp_size=1 release_inference_cache=False pin_parameters=True tp_gather_partition_size=8 +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] initial_dynamic_scale ........ 1 +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] load_universal_checkpoint .... False +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] loss_scale ................... 1.0 +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] memory_breakdown ............. False +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] mics_hierarchial_params_gather False +[2026-01-26 13:48:03,226] [INFO] [config.py:991:print] mics_shard_size .............. -1 +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] monitor_config ............... tensorboard=TensorBoardConfig(enabled=False, output_path='', job_name='DeepSpeedJobName') wandb=WandbConfig(enabled=False, group=None, team=None, project='deepspeed') csv_monitor=CSVConfig(enabled=False, output_path='', job_name='DeepSpeedJobName') enabled=False +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] nebula_config ................ { + "enabled": false, + "persistent_storage_path": null, + "persistent_time_interval": 100, + "num_of_version_in_retention": 2, + "enable_nebula_load": true, + "load_path": null +} +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] optimizer_legacy_fusion ...... False +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] optimizer_name ............... None +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] optimizer_params ............. None +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0, 'pipe_partitioned': True, 'grad_partitioned': True} +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] pld_enabled .................. False +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] pld_params ................... False +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] prescale_gradients ........... False +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] scheduler_name ............... None +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] scheduler_params ............. None +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] seq_parallel_communication_data_type torch.float32 +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] sparse_attention ............. None +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] sparse_gradients_enabled ..... False +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] steps_per_print .............. 100 +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] train_batch_size ............. 56 +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] train_micro_batch_size_per_gpu 14 +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] use_data_before_expert_parallel_ False +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] use_node_local_storage ....... False +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] wall_clock_breakdown ......... False +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] weight_quantization_config ... None +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] world_size ................... 4 +[2026-01-26 13:48:03,227] [INFO] [config.py:991:print] zero_allow_untested_optimizer False +[2026-01-26 13:48:03,228] [INFO] [config.py:991:print] zero_config .................. stage=2 contiguous_gradients=True reduce_scatter=True reduce_bucket_size=500,000,000 use_multi_rank_bucket_allreduce=True allgather_partitions=True allgather_bucket_size=500,000,000 overlap_comm=False load_from_fp32_weights=True elastic_checkpoint=False offload_param=DeepSpeedZeroOffloadParamConfig(device='none', nvme_path=None, buffer_count=5, buffer_size=100,000,000, max_in_cpu=1,000,000,000, pin_memory=False) offload_optimizer=DeepSpeedZeroOffloadOptimizerConfig(device='none', nvme_path=None, buffer_count=4, pin_memory=True, pipeline=False, pipeline_read=False, pipeline_write=False, fast_init=False, ratio=1.0) sub_group_size=1,000,000,000 cpu_offload_param=None cpu_offload_use_pin_memory=None cpu_offload=None prefetch_bucket_size=50,000,000 param_persistence_threshold=100,000 model_persistence_threshold=sys.maxsize max_live_parameters=1,000,000,000 max_reuse_distance=1,000,000,000 gather_16bit_weights_on_model_save=False stage3_gather_fp16_weights_on_model_save=False ignore_unused_parameters=True legacy_stage1=False round_robin_gradients=False zero_hpz_partition_size=1 zero_quantized_weights=False zero_quantized_nontrainable_weights=False zero_quantized_gradients=False mics_shard_size=-1 mics_hierarchical_params_gather=False memory_efficient_linear=True pipeline_loading_checkpoint=False override_module_apply=True +[2026-01-26 13:48:03,228] [INFO] [config.py:991:print] zero_enabled ................. True +[2026-01-26 13:48:03,228] [INFO] [config.py:991:print] zero_force_ds_cpu_optimizer .. True +[2026-01-26 13:48:03,228] [INFO] [config.py:991:print] zero_optimization_stage ...... 2 +[2026-01-26 13:48:03,228] [INFO] [config.py:977:print_user_config] json = { + "steps_per_print": 100, + "zero_optimization": { + "stage": 2, + "offload_param": { + "device": "none" + }, + "offload_optimizer": { + "device": "none", + "pin_memory": true + }, + "sub_group_size": "auto", + "stage3_max_live_parameters": "auto", + "stage3_max_reuse_distance": "auto", + "stage3_param_persistence_threshold": "auto", + "stage3_prefetch_bucket_size": "auto", + "reduce_bucket_size": "auto", + "zero_hpz_partition_size": 1, + "zero_quantized_weights": false, + "zero_quantized_gradients": false + }, + "bf16": { + "enabled": true + }, + "gradient_clipping": 1.0, + "prescale_gradients": false, + "wall_clock_breakdown": false, + "data_types": { + "grad_accum_dtype": "fp32" + }, + "train_micro_batch_size_per_gpu": 14, + "train_batch_size": 56 +} +Load checkpoint: /fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B +[2026-01-26 13:48:03,228] [WARNING] [engine.py:2740:load_checkpoint] Unable to find latest file at /fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/ckpts/latest, if trying to load latest checkpoint please ensure this file exists or pass an explicit checkpoint tag when loading a checkpoint. +> setting tensorboard ... +[2026-01-26 13:48:03,249] [WARNING] [engine.py:2740:load_checkpoint] Unable to find latest file at /fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/ckpts/latest, if trying to load latest checkpoint please ensure this file exists or pass an explicit checkpoint tag when loading a checkpoint. +[2026-01-26 13:48:03,345] [WARNING] [engine.py:2740:load_checkpoint] Unable to find latest file at /fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/ckpts/latest, if trying to load latest checkpoint please ensure this file exists or pass an explicit checkpoint tag when loading a checkpoint. +[2026-01-26 13:48:03,348] [WARNING] [engine.py:2740:load_checkpoint] Unable to find latest file at /fs/nlp/common_intern/meiyuxiang/assets/multilingual/qwen3-1.7b-whisper-0126_12x1000h_lite1h_zipper_soft_lora_audio_init_baseline_with_lid_embedding_fix_init_B/ckpts/latest, if trying to load latest checkpoint please ensure this file exists or pass an explicit checkpoint tag when loading a checkpoint. +2026-01-26 13:48:03.416 | INFO | trainer.unigpt_pretrain_trainer:fit:130 - [Zipper LoRA] Freeze B weights for 0 steps. + Train epoch: 0%| | 0/3 [00:00