Upload folder using huggingface_hub

.gitattributes

@@ -35,3 +35,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 train/run.log filter=lfs diff=lfs merge=lfs -text
 train/run_inference.log filter=lfs diff=lfs merge=lfs -text
+train_no_diff/run.log filter=lfs diff=lfs merge=lfs -text

train_no_diff/.hydra/config.yaml

@@ -0,0 +1,114 @@
+data_wrapper:
+  dataset:
+    garment_tokenizer:
+      standardize:
+        rotations:
+          _target_: data.datasets.panel_configs.StatsConfig
+          scale:
+          - 1
+          - 1
+          - 1
+          - 1
+          shift:
+          - 0
+          - 0
+          - 0
+          - 0
+        translations:
+          _target_: data.datasets.panel_configs.StatsConfig
+          scale:
+          - 26.06867645
+          - 32.42920198
+          - 22.29905009
+          shift:
+          - -0.0125378371
+          - 113.507532
+          - 2.63046369
+        vertices:
+          _target_: data.datasets.panel_configs.StatsConfig
+          scale:
+          - 24.4920733
+          - 26.60402835
+          shift:
+          - 8.44428116
+          - 16.84081321
+        _target_: data.datasets.panel_configs.StandardizeConfig
+      _target_: data.garment_tokenizers.gcd_garment_tokenizer.GCDGarmentTokenizer
+      random_tag: true
+      num_tags: 108
+    _target_: data.datasets.gcd_mm_dataset.GCDMM
+    root_dir: datadir/GarmentCodeData_v2/GarmentCodeData_v2
+    editing_dir: datadir/GarmentCodeData_v2/GCD-MM/editing_dir/garmentcodedata_editing
+    caption_dir: datadir/GarmentCodeData_v2/GCD-MM/caption_dir/long-caption-processed
+    editing_flip_prob: 0.5
+    sampling_rate:
+    - 0.2
+    - 0.2
+    - 0.2
+    - 0.2
+    - 0.2
+    panel_classification: assets/data_configs/panel_classes_garmentcodedata.json
+    load_by_dataname: assets/data_configs/garmentcodedata_list.txt
+    image_size: 448
+    max_num_patches: 12
+    conv_template: internvl2_5
+  _target_: data.data_wrappers.data_wrapper.DataWrapper
+  _recursive_: false
+  collate_fn: data.data_wrappers.collate_fns.collate_fn
+  split_file: assets/data_configs/garmentcodedata_datasplit.json
+trainer:
+  _target_: trainers.trainer.Trainer
+  lr: 1.0e-05
+  beta1: 0.9
+  beta2: 0.95
+  grad_accumulation_steps: 10
+  batch_size: 4
+  num_steps: 12750
+  save_freq: 1000
+experiment:
+  wandb_info:
+    wandb_dir: wandb
+    wandb_cache_dir: wandb_cache
+  project_name: AIpparel
+  run_name: train
+  run_id: null
+  local_dir: null
+model:
+  edge_loss_weight: 0.1
+  pos_embed: true
+  num_freq: 0
+  pos_embed_type: sinusoidal
+  verts_bounds:
+  - -4
+  - -4
+  - 4
+  - 4
+  transf_bounds:
+  - -4
+  - -4
+  - -4
+  - -1
+  - -1
+  - -1
+  - -1
+  - 4
+  - 4
+  - 4
+  - 1
+  - 1
+  - 1
+  - 1
+  denormalize_for_loss: false
+  num_regression_layers: 2
+  discretize: true
+  bin_num: 256
+my_pretrained_path: null
+llm_pretrained_path: cache/InternVL3-2B-Instruct
+resume_path: null
+sampling_type: default
+pretrained_config: null
+precision: bf16
+evaluate: false
+conv_type: internvl2_5
+from_start: false
+storage_dir: output/

train_no_diff/.hydra/hydra.yaml

@@ -0,0 +1,168 @@
+hydra:
+  run:
+    dir: output/train/
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - experiment.project_name=AIpparel
+    - experiment.run_name=train
+    - evaluate=False
+  job:
+    name: diff_run
+    chdir: null
+    override_dirname: evaluate=False,experiment.project_name=AIpparel,experiment.run_name=train
+    id: ???
+    num: ???
+    config_name: aipparel_diff
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /root/workspace/SwiftTailor3
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /root/workspace/SwiftTailor3/configs
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /root/workspace/SwiftTailor3/output/train
+    choices:
+      model: aipparel_model
+      experiment: experiment
+      experiment/wandb_info: wandb
+      trainer: trainer
+      data_wrapper: gcd_datawrapper
+      data_wrapper/dataset: gcd_mm
+      data_wrapper/dataset/garment_tokenizer: gcd_garment_tokenizer
+      data_wrapper/dataset/garment_tokenizer/standardize: gcd_stats
+      data_wrapper/dataset/garment_tokenizer/standardize/vertices: gcd_verts_stats
+      data_wrapper/dataset/garment_tokenizer/standardize/translations: gcd_transl_stats
+      data_wrapper/dataset/garment_tokenizer/standardize/rotations: gcd_rotation_stats
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

train_no_diff/.hydra/overrides.yaml

@@ -0,0 +1,3 @@
+- experiment.project_name=AIpparel
+- experiment.run_name=train
+- evaluate=False

train_no_diff/ckpt_29000/global_step29001/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14456252e1c47d363caab10d2cbf9d5337b9e81f2defd46ddcec14f68525eccb
+size 5358433520

train_no_diff/ckpt_29000/global_step29001/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f5b84e448135c7ca98225fef07fe659b8804e5ae93fab09fd4eaa51b8a9da931
+size 5358437872

train_no_diff/ckpt_29000/global_step29001/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f73962f98a9f47154957208420ed79362e9bb3d41e1cb70ce1d26bbfb02c0af1
+size 5358438832

train_no_diff/ckpt_29000/global_step29001/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f536d84489944f78c2c4d247ad90b74f7d3f3b3624062b07496b55c1eb0d5f71
+size 5358432176

train_no_diff/ckpt_29000/global_step29001/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08191d11e0d4427b45565199e5d89b5955bbaf23c5bf3a091f6ac4e1669ac805
+size 4823376020

train_no_diff/ckpt_29000/latest

@@ -0,0 +1 @@
+global_step29001

train_no_diff/ckpt_29000/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        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 tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # 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}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] 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")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+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 = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * 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
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        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 to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=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
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    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.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    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")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

train_no_diff/ckpt_29999/global_step30000/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a7b43d3497b5c861dd263826e96412a83e78acb163dec0fcaafc76252631a7b
+size 5358433520

train_no_diff/ckpt_29999/global_step30000/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53aed9ab192f6ee5e2b7fd684f8ebef51dbf5b0a605e20a2b176f9e303182cb0
+size 5358437872

train_no_diff/ckpt_29999/global_step30000/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0714536548f0e7e2afbd1f78eb701e9751824c485059c83823b88a17a5c940cb
+size 5358438832

train_no_diff/ckpt_29999/global_step30000/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4108f6d4f0e14368b8fe52f0c75f1deeaa9f78e4ec7b869e6c0ba423aa8f43e1
+size 5358432176

train_no_diff/ckpt_29999/global_step30000/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bdc8d23e7b7be29859f048c1cb78fc40fbb30216c340b63d1c0efb497d75ee06
+size 4823376020

train_no_diff/ckpt_29999/latest

@@ -0,0 +1 @@
+global_step30000

train_no_diff/ckpt_29999/output_dir/config.json

@@ -0,0 +1,220 @@
+{
+    "_commit_hash": null,
+    "_name_or_path": "/mnt/petrelfs/share_data/wangweiyun/share_internvl_preview/InternVL3-2B-Pretrain",
+    "architectures": [
+        "InternVLChatModel"
+    ],
+    "auto_map": {
+        "AutoConfig": "configuration_internvl_chat.InternVLChatConfig",
+        "AutoModel": "modeling_internvl_chat.InternVLChatModel",
+        "AutoModelForCausalLM": "modeling_internvl_chat.InternVLChatModel"
+    },
+    "downsample_ratio": 0.5,
+    "dynamic_image_size": true,
+    "force_image_size": 448,
+    "image_fold": null,
+    "llm_config": {
+        "_name_or_path": "./pretrained/Qwen2.5-32B-Instruct",
+        "add_cross_attention": false,
+        "architectures": [
+            "Qwen2ForCausalLM"
+        ],
+        "attention_dropout": 0.0,
+        "bad_words_ids": null,
+        "begin_suppress_tokens": null,
+        "bos_token_id": 151643,
+        "chunk_size_feed_forward": 0,
+        "cross_attention_hidden_size": null,
+        "decoder_start_token_id": null,
+        "diversity_penalty": 0.0,
+        "do_sample": false,
+        "early_stopping": false,
+        "encoder_no_repeat_ngram_size": 0,
+        "eos_token_id": 151643,
+        "exponential_decay_length_penalty": null,
+        "finetuning_task": null,
+        "forced_bos_token_id": null,
+        "forced_eos_token_id": null,
+        "hidden_act": "silu",
+        "hidden_size": 1536,
+        "id2label": {
+            "0": "LABEL_0",
+            "1": "LABEL_1"
+        },
+        "initializer_range": 0.02,
+        "intermediate_size": 8960,
+        "is_decoder": false,
+        "is_encoder_decoder": false,
+        "label2id": {
+            "LABEL_0": 0,
+            "LABEL_1": 1
+        },
+        "length_penalty": 1.0,
+        "max_length": 20,
+        "max_position_embeddings": 32768,
+        "max_window_layers": 70,
+        "min_length": 0,
+        "model_type": "qwen2",
+        "moe_config": null,
+        "no_repeat_ngram_size": 0,
+        "num_attention_heads": 12,
+        "num_beam_groups": 1,
+        "num_beams": 1,
+        "num_hidden_layers": 28,
+        "num_key_value_heads": 2,
+        "num_return_sequences": 1,
+        "output_attentions": false,
+        "output_hidden_states": false,
+        "output_scores": false,
+        "pad_token_id": null,
+        "prefix": null,
+        "problem_type": null,
+        "pruned_heads": {},
+        "remove_invalid_values": false,
+        "repetition_penalty": 1.0,
+        "return_dict": true,
+        "return_dict_in_generate": false,
+        "rms_norm_eps": 1e-06,
+        "rope_scaling": {
+            "factor": 2.0,
+            "rope_type": "dynamic",
+            "type": "dynamic"
+        },
+        "rope_theta": 1000000.0,
+        "sep_token_id": null,
+        "sliding_window": null,
+        "suppress_tokens": null,
+        "task_specific_params": null,
+        "temperature": 1.0,
+        "tf_legacy_loss": false,
+        "tie_encoder_decoder": false,
+        "tie_word_embeddings": false,
+        "tokenizer_class": null,
+        "top_k": 50,
+        "top_p": 1.0,
+        "torch_dtype": "bfloat16",
+        "torchscript": false,
+        "transformers_version": "4.45.1",
+        "typical_p": 1.0,
+        "use_bfloat16": true,
+        "use_cache": true,
+        "use_sliding_window": false,
+        "vocab_size": 151796
+    },
+    "max_dynamic_patch": 12,
+    "min_dynamic_patch": 1,
+    "model_type": "internvl_chat",
+    "pad2square": false,
+    "ps_version": "v2",
+    "select_layer": -1,
+    "template": "internvl2_5",
+    "tie_word_embeddings": false,
+    "torch_dtype": "bfloat16",
+    "transformers_version": null,
+    "use_backbone_lora": 0,
+    "use_llm_lora": 0,
+    "use_thumbnail": true,
+    "vision_config": {
+        "_name_or_path": "OpenGVLab/InternViT-6B-448px-V1-5",
+        "add_cross_attention": false,
+        "architectures": [
+            "InternVisionModel"
+        ],
+        "attention_dropout": 0.0,
+        "auto_map": {
+            "AutoConfig": "configuration_intern_vit.InternVisionConfig",
+            "AutoModel": "modeling_intern_vit.InternVisionModel"
+        },
+        "bad_words_ids": null,
+        "begin_suppress_tokens": null,
+        "bos_token_id": null,
+        "capacity_factor": 1.2,
+        "chunk_size_feed_forward": 0,
+        "cross_attention_hidden_size": null,
+        "decoder_start_token_id": null,
+        "diversity_penalty": 0.0,
+        "do_sample": false,
+        "drop_path_rate": 0.0,
+        "dropout": 0.0,
+        "early_stopping": false,
+        "encoder_no_repeat_ngram_size": 0,
+        "eos_token_id": null,
+        "eval_capacity_factor": 1.4,
+        "exponential_decay_length_penalty": null,
+        "finetuning_task": null,
+        "forced_bos_token_id": null,
+        "forced_eos_token_id": null,
+        "hidden_act": "gelu",
+        "hidden_size": 1024,
+        "id2label": {
+            "0": "LABEL_0",
+            "1": "LABEL_1"
+        },
+        "image_size": 448,
+        "initializer_factor": 0.1,
+        "initializer_range": 1e-10,
+        "intermediate_size": 4096,
+        "is_decoder": false,
+        "is_encoder_decoder": false,
+        "label2id": {
+            "LABEL_0": 0,
+            "LABEL_1": 1
+        },
+        "laux_allreduce": "all_nodes",
+        "layer_norm_eps": 1e-06,
+        "length_penalty": 1.0,
+        "max_length": 20,
+        "min_length": 0,
+        "model_type": "intern_vit_6b",
+        "moe_coeff_ratio": 0.5,
+        "moe_intermediate_size": 768,
+        "moe_output_scale": 4.0,
+        "no_repeat_ngram_size": 0,
+        "noisy_gate_policy": "RSample_before",
+        "norm_type": "layer_norm",
+        "num_attention_heads": 16,
+        "num_beam_groups": 1,
+        "num_beams": 1,
+        "num_channels": 3,
+        "num_experts": 8,
+        "num_hidden_layers": 24,
+        "num_return_sequences": 1,
+        "num_routed_experts": 4,
+        "num_shared_experts": 4,
+        "output_attentions": false,
+        "output_hidden_states": false,
+        "output_scores": false,
+        "pad_token_id": null,
+        "patch_size": 14,
+        "prefix": null,
+        "problem_type": null,
+        "pruned_heads": {},
+        "qk_normalization": false,
+        "qkv_bias": true,
+        "remove_invalid_values": false,
+        "repetition_penalty": 1.0,
+        "return_dict": true,
+        "return_dict_in_generate": false,
+        "sep_token_id": null,
+        "shared_expert_intermediate_size": 3072,
+        "suppress_tokens": null,
+        "task_specific_params": null,
+        "temperature": 1.0,
+        "tf_legacy_loss": false,
+        "tie_encoder_decoder": false,
+        "tie_word_embeddings": true,
+        "tokenizer_class": null,
+        "top_k": 50,
+        "top_p": 1.0,
+        "torch_dtype": "bfloat16",
+        "torchscript": false,
+        "transformers_version": "4.45.1",
+        "typical_p": 1.0,
+        "use_bfloat16": true,
+        "use_flash_attn": true,
+        "use_moe": false,
+        "use_residual": true,
+        "use_rts": false,
+        "use_weighted_residual": false
+    }
+}

train_no_diff/ckpt_29999/output_dir/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b0998b45d71a1b5de7cf1ba5c38be2a272260c1ab0d094733fda21ce7f27502
+size 7770003116

train_no_diff/ckpt_29999/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        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 tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # 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}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] 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")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+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 = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * 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
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        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 to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=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
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    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.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    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")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

train_no_diff/diff_run.log

@@ -0,0 +1,46 @@
+[2025-09-03 21:58:38,725][__main__][INFO] - Working directory : /root/workspace/SwiftTailor3
+[2025-09-03 21:58:38,726][__main__][INFO] - Output directory : /root/workspace/SwiftTailor3/output/train
+[2025-09-03 21:58:38,734][data.data_wrappers.data_wrapper][INFO] - Loading data split from assets/data_configs/garmentcodedata_datasplit.json
+[2025-09-03 21:59:02,452][data.data_wrappers.data_wrapper][INFO] - Dataset split: 114869 / 6381 / 6381
+[2025-09-03 21:59:02,453][__main__][INFO] - Dataset created.
+[2025-09-03 21:59:02,752][__main__][INFO] - Added 122 tokens to the tokenizer.
+[2025-09-03 21:59:02,766][__main__][INFO] - Token name to index dictionary: {'<pattern_cmd_MOVE>': 151674, '<pattern_cmd_LINE>': 151675, '<pattern_cmd_CLINE>': 151676, '<pattern_cmd_CURVE>': 151677, '<pattern_cmd_CCURVE>': 151678, '<pattern_cmd_CUBIC>': 151679, '<pattern_cmd_CCUBIC>': 151680, '<pattern_cmd_ARC>': 151681, '<pattern_cmd_CARC>': 151682, '<panel_start>': 151683, '<panel_end>': 151684, '<pattern_start>': 151685, '<pattern_end>': 151686, '<stitch_tag_0>': 151687, '<stitch_tag_1>': 151688, '<stitch_tag_2>': 151689, '<stitch_tag_3>': 151690, '<stitch_tag_4>': 151691, '<stitch_tag_5>': 151692, '<stitch_tag_6>': 151693, '<stitch_tag_7>': 151694, '<stitch_tag_8>': 151695, '<stitch_tag_9>': 151696, '<stitch_tag_10>': 151697, '<stitch_tag_11>': 151698, '<stitch_tag_12>': 151699, '<stitch_tag_13>': 151700, '<stitch_tag_14>': 151701, '<stitch_tag_15>': 151702, '<stitch_tag_16>': 151703, '<stitch_tag_17>': 151704, '<stitch_tag_18>': 151705, '<stitch_tag_19>': 151706, '<stitch_tag_20>': 151707, '<stitch_tag_21>': 151708, '<stitch_tag_22>': 151709, '<stitch_tag_23>': 151710, '<stitch_tag_24>': 151711, '<stitch_tag_25>': 151712, '<stitch_tag_26>': 151713, '<stitch_tag_27>': 151714, '<stitch_tag_28>': 151715, '<stitch_tag_29>': 151716, '<stitch_tag_30>': 151717, '<stitch_tag_31>': 151718, '<stitch_tag_32>': 151719, '<stitch_tag_33>': 151720, '<stitch_tag_34>': 151721, '<stitch_tag_35>': 151722, '<stitch_tag_36>': 151723, '<stitch_tag_37>': 151724, '<stitch_tag_38>': 151725, '<stitch_tag_39>': 151726, '<stitch_tag_40>': 151727, '<stitch_tag_41>': 151728, '<stitch_tag_42>': 151729, '<stitch_tag_43>': 151730, '<stitch_tag_44>': 151731, '<stitch_tag_45>': 151732, '<stitch_tag_46>': 151733, '<stitch_tag_47>': 151734, '<stitch_tag_48>': 151735, '<stitch_tag_49>': 151736, '<stitch_tag_50>': 151737, '<stitch_tag_51>': 151738, '<stitch_tag_52>': 151739, '<stitch_tag_53>': 151740, '<stitch_tag_54>': 151741, '<stitch_tag_55>': 151742, '<stitch_tag_56>': 151743, '<stitch_tag_57>': 151744, '<stitch_tag_58>': 151745, '<stitch_tag_59>': 151746, '<stitch_tag_60>': 151747, '<stitch_tag_61>': 151748, '<stitch_tag_62>': 151749, '<stitch_tag_63>': 151750, '<stitch_tag_64>': 151751, '<stitch_tag_65>': 151752, '<stitch_tag_66>': 151753, '<stitch_tag_67>': 151754, '<stitch_tag_68>': 151755, '<stitch_tag_69>': 151756, '<stitch_tag_70>': 151757, '<stitch_tag_71>': 151758, '<stitch_tag_72>': 151759, '<stitch_tag_73>': 151760, '<stitch_tag_74>': 151761, '<stitch_tag_75>': 151762, '<stitch_tag_76>': 151763, '<stitch_tag_77>': 151764, '<stitch_tag_78>': 151765, '<stitch_tag_79>': 151766, '<stitch_tag_80>': 151767, '<stitch_tag_81>': 151768, '<stitch_tag_82>': 151769, '<stitch_tag_83>': 151770, '<stitch_tag_84>': 151771, '<stitch_tag_85>': 151772, '<stitch_tag_86>': 151773, '<stitch_tag_87>': 151774, '<stitch_tag_88>': 151775, '<stitch_tag_89>': 151776, '<stitch_tag_90>': 151777, '<stitch_tag_91>': 151778, '<stitch_tag_92>': 151779, '<stitch_tag_93>': 151780, '<stitch_tag_94>': 151781, '<stitch_tag_95>': 151782, '<stitch_tag_96>': 151783, '<stitch_tag_97>': 151784, '<stitch_tag_98>': 151785, '<stitch_tag_99>': 151786, '<stitch_tag_100>': 151787, '<stitch_tag_101>': 151788, '<stitch_tag_102>': 151789, '<stitch_tag_103>': 151790, '<stitch_tag_104>': 151791, '<stitch_tag_105>': 151792, '<stitch_tag_106>': 151793, '<stitch_tag_107>': 151794, '<stitch_tag_null>': 151795}
+[2025-09-03 21:59:02,768][models.internvl.configuration_internvl_chat][INFO] - vision_select_layer: -1
+[2025-09-03 21:59:02,768][models.internvl.configuration_internvl_chat][INFO] - ps_version: v2
+[2025-09-03 21:59:02,768][models.internvl.configuration_internvl_chat][INFO] - min_dynamic_patch: 1
+[2025-09-03 21:59:02,768][models.internvl.configuration_internvl_chat][INFO] - max_dynamic_patch: 12
+[2025-09-03 21:59:02,809][models.st_model_diff][INFO] - num_image_token: 256
+[2025-09-03 21:59:02,810][models.st_model_diff][INFO] - ps_version: v2
+[2025-09-03 21:59:24,668][trainers.trainer][INFO] - NN training Using device: cuda:0
+[2025-09-03 21:59:24,751][trainers.trainer][INFO] - Start training from step 0
+[2025-09-03 22:00:53,559][__main__][INFO] - Working directory : /root/workspace/SwiftTailor3
+[2025-09-03 22:00:53,559][__main__][INFO] - Output directory : /root/workspace/SwiftTailor3/output/train
+[2025-09-03 22:00:53,567][data.data_wrappers.data_wrapper][INFO] - Loading data split from assets/data_configs/garmentcodedata_datasplit.json
+[2025-09-03 22:01:17,294][data.data_wrappers.data_wrapper][INFO] - Dataset split: 114869 / 6381 / 6381
+[2025-09-03 22:01:17,296][__main__][INFO] - Dataset created.
+[2025-09-03 22:01:17,569][__main__][INFO] - Added 122 tokens to the tokenizer.
+[2025-09-03 22:01:17,579][__main__][INFO] - Token name to index dictionary: {'<pattern_cmd_MOVE>': 151674, '<pattern_cmd_LINE>': 151675, '<pattern_cmd_CLINE>': 151676, '<pattern_cmd_CURVE>': 151677, '<pattern_cmd_CCURVE>': 151678, '<pattern_cmd_CUBIC>': 151679, '<pattern_cmd_CCUBIC>': 151680, '<pattern_cmd_ARC>': 151681, '<pattern_cmd_CARC>': 151682, '<panel_start>': 151683, '<panel_end>': 151684, '<pattern_start>': 151685, '<pattern_end>': 151686, '<stitch_tag_0>': 151687, '<stitch_tag_1>': 151688, '<stitch_tag_2>': 151689, '<stitch_tag_3>': 151690, '<stitch_tag_4>': 151691, '<stitch_tag_5>': 151692, '<stitch_tag_6>': 151693, '<stitch_tag_7>': 151694, '<stitch_tag_8>': 151695, '<stitch_tag_9>': 151696, '<stitch_tag_10>': 151697, '<stitch_tag_11>': 151698, '<stitch_tag_12>': 151699, '<stitch_tag_13>': 151700, '<stitch_tag_14>': 151701, '<stitch_tag_15>': 151702, '<stitch_tag_16>': 151703, '<stitch_tag_17>': 151704, '<stitch_tag_18>': 151705, '<stitch_tag_19>': 151706, '<stitch_tag_20>': 151707, '<stitch_tag_21>': 151708, '<stitch_tag_22>': 151709, '<stitch_tag_23>': 151710, '<stitch_tag_24>': 151711, '<stitch_tag_25>': 151712, '<stitch_tag_26>': 151713, '<stitch_tag_27>': 151714, '<stitch_tag_28>': 151715, '<stitch_tag_29>': 151716, '<stitch_tag_30>': 151717, '<stitch_tag_31>': 151718, '<stitch_tag_32>': 151719, '<stitch_tag_33>': 151720, '<stitch_tag_34>': 151721, '<stitch_tag_35>': 151722, '<stitch_tag_36>': 151723, '<stitch_tag_37>': 151724, '<stitch_tag_38>': 151725, '<stitch_tag_39>': 151726, '<stitch_tag_40>': 151727, '<stitch_tag_41>': 151728, '<stitch_tag_42>': 151729, '<stitch_tag_43>': 151730, '<stitch_tag_44>': 151731, '<stitch_tag_45>': 151732, '<stitch_tag_46>': 151733, '<stitch_tag_47>': 151734, '<stitch_tag_48>': 151735, '<stitch_tag_49>': 151736, '<stitch_tag_50>': 151737, '<stitch_tag_51>': 151738, '<stitch_tag_52>': 151739, '<stitch_tag_53>': 151740, '<stitch_tag_54>': 151741, '<stitch_tag_55>': 151742, '<stitch_tag_56>': 151743, '<stitch_tag_57>': 151744, '<stitch_tag_58>': 151745, '<stitch_tag_59>': 151746, '<stitch_tag_60>': 151747, '<stitch_tag_61>': 151748, '<stitch_tag_62>': 151749, '<stitch_tag_63>': 151750, '<stitch_tag_64>': 151751, '<stitch_tag_65>': 151752, '<stitch_tag_66>': 151753, '<stitch_tag_67>': 151754, '<stitch_tag_68>': 151755, '<stitch_tag_69>': 151756, '<stitch_tag_70>': 151757, '<stitch_tag_71>': 151758, '<stitch_tag_72>': 151759, '<stitch_tag_73>': 151760, '<stitch_tag_74>': 151761, '<stitch_tag_75>': 151762, '<stitch_tag_76>': 151763, '<stitch_tag_77>': 151764, '<stitch_tag_78>': 151765, '<stitch_tag_79>': 151766, '<stitch_tag_80>': 151767, '<stitch_tag_81>': 151768, '<stitch_tag_82>': 151769, '<stitch_tag_83>': 151770, '<stitch_tag_84>': 151771, '<stitch_tag_85>': 151772, '<stitch_tag_86>': 151773, '<stitch_tag_87>': 151774, '<stitch_tag_88>': 151775, '<stitch_tag_89>': 151776, '<stitch_tag_90>': 151777, '<stitch_tag_91>': 151778, '<stitch_tag_92>': 151779, '<stitch_tag_93>': 151780, '<stitch_tag_94>': 151781, '<stitch_tag_95>': 151782, '<stitch_tag_96>': 151783, '<stitch_tag_97>': 151784, '<stitch_tag_98>': 151785, '<stitch_tag_99>': 151786, '<stitch_tag_100>': 151787, '<stitch_tag_101>': 151788, '<stitch_tag_102>': 151789, '<stitch_tag_103>': 151790, '<stitch_tag_104>': 151791, '<stitch_tag_105>': 151792, '<stitch_tag_106>': 151793, '<stitch_tag_107>': 151794, '<stitch_tag_null>': 151795}
+[2025-09-03 22:01:17,581][models.internvl.configuration_internvl_chat][INFO] - vision_select_layer: -1
+[2025-09-03 22:01:17,581][models.internvl.configuration_internvl_chat][INFO] - ps_version: v2
+[2025-09-03 22:01:17,581][models.internvl.configuration_internvl_chat][INFO] - min_dynamic_patch: 1
+[2025-09-03 22:01:17,581][models.internvl.configuration_internvl_chat][INFO] - max_dynamic_patch: 12
+[2025-09-03 22:01:17,595][models.st_model_diff][INFO] - num_image_token: 256
+[2025-09-03 22:01:17,595][models.st_model_diff][INFO] - ps_version: v2
+[2025-09-03 22:01:36,641][trainers.trainer][INFO] - NN training Using device: cuda:0
+[2025-09-03 22:01:36,770][trainers.trainer][INFO] - Start training from step 0
+[2025-09-03 22:29:56,518][__main__][INFO] - Working directory : /root/workspace/SwiftTailor3
+[2025-09-03 22:29:56,518][__main__][INFO] - Output directory : /root/workspace/SwiftTailor3/output/train
+[2025-09-03 22:29:56,526][data.data_wrappers.data_wrapper][INFO] - Loading data split from assets/data_configs/garmentcodedata_datasplit.json
+[2025-09-03 22:30:20,298][data.data_wrappers.data_wrapper][INFO] - Dataset split: 114869 / 6381 / 6381
+[2025-09-03 22:30:20,300][__main__][INFO] - Dataset created.
+[2025-09-03 22:30:20,573][__main__][INFO] - Added 122 tokens to the tokenizer.
+[2025-09-03 22:30:20,582][__main__][INFO] - Token name to index dictionary: {'<pattern_cmd_MOVE>': 151674, '<pattern_cmd_LINE>': 151675, '<pattern_cmd_CLINE>': 151676, '<pattern_cmd_CURVE>': 151677, '<pattern_cmd_CCURVE>': 151678, '<pattern_cmd_CUBIC>': 151679, '<pattern_cmd_CCUBIC>': 151680, '<pattern_cmd_ARC>': 151681, '<pattern_cmd_CARC>': 151682, '<panel_start>': 151683, '<panel_end>': 151684, '<pattern_start>': 151685, '<pattern_end>': 151686, '<stitch_tag_0>': 151687, '<stitch_tag_1>': 151688, '<stitch_tag_2>': 151689, '<stitch_tag_3>': 151690, '<stitch_tag_4>': 151691, '<stitch_tag_5>': 151692, '<stitch_tag_6>': 151693, '<stitch_tag_7>': 151694, '<stitch_tag_8>': 151695, '<stitch_tag_9>': 151696, '<stitch_tag_10>': 151697, '<stitch_tag_11>': 151698, '<stitch_tag_12>': 151699, '<stitch_tag_13>': 151700, '<stitch_tag_14>': 151701, '<stitch_tag_15>': 151702, '<stitch_tag_16>': 151703, '<stitch_tag_17>': 151704, '<stitch_tag_18>': 151705, '<stitch_tag_19>': 151706, '<stitch_tag_20>': 151707, '<stitch_tag_21>': 151708, '<stitch_tag_22>': 151709, '<stitch_tag_23>': 151710, '<stitch_tag_24>': 151711, '<stitch_tag_25>': 151712, '<stitch_tag_26>': 151713, '<stitch_tag_27>': 151714, '<stitch_tag_28>': 151715, '<stitch_tag_29>': 151716, '<stitch_tag_30>': 151717, '<stitch_tag_31>': 151718, '<stitch_tag_32>': 151719, '<stitch_tag_33>': 151720, '<stitch_tag_34>': 151721, '<stitch_tag_35>': 151722, '<stitch_tag_36>': 151723, '<stitch_tag_37>': 151724, '<stitch_tag_38>': 151725, '<stitch_tag_39>': 151726, '<stitch_tag_40>': 151727, '<stitch_tag_41>': 151728, '<stitch_tag_42>': 151729, '<stitch_tag_43>': 151730, '<stitch_tag_44>': 151731, '<stitch_tag_45>': 151732, '<stitch_tag_46>': 151733, '<stitch_tag_47>': 151734, '<stitch_tag_48>': 151735, '<stitch_tag_49>': 151736, '<stitch_tag_50>': 151737, '<stitch_tag_51>': 151738, '<stitch_tag_52>': 151739, '<stitch_tag_53>': 151740, '<stitch_tag_54>': 151741, '<stitch_tag_55>': 151742, '<stitch_tag_56>': 151743, '<stitch_tag_57>': 151744, '<stitch_tag_58>': 151745, '<stitch_tag_59>': 151746, '<stitch_tag_60>': 151747, '<stitch_tag_61>': 151748, '<stitch_tag_62>': 151749, '<stitch_tag_63>': 151750, '<stitch_tag_64>': 151751, '<stitch_tag_65>': 151752, '<stitch_tag_66>': 151753, '<stitch_tag_67>': 151754, '<stitch_tag_68>': 151755, '<stitch_tag_69>': 151756, '<stitch_tag_70>': 151757, '<stitch_tag_71>': 151758, '<stitch_tag_72>': 151759, '<stitch_tag_73>': 151760, '<stitch_tag_74>': 151761, '<stitch_tag_75>': 151762, '<stitch_tag_76>': 151763, '<stitch_tag_77>': 151764, '<stitch_tag_78>': 151765, '<stitch_tag_79>': 151766, '<stitch_tag_80>': 151767, '<stitch_tag_81>': 151768, '<stitch_tag_82>': 151769, '<stitch_tag_83>': 151770, '<stitch_tag_84>': 151771, '<stitch_tag_85>': 151772, '<stitch_tag_86>': 151773, '<stitch_tag_87>': 151774, '<stitch_tag_88>': 151775, '<stitch_tag_89>': 151776, '<stitch_tag_90>': 151777, '<stitch_tag_91>': 151778, '<stitch_tag_92>': 151779, '<stitch_tag_93>': 151780, '<stitch_tag_94>': 151781, '<stitch_tag_95>': 151782, '<stitch_tag_96>': 151783, '<stitch_tag_97>': 151784, '<stitch_tag_98>': 151785, '<stitch_tag_99>': 151786, '<stitch_tag_100>': 151787, '<stitch_tag_101>': 151788, '<stitch_tag_102>': 151789, '<stitch_tag_103>': 151790, '<stitch_tag_104>': 151791, '<stitch_tag_105>': 151792, '<stitch_tag_106>': 151793, '<stitch_tag_107>': 151794, '<stitch_tag_null>': 151795}
+[2025-09-03 22:30:20,584][models.internvl.configuration_internvl_chat][INFO] - vision_select_layer: -1
+[2025-09-03 22:30:20,584][models.internvl.configuration_internvl_chat][INFO] - ps_version: v2
+[2025-09-03 22:30:20,584][models.internvl.configuration_internvl_chat][INFO] - min_dynamic_patch: 1
+[2025-09-03 22:30:20,584][models.internvl.configuration_internvl_chat][INFO] - max_dynamic_patch: 12
+[2025-09-03 22:30:20,598][models.st_model_diff][INFO] - num_image_token: 256
+[2025-09-03 22:30:20,598][models.st_model_diff][INFO] - ps_version: v2
+[2025-09-03 22:30:39,345][trainers.trainer][INFO] - NN training Using device: cuda:0
+[2025-09-03 22:30:39,463][trainers.trainer][INFO] - Start training from step 0
+[2025-09-03 22:30:53,626][trainers.trainer][INFO] - Step: [1]	Time  1.714 ( 1.406)	Data  0.001 ( 0.069)	Total Loss 18.1910 (22.1247)	CE Loss 13.3750 (13.9187)	Edge Loss 48.1604 (82.0597)	LR 0.000000 (0.000000)	<pattern_cmd_MOVE> Loss 3.8395 (4.5083)	<pattern_cmd_LINE> Loss 7.9621 (17.1630)	<pattern_cmd_CLINE> Loss 0.0000 (0.0000)	<pattern_cmd_CURVE> Loss 6.2891 (18.7915)	<pattern_cmd_CCURVE> Loss 2.1175 (7.6664)	<pattern_cmd_CUBIC> Loss 7.5026 (16.7568)	<pattern_cmd_CCUBIC> Loss 4.5007 (4.5007)	<pattern_cmd_ARC> Loss 8.6430 (8.0111)	<pattern_cmd_CARC> Loss 11.8064 (11.8491)

train_no_diff/run.log

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e76e9b16a075fc2bdba9dc69eda4b010abff19c6f7dbde89e347ba4dd3941986
+size 69122454

train_no_diff/run_debug.log

@@ -0,0 +1,458 @@
+[2025-08-28 10:41:15,466][__main__][INFO] - Working directory : /root/workspace/SwiftTailor3
+[2025-08-28 10:41:15,468][__main__][INFO] - Output directory : /root/workspace/SwiftTailor3/output/train
+[2025-08-28 10:41:15,499][data.data_wrappers.data_wrapper][INFO] - Loading data split from assets/data_configs/garmentcodedata_datasplit.json
+[2025-08-28 10:48:57,669][__main__][INFO] - Working directory : /root/workspace/SwiftTailor3
+[2025-08-28 10:48:57,671][__main__][INFO] - Output directory : /root/workspace/SwiftTailor3/output/train
+[2025-08-28 10:48:57,702][data.data_wrappers.data_wrapper][INFO] - Loading data split from assets/data_configs/garmentcodedata_datasplit.json
+[2025-08-28 10:52:13,818][__main__][INFO] - Working directory : /root/workspace/SwiftTailor3
+[2025-08-28 10:52:13,819][__main__][INFO] - Output directory : /root/workspace/SwiftTailor3/output/train
+[2025-08-28 10:52:13,851][data.data_wrappers.data_wrapper][INFO] - Loading data split from assets/data_configs/garmentcodedata_datasplit.json
+[2025-08-28 10:54:05,898][data.data_wrappers.data_wrapper][INFO] - Dataset split: 114869 / 6381 / 6381
+[2025-08-28 10:54:05,899][__main__][INFO] - Dataset created.
+[2025-08-28 10:54:06,241][__main__][INFO] - Added 122 tokens to the tokenizer.
+[2025-08-28 10:54:06,270][__main__][INFO] - Token name to index dictionary: {'<pattern_cmd_MOVE>': 151674, '<pattern_cmd_LINE>': 151675, '<pattern_cmd_CLINE>': 151676, '<pattern_cmd_CURVE>': 151677, '<pattern_cmd_CCURVE>': 151678, '<pattern_cmd_CUBIC>': 151679, '<pattern_cmd_CCUBIC>': 151680, '<pattern_cmd_ARC>': 151681, '<pattern_cmd_CARC>': 151682, '<panel_start>': 151683, '<panel_end>': 151684, '<pattern_start>': 151685, '<pattern_end>': 151686, '<stitch_tag_0>': 151687, '<stitch_tag_1>': 151688, '<stitch_tag_2>': 151689, '<stitch_tag_3>': 151690, '<stitch_tag_4>': 151691, '<stitch_tag_5>': 151692, '<stitch_tag_6>': 151693, '<stitch_tag_7>': 151694, '<stitch_tag_8>': 151695, '<stitch_tag_9>': 151696, '<stitch_tag_10>': 151697, '<stitch_tag_11>': 151698, '<stitch_tag_12>': 151699, '<stitch_tag_13>': 151700, '<stitch_tag_14>': 151701, '<stitch_tag_15>': 151702, '<stitch_tag_16>': 151703, '<stitch_tag_17>': 151704, '<stitch_tag_18>': 151705, '<stitch_tag_19>': 151706, '<stitch_tag_20>': 151707, '<stitch_tag_21>': 151708, '<stitch_tag_22>': 151709, '<stitch_tag_23>': 151710, '<stitch_tag_24>': 151711, '<stitch_tag_25>': 151712, '<stitch_tag_26>': 151713, '<stitch_tag_27>': 151714, '<stitch_tag_28>': 151715, '<stitch_tag_29>': 151716, '<stitch_tag_30>': 151717, '<stitch_tag_31>': 151718, '<stitch_tag_32>': 151719, '<stitch_tag_33>': 151720, '<stitch_tag_34>': 151721, '<stitch_tag_35>': 151722, '<stitch_tag_36>': 151723, '<stitch_tag_37>': 151724, '<stitch_tag_38>': 151725, '<stitch_tag_39>': 151726, '<stitch_tag_40>': 151727, '<stitch_tag_41>': 151728, '<stitch_tag_42>': 151729, '<stitch_tag_43>': 151730, '<stitch_tag_44>': 151731, '<stitch_tag_45>': 151732, '<stitch_tag_46>': 151733, '<stitch_tag_47>': 151734, '<stitch_tag_48>': 151735, '<stitch_tag_49>': 151736, '<stitch_tag_50>': 151737, '<stitch_tag_51>': 151738, '<stitch_tag_52>': 151739, '<stitch_tag_53>': 151740, '<stitch_tag_54>': 151741, '<stitch_tag_55>': 151742, '<stitch_tag_56>': 151743, '<stitch_tag_57>': 151744, '<stitch_tag_58>': 151745, '<stitch_tag_59>': 151746, '<stitch_tag_60>': 151747, '<stitch_tag_61>': 151748, '<stitch_tag_62>': 151749, '<stitch_tag_63>': 151750, '<stitch_tag_64>': 151751, '<stitch_tag_65>': 151752, '<stitch_tag_66>': 151753, '<stitch_tag_67>': 151754, '<stitch_tag_68>': 151755, '<stitch_tag_69>': 151756, '<stitch_tag_70>': 151757, '<stitch_tag_71>': 151758, '<stitch_tag_72>': 151759, '<stitch_tag_73>': 151760, '<stitch_tag_74>': 151761, '<stitch_tag_75>': 151762, '<stitch_tag_76>': 151763, '<stitch_tag_77>': 151764, '<stitch_tag_78>': 151765, '<stitch_tag_79>': 151766, '<stitch_tag_80>': 151767, '<stitch_tag_81>': 151768, '<stitch_tag_82>': 151769, '<stitch_tag_83>': 151770, '<stitch_tag_84>': 151771, '<stitch_tag_85>': 151772, '<stitch_tag_86>': 151773, '<stitch_tag_87>': 151774, '<stitch_tag_88>': 151775, '<stitch_tag_89>': 151776, '<stitch_tag_90>': 151777, '<stitch_tag_91>': 151778, '<stitch_tag_92>': 151779, '<stitch_tag_93>': 151780, '<stitch_tag_94>': 151781, '<stitch_tag_95>': 151782, '<stitch_tag_96>': 151783, '<stitch_tag_97>': 151784, '<stitch_tag_98>': 151785, '<stitch_tag_99>': 151786, '<stitch_tag_100>': 151787, '<stitch_tag_101>': 151788, '<stitch_tag_102>': 151789, '<stitch_tag_103>': 151790, '<stitch_tag_104>': 151791, '<stitch_tag_105>': 151792, '<stitch_tag_106>': 151793, '<stitch_tag_107>': 151794, '<stitch_tag_null>': 151795}
+[2025-08-28 10:54:06,275][models.internvl.configuration_internvl_chat][INFO] - vision_select_layer: -1
+[2025-08-28 10:54:06,276][models.internvl.configuration_internvl_chat][INFO] - ps_version: v2
+[2025-08-28 10:54:06,276][models.internvl.configuration_internvl_chat][INFO] - min_dynamic_patch: 1
+[2025-08-28 10:54:06,276][models.internvl.configuration_internvl_chat][INFO] - max_dynamic_patch: 12
+[2025-08-28 10:54:06,276][__main__][INFO] - Loading model from output/train/ckpt_29999/output_dir
+[2025-08-28 10:54:06,321][models.st_model][INFO] - num_image_token: 256
+[2025-08-28 10:54:06,321][models.st_model][INFO] - ps_version: v2
+[2025-08-28 11:00:43,936][trainers.trainer][INFO] - Rank: 0, Progress: [0/6381]
+[2025-08-28 11:00:52,549][trainers.trainer][INFO] - Rank: 0, Progress: [1/6381]
+[2025-08-28 11:00:52,681][trainers.trainer][INFO] - Rank: 0, Progress: [2/6381]
+[2025-08-28 11:00:52,801][trainers.trainer][INFO] - Rank: 0, Progress: [3/6381]
+[2025-08-28 11:00:53,026][trainers.trainer][INFO] - Rank: 0, Progress: [4/6381]
+[2025-08-28 11:00:53,173][trainers.trainer][INFO] - Rank: 0, Progress: [5/6381]
+[2025-08-28 11:00:53,774][trainers.trainer][INFO] - Rank: 0, Progress: [6/6381]
+[2025-08-28 11:00:53,878][trainers.trainer][INFO] - Rank: 0, Progress: [7/6381]
+[2025-08-28 11:00:53,943][trainers.trainer][INFO] - Rank: 0, Progress: [8/6381]
+[2025-08-28 11:00:54,044][trainers.trainer][INFO] - Rank: 0, Progress: [9/6381]
+[2025-08-28 11:00:54,175][trainers.trainer][INFO] - Rank: 0, Progress: [10/6381]
+[2025-08-28 11:00:54,296][trainers.trainer][INFO] - Rank: 0, Progress: [11/6381]
+[2025-08-28 11:00:54,530][trainers.trainer][INFO] - Rank: 0, Progress: [12/6381]
+[2025-08-28 11:00:54,602][trainers.trainer][INFO] - Rank: 0, Progress: [13/6381]
+[2025-08-28 11:00:54,671][trainers.trainer][INFO] - Rank: 0, Progress: [14/6381]
+[2025-08-28 11:00:54,743][trainers.trainer][INFO] - Rank: 0, Progress: [15/6381]
+[2025-08-28 11:00:55,718][trainers.trainer][INFO] - Rank: 0, Progress: [16/6381]
+[2025-08-28 11:00:55,799][trainers.trainer][INFO] - Rank: 0, Progress: [17/6381]
+[2025-08-28 11:00:55,912][trainers.trainer][INFO] - Rank: 0, Progress: [18/6381]
+[2025-08-28 11:00:55,983][trainers.trainer][INFO] - Rank: 0, Progress: [19/6381]
+[2025-08-28 11:00:56,061][trainers.trainer][INFO] - Rank: 0, Progress: [20/6381]
+[2025-08-28 11:00:56,191][trainers.trainer][INFO] - Rank: 0, Progress: [21/6381]
+[2025-08-28 11:00:56,394][trainers.trainer][INFO] - Rank: 0, Progress: [22/6381]
+[2025-08-28 11:00:56,502][trainers.trainer][INFO] - Rank: 0, Progress: [23/6381]
+[2025-08-28 11:00:56,643][trainers.trainer][INFO] - Rank: 0, Progress: [24/6381]
+[2025-08-28 11:00:56,757][trainers.trainer][INFO] - Rank: 0, Progress: [25/6381]
+[2025-08-28 11:00:56,871][trainers.trainer][INFO] - Rank: 0, Progress: [26/6381]
+[2025-08-28 11:00:56,964][trainers.trainer][INFO] - Rank: 0, Progress: [27/6381]
+[2025-08-28 11:00:57,528][trainers.trainer][INFO] - Rank: 0, Progress: [28/6381]
+[2025-08-28 11:00:57,630][trainers.trainer][INFO] - Rank: 0, Progress: [29/6381]
+[2025-08-28 11:00:57,854][trainers.trainer][INFO] - Rank: 0, Progress: [30/6381]
+[2025-08-28 11:00:57,967][trainers.trainer][INFO] - Rank: 0, Progress: [31/6381]
+[2025-08-28 11:00:58,039][trainers.trainer][INFO] - Rank: 0, Progress: [32/6381]
+[2025-08-28 11:00:58,159][trainers.trainer][INFO] - Rank: 0, Progress: [33/6381]
+[2025-08-28 11:00:58,300][trainers.trainer][INFO] - Rank: 0, Progress: [34/6381]
+[2025-08-28 11:00:58,448][trainers.trainer][INFO] - Rank: 0, Progress: [35/6381]
+[2025-08-28 11:00:58,576][trainers.trainer][INFO] - Rank: 0, Progress: [36/6381]
+[2025-08-28 11:00:58,718][trainers.trainer][INFO] - Rank: 0, Progress: [37/6381]
+[2025-08-28 11:00:58,832][trainers.trainer][INFO] - Rank: 0, Progress: [38/6381]
+[2025-08-28 11:00:58,999][trainers.trainer][INFO] - Rank: 0, Progress: [39/6381]
+[2025-08-28 11:00:59,328][trainers.trainer][INFO] - Rank: 0, Progress: [40/6381]
+[2025-08-28 11:00:59,412][trainers.trainer][INFO] - Rank: 0, Progress: [41/6381]
+[2025-08-28 11:00:59,587][trainers.trainer][INFO] - Rank: 0, Progress: [42/6381]
+[2025-08-28 11:00:59,672][trainers.trainer][INFO] - Rank: 0, Progress: [43/6381]
+[2025-08-28 11:00:59,780][trainers.trainer][INFO] - Rank: 0, Progress: [44/6381]
+[2025-08-28 11:00:59,900][trainers.trainer][INFO] - Rank: 0, Progress: [45/6381]
+[2025-08-28 11:01:00,022][trainers.trainer][INFO] - Rank: 0, Progress: [46/6381]
+[2025-08-28 11:01:00,132][trainers.trainer][INFO] - Rank: 0, Progress: [47/6381]
+[2025-08-28 11:01:00,259][trainers.trainer][INFO] - Rank: 0, Progress: [48/6381]
+[2025-08-28 11:01:00,415][trainers.trainer][INFO] - Rank: 0, Progress: [49/6381]
+[2025-08-28 11:01:00,546][trainers.trainer][INFO] - Rank: 0, Progress: [50/6381]
+[2025-08-28 11:01:00,637][trainers.trainer][INFO] - Rank: 0, Progress: [51/6381]
+[2025-08-28 11:01:00,896][trainers.trainer][INFO] - Rank: 0, Progress: [52/6381]
+[2025-08-28 11:01:01,047][trainers.trainer][INFO] - Rank: 0, Progress: [53/6381]
+[2025-08-28 11:01:01,524][trainers.trainer][INFO] - Rank: 0, Progress: [54/6381]
+[2025-08-28 11:01:01,611][trainers.trainer][INFO] - Rank: 0, Progress: [55/6381]
+[2025-08-28 11:01:01,747][trainers.trainer][INFO] - Rank: 0, Progress: [56/6381]
+[2025-08-28 11:01:01,824][trainers.trainer][INFO] - Rank: 0, Progress: [57/6381]
+[2025-08-28 11:01:01,901][trainers.trainer][INFO] - Rank: 0, Progress: [58/6381]
+[2025-08-28 11:01:02,018][trainers.trainer][INFO] - Rank: 0, Progress: [59/6381]
+[2025-08-28 11:01:02,202][trainers.trainer][INFO] - Rank: 0, Progress: [60/6381]
+[2025-08-28 11:01:02,333][trainers.trainer][INFO] - Rank: 0, Progress: [61/6381]
+[2025-08-28 11:01:02,476][trainers.trainer][INFO] - Rank: 0, Progress: [62/6381]
+[2025-08-28 11:01:02,585][trainers.trainer][INFO] - Rank: 0, Progress: [63/6381]
+[2025-08-28 11:01:02,882][trainers.trainer][INFO] - Rank: 0, Progress: [64/6381]
+[2025-08-28 11:01:03,042][trainers.trainer][INFO] - Rank: 0, Progress: [65/6381]
+[2025-08-28 11:01:03,297][trainers.trainer][INFO] - Rank: 0, Progress: [66/6381]
+[2025-08-28 11:01:03,383][trainers.trainer][INFO] - Rank: 0, Progress: [67/6381]
+[2025-08-28 11:01:03,497][trainers.trainer][INFO] - Rank: 0, Progress: [68/6381]
+[2025-08-28 11:01:03,581][trainers.trainer][INFO] - Rank: 0, Progress: [69/6381]
+[2025-08-28 11:01:03,670][trainers.trainer][INFO] - Rank: 0, Progress: [70/6381]
+[2025-08-28 11:01:03,766][trainers.trainer][INFO] - Rank: 0, Progress: [71/6381]
+[2025-08-28 11:01:03,940][trainers.trainer][INFO] - Rank: 0, Progress: [72/6381]
+[2025-08-28 11:01:04,066][trainers.trainer][INFO] - Rank: 0, Progress: [73/6381]
+[2025-08-28 11:01:04,192][trainers.trainer][INFO] - Rank: 0, Progress: [74/6381]
+[2025-08-28 11:01:04,340][trainers.trainer][INFO] - Rank: 0, Progress: [75/6381]
+[2025-08-28 11:01:04,699][trainers.trainer][INFO] - Rank: 0, Progress: [76/6381]
+[2025-08-28 11:01:04,816][trainers.trainer][INFO] - Rank: 0, Progress: [77/6381]
+[2025-08-28 11:01:05,033][trainers.trainer][INFO] - Rank: 0, Progress: [78/6381]
+[2025-08-28 11:01:05,112][trainers.trainer][INFO] - Rank: 0, Progress: [79/6381]
+[2025-08-28 11:01:05,211][trainers.trainer][INFO] - Rank: 0, Progress: [80/6381]
+[2025-08-28 11:01:05,330][trainers.trainer][INFO] - Rank: 0, Progress: [81/6381]
+[2025-08-28 11:01:05,446][trainers.trainer][INFO] - Rank: 0, Progress: [82/6381]
+[2025-08-28 11:01:05,592][trainers.trainer][INFO] - Rank: 0, Progress: [83/6381]
+[2025-08-28 11:01:05,781][trainers.trainer][INFO] - Rank: 0, Progress: [84/6381]
+[2025-08-28 11:01:05,960][trainers.trainer][INFO] - Rank: 0, Progress: [85/6381]
+[2025-08-28 11:01:06,067][trainers.trainer][INFO] - Rank: 0, Progress: [86/6381]
+[2025-08-28 11:01:06,281][trainers.trainer][INFO] - Rank: 0, Progress: [87/6381]
+[2025-08-28 11:01:06,664][trainers.trainer][INFO] - Rank: 0, Progress: [88/6381]
+[2025-08-28 11:01:06,742][trainers.trainer][INFO] - Rank: 0, Progress: [89/6381]
+[2025-08-28 11:01:06,898][trainers.trainer][INFO] - Rank: 0, Progress: [90/6381]
+[2025-08-28 11:01:07,003][trainers.trainer][INFO] - Rank: 0, Progress: [91/6381]
+[2025-08-28 11:01:07,129][trainers.trainer][INFO] - Rank: 0, Progress: [92/6381]
+[2025-08-28 11:01:07,226][trainers.trainer][INFO] - Rank: 0, Progress: [93/6381]
+[2025-08-28 11:01:07,388][trainers.trainer][INFO] - Rank: 0, Progress: [94/6381]
+[2025-08-28 11:01:07,897][trainers.trainer][INFO] - Rank: 0, Progress: [95/6381]
+[2025-08-28 11:01:08,101][trainers.trainer][INFO] - Rank: 0, Progress: [96/6381]
+[2025-08-28 11:01:08,313][trainers.trainer][INFO] - Rank: 0, Progress: [97/6381]
+[2025-08-28 11:01:08,563][trainers.trainer][INFO] - Rank: 0, Progress: [98/6381]
+[2025-08-28 11:01:08,704][trainers.trainer][INFO] - Rank: 0, Progress: [99/6381]
+[2025-08-28 11:01:08,900][trainers.trainer][INFO] - Rank: 0, Progress: [100/6381]
+[2025-08-28 11:01:09,006][trainers.trainer][INFO] - Rank: 0, Progress: [101/6381]
+[2025-08-28 11:01:09,179][trainers.trainer][INFO] - Rank: 0, Progress: [102/6381]
+[2025-08-28 11:01:09,324][trainers.trainer][INFO] - Rank: 0, Progress: [103/6381]
+[2025-08-28 11:01:09,423][trainers.trainer][INFO] - Rank: 0, Progress: [104/6381]
+[2025-08-28 11:01:09,554][trainers.trainer][INFO] - Rank: 0, Progress: [105/6381]
+[2025-08-28 11:01:09,736][trainers.trainer][INFO] - Rank: 0, Progress: [106/6381]
+[2025-08-28 11:01:09,873][trainers.trainer][INFO] - Rank: 0, Progress: [107/6381]
+[2025-08-28 11:01:10,035][trainers.trainer][INFO] - Rank: 0, Progress: [108/6381]
+[2025-08-28 11:01:10,136][trainers.trainer][INFO] - Rank: 0, Progress: [109/6381]
+[2025-08-28 11:01:10,263][trainers.trainer][INFO] - Rank: 0, Progress: [110/6381]
+[2025-08-28 11:01:10,435][trainers.trainer][INFO] - Rank: 0, Progress: [111/6381]
+[2025-08-28 11:01:10,654][trainers.trainer][INFO] - Rank: 0, Progress: [112/6381]
+[2025-08-28 11:01:10,924][trainers.trainer][INFO] - Rank: 0, Progress: [113/6381]
+[2025-08-28 11:01:11,288][trainers.trainer][INFO] - Rank: 0, Progress: [114/6381]
+[2025-08-28 11:01:11,386][trainers.trainer][INFO] - Rank: 0, Progress: [115/6381]
+[2025-08-28 11:01:11,497][trainers.trainer][INFO] - Rank: 0, Progress: [116/6381]
+[2025-08-28 11:01:11,579][trainers.trainer][INFO] - Rank: 0, Progress: [117/6381]
+[2025-08-28 11:01:11,784][trainers.trainer][INFO] - Rank: 0, Progress: [118/6381]
+[2025-08-28 11:01:11,884][trainers.trainer][INFO] - Rank: 0, Progress: [119/6381]
+[2025-08-28 11:01:12,018][trainers.trainer][INFO] - Rank: 0, Progress: [120/6381]
+[2025-08-28 11:01:12,099][trainers.trainer][INFO] - Rank: 0, Progress: [121/6381]
+[2025-08-28 11:01:12,217][trainers.trainer][INFO] - Rank: 0, Progress: [122/6381]
+[2025-08-28 11:01:12,349][trainers.trainer][INFO] - Rank: 0, Progress: [123/6381]
+[2025-08-28 11:01:12,549][trainers.trainer][INFO] - Rank: 0, Progress: [124/6381]
+[2025-08-28 11:01:12,652][trainers.trainer][INFO] - Rank: 0, Progress: [125/6381]
+[2025-08-28 11:01:12,952][trainers.trainer][INFO] - Rank: 0, Progress: [126/6381]
+[2025-08-28 11:01:13,056][trainers.trainer][INFO] - Rank: 0, Progress: [127/6381]
+[2025-08-28 11:01:13,130][trainers.trainer][INFO] - Rank: 0, Progress: [128/6381]
+[2025-08-28 11:01:13,267][trainers.trainer][INFO] - Rank: 0, Progress: [129/6381]
+[2025-08-28 11:01:13,350][trainers.trainer][INFO] - Rank: 0, Progress: [130/6381]
+[2025-08-28 11:01:13,536][trainers.trainer][INFO] - Rank: 0, Progress: [131/6381]
+[2025-08-28 11:01:13,678][trainers.trainer][INFO] - Rank: 0, Progress: [132/6381]
+[2025-08-28 11:01:13,765][trainers.trainer][INFO] - Rank: 0, Progress: [133/6381]
+[2025-08-28 11:01:13,940][trainers.trainer][INFO] - Rank: 0, Progress: [134/6381]
+[2025-08-28 11:01:14,051][trainers.trainer][INFO] - Rank: 0, Progress: [135/6381]
+[2025-08-28 11:01:14,290][trainers.trainer][INFO] - Rank: 0, Progress: [136/6381]
+[2025-08-28 11:01:14,449][trainers.trainer][INFO] - Rank: 0, Progress: [137/6381]
+[2025-08-28 11:01:14,614][trainers.trainer][INFO] - Rank: 0, Progress: [138/6381]
+[2025-08-28 11:01:14,747][trainers.trainer][INFO] - Rank: 0, Progress: [139/6381]
+[2025-08-28 11:01:14,964][trainers.trainer][INFO] - Rank: 0, Progress: [140/6381]
+[2025-08-28 11:01:15,053][trainers.trainer][INFO] - Rank: 0, Progress: [141/6381]
+[2025-08-28 11:01:15,159][trainers.trainer][INFO] - Rank: 0, Progress: [142/6381]
+[2025-08-28 11:01:15,271][trainers.trainer][INFO] - Rank: 0, Progress: [143/6381]
+[2025-08-28 11:01:15,468][trainers.trainer][INFO] - Rank: 0, Progress: [144/6381]
+[2025-08-28 11:01:15,657][trainers.trainer][INFO] - Rank: 0, Progress: [145/6381]
+[2025-08-28 11:01:15,756][trainers.trainer][INFO] - Rank: 0, Progress: [146/6381]
+[2025-08-28 11:01:15,828][trainers.trainer][INFO] - Rank: 0, Progress: [147/6381]
+[2025-08-28 11:01:15,990][trainers.trainer][INFO] - Rank: 0, Progress: [148/6381]
+[2025-08-28 11:01:16,097][trainers.trainer][INFO] - Rank: 0, Progress: [149/6381]
+[2025-08-28 11:01:16,417][trainers.trainer][INFO] - Rank: 0, Progress: [150/6381]
+[2025-08-28 11:01:16,501][trainers.trainer][INFO] - Rank: 0, Progress: [151/6381]
+[2025-08-28 11:01:16,574][trainers.trainer][INFO] - Rank: 0, Progress: [152/6381]
+[2025-08-28 11:01:16,687][trainers.trainer][INFO] - Rank: 0, Progress: [153/6381]
+[2025-08-28 11:01:16,783][trainers.trainer][INFO] - Rank: 0, Progress: [154/6381]
+[2025-08-28 11:01:16,905][trainers.trainer][INFO] - Rank: 0, Progress: [155/6381]
+[2025-08-28 11:01:17,066][trainers.trainer][INFO] - Rank: 0, Progress: [156/6381]
+[2025-08-28 11:01:17,214][trainers.trainer][INFO] - Rank: 0, Progress: [157/6381]
+[2025-08-28 11:01:17,529][trainers.trainer][INFO] - Rank: 0, Progress: [158/6381]
+[2025-08-28 11:01:17,632][trainers.trainer][INFO] - Rank: 0, Progress: [159/6381]
+[2025-08-28 11:01:17,797][trainers.trainer][INFO] - Rank: 0, Progress: [160/6381]
+[2025-08-28 11:01:17,910][trainers.trainer][INFO] - Rank: 0, Progress: [161/6381]
+[2025-08-28 11:01:18,093][trainers.trainer][INFO] - Rank: 0, Progress: [162/6381]
+[2025-08-28 11:01:18,375][trainers.trainer][INFO] - Rank: 0, Progress: [163/6381]
+[2025-08-28 11:01:18,544][trainers.trainer][INFO] - Rank: 0, Progress: [164/6381]
+[2025-08-28 11:01:18,626][trainers.trainer][INFO] - Rank: 0, Progress: [165/6381]
+[2025-08-28 11:01:18,725][trainers.trainer][INFO] - Rank: 0, Progress: [166/6381]
+[2025-08-28 11:01:18,843][trainers.trainer][INFO] - Rank: 0, Progress: [167/6381]
+[2025-08-28 11:01:19,164][trainers.trainer][INFO] - Rank: 0, Progress: [168/6381]
+[2025-08-28 11:01:19,316][trainers.trainer][INFO] - Rank: 0, Progress: [169/6381]
+[2025-08-28 11:01:19,439][trainers.trainer][INFO] - Rank: 0, Progress: [170/6381]
+[2025-08-28 11:01:19,582][trainers.trainer][INFO] - Rank: 0, Progress: [171/6381]
+[2025-08-28 11:01:19,799][trainers.trainer][INFO] - Rank: 0, Progress: [172/6381]
+[2025-08-28 11:01:19,908][trainers.trainer][INFO] - Rank: 0, Progress: [173/6381]
+[2025-08-28 11:01:20,085][trainers.trainer][INFO] - Rank: 0, Progress: [174/6381]
+[2025-08-28 11:01:20,342][trainers.trainer][INFO] - Rank: 0, Progress: [175/6381]
+[2025-08-28 11:01:20,492][trainers.trainer][INFO] - Rank: 0, Progress: [176/6381]
+[2025-08-28 11:01:20,622][trainers.trainer][INFO] - Rank: 0, Progress: [177/6381]
+[2025-08-28 11:01:20,725][trainers.trainer][INFO] - Rank: 0, Progress: [178/6381]
+[2025-08-28 11:01:20,845][trainers.trainer][INFO] - Rank: 0, Progress: [179/6381]
+[2025-08-28 11:01:21,085][trainers.trainer][INFO] - Rank: 0, Progress: [180/6381]
+[2025-08-28 11:01:21,197][trainers.trainer][INFO] - Rank: 0, Progress: [181/6381]
+[2025-08-28 11:01:21,335][trainers.trainer][INFO] - Rank: 0, Progress: [182/6381]
+[2025-08-28 11:01:21,594][trainers.trainer][INFO] - Rank: 0, Progress: [183/6381]
+[2025-08-28 11:01:21,744][trainers.trainer][INFO] - Rank: 0, Progress: [184/6381]
+[2025-08-28 11:01:21,825][trainers.trainer][INFO] - Rank: 0, Progress: [185/6381]
+[2025-08-28 11:01:21,964][trainers.trainer][INFO] - Rank: 0, Progress: [186/6381]
+[2025-08-28 11:01:22,039][trainers.trainer][INFO] - Rank: 0, Progress: [187/6381]
+[2025-08-28 11:01:22,107][trainers.trainer][INFO] - Rank: 0, Progress: [188/6381]
+[2025-08-28 11:01:22,201][trainers.trainer][INFO] - Rank: 0, Progress: [189/6381]
+[2025-08-28 11:01:22,300][trainers.trainer][INFO] - Rank: 0, Progress: [190/6381]
+[2025-08-28 11:01:22,452][trainers.trainer][INFO] - Rank: 0, Progress: [191/6381]
+[2025-08-28 11:01:22,696][trainers.trainer][INFO] - Rank: 0, Progress: [192/6381]
+[2025-08-28 11:01:22,808][trainers.trainer][INFO] - Rank: 0, Progress: [193/6381]
+[2025-08-28 11:01:22,942][trainers.trainer][INFO] - Rank: 0, Progress: [194/6381]
+[2025-08-28 11:01:23,059][trainers.trainer][INFO] - Rank: 0, Progress: [195/6381]
+[2025-08-28 11:01:23,266][trainers.trainer][INFO] - Rank: 0, Progress: [196/6381]
+[2025-08-28 11:01:23,386][trainers.trainer][INFO] - Rank: 0, Progress: [197/6381]
+[2025-08-28 11:01:23,577][trainers.trainer][INFO] - Rank: 0, Progress: [198/6381]
+[2025-08-28 11:01:23,770][trainers.trainer][INFO] - Rank: 0, Progress: [199/6381]
+[2025-08-28 11:01:23,873][trainers.trainer][INFO] - Rank: 0, Progress: [200/6381]
+[2025-08-28 11:01:23,979][trainers.trainer][INFO] - Rank: 0, Progress: [201/6381]
+[2025-08-28 11:01:24,121][trainers.trainer][INFO] - Rank: 0, Progress: [202/6381]
+[2025-08-28 11:01:24,250][trainers.trainer][INFO] - Rank: 0, Progress: [203/6381]
+[2025-08-28 11:01:24,463][trainers.trainer][INFO] - Rank: 0, Progress: [204/6381]
+[2025-08-28 11:01:24,592][trainers.trainer][INFO] - Rank: 0, Progress: [205/6381]
+[2025-08-28 11:01:24,922][trainers.trainer][INFO] - Rank: 0, Progress: [206/6381]
+[2025-08-28 11:01:25,125][trainers.trainer][INFO] - Rank: 0, Progress: [207/6381]
+[2025-08-28 11:01:25,248][trainers.trainer][INFO] - Rank: 0, Progress: [208/6381]
+[2025-08-28 11:01:25,456][trainers.trainer][INFO] - Rank: 0, Progress: [209/6381]
+[2025-08-28 11:01:25,587][trainers.trainer][INFO] - Rank: 0, Progress: [210/6381]
+[2025-08-28 11:01:25,845][trainers.trainer][INFO] - Rank: 0, Progress: [211/6381]
+[2025-08-28 11:01:25,951][trainers.trainer][INFO] - Rank: 0, Progress: [212/6381]
+[2025-08-28 11:01:26,036][trainers.trainer][INFO] - Rank: 0, Progress: [213/6381]
+[2025-08-28 11:01:26,185][trainers.trainer][INFO] - Rank: 0, Progress: [214/6381]
+[2025-08-28 11:01:26,324][trainers.trainer][INFO] - Rank: 0, Progress: [215/6381]
+[2025-08-28 11:01:26,536][trainers.trainer][INFO] - Rank: 0, Progress: [216/6381]
+[2025-08-28 11:01:26,676][trainers.trainer][INFO] - Rank: 0, Progress: [217/6381]
+[2025-08-28 11:01:26,757][trainers.trainer][INFO] - Rank: 0, Progress: [218/6381]
+[2025-08-28 11:01:26,838][trainers.trainer][INFO] - Rank: 0, Progress: [219/6381]
+[2025-08-28 11:01:26,951][trainers.trainer][INFO] - Rank: 0, Progress: [220/6381]
+[2025-08-28 11:01:27,082][trainers.trainer][INFO] - Rank: 0, Progress: [221/6381]
+[2025-08-28 11:01:27,256][trainers.trainer][INFO] - Rank: 0, Progress: [222/6381]
+[2025-08-28 11:01:27,460][trainers.trainer][INFO] - Rank: 0, Progress: [223/6381]
+[2025-08-28 11:01:27,596][trainers.trainer][INFO] - Rank: 0, Progress: [224/6381]
+[2025-08-28 11:01:27,739][trainers.trainer][INFO] - Rank: 0, Progress: [225/6381]
+[2025-08-28 11:01:27,855][trainers.trainer][INFO] - Rank: 0, Progress: [226/6381]
+[2025-08-28 11:01:27,936][trainers.trainer][INFO] - Rank: 0, Progress: [227/6381]
+[2025-08-28 11:01:28,132][trainers.trainer][INFO] - Rank: 0, Progress: [228/6381]
+[2025-08-28 11:01:28,203][trainers.trainer][INFO] - Rank: 0, Progress: [229/6381]
+[2025-08-28 11:01:28,310][trainers.trainer][INFO] - Rank: 0, Progress: [230/6381]
+[2025-08-28 11:01:28,688][trainers.trainer][INFO] - Rank: 0, Progress: [231/6381]
+[2025-08-28 11:01:28,797][trainers.trainer][INFO] - Rank: 0, Progress: [232/6381]
+[2025-08-28 11:01:28,928][trainers.trainer][INFO] - Rank: 0, Progress: [233/6381]
+[2025-08-28 11:01:29,144][trainers.trainer][INFO] - Rank: 0, Progress: [234/6381]
+[2025-08-28 11:01:29,262][trainers.trainer][INFO] - Rank: 0, Progress: [235/6381]
+[2025-08-28 11:01:29,352][trainers.trainer][INFO] - Rank: 0, Progress: [236/6381]
+[2025-08-28 11:01:29,435][trainers.trainer][INFO] - Rank: 0, Progress: [237/6381]
+[2025-08-28 11:01:29,550][trainers.trainer][INFO] - Rank: 0, Progress: [238/6381]
+[2025-08-28 11:01:29,701][trainers.trainer][INFO] - Rank: 0, Progress: [239/6381]
+[2025-08-28 11:01:29,869][trainers.trainer][INFO] - Rank: 0, Progress: [240/6381]
+[2025-08-28 11:01:29,984][trainers.trainer][INFO] - Rank: 0, Progress: [241/6381]
+[2025-08-28 11:01:30,257][trainers.trainer][INFO] - Rank: 0, Progress: [242/6381]
+[2025-08-28 11:01:30,467][trainers.trainer][INFO] - Rank: 0, Progress: [243/6381]
+[2025-08-28 11:01:30,679][trainers.trainer][INFO] - Rank: 0, Progress: [244/6381]
+[2025-08-28 11:01:30,889][trainers.trainer][INFO] - Rank: 0, Progress: [245/6381]
+[2025-08-28 11:01:31,327][trainers.trainer][INFO] - Rank: 0, Progress: [246/6381]
+[2025-08-28 11:01:31,411][trainers.trainer][INFO] - Rank: 0, Progress: [247/6381]
+[2025-08-28 11:01:31,610][trainers.trainer][INFO] - Rank: 0, Progress: [248/6381]
+[2025-08-28 11:01:31,693][trainers.trainer][INFO] - Rank: 0, Progress: [249/6381]
+[2025-08-28 11:01:31,816][trainers.trainer][INFO] - Rank: 0, Progress: [250/6381]
+[2025-08-28 11:01:31,915][trainers.trainer][INFO] - Rank: 0, Progress: [251/6381]
+[2025-08-28 11:01:32,077][trainers.trainer][INFO] - Rank: 0, Progress: [252/6381]
+[2025-08-28 11:01:32,169][trainers.trainer][INFO] - Rank: 0, Progress: [253/6381]
+[2025-08-28 11:01:32,242][trainers.trainer][INFO] - Rank: 0, Progress: [254/6381]
+[2025-08-28 11:01:32,325][trainers.trainer][INFO] - Rank: 0, Progress: [255/6381]
+[2025-08-28 11:01:32,510][trainers.trainer][INFO] - Rank: 0, Progress: [256/6381]
+[2025-08-28 11:01:32,633][trainers.trainer][INFO] - Rank: 0, Progress: [257/6381]
+[2025-08-28 11:01:33,043][trainers.trainer][INFO] - Rank: 0, Progress: [258/6381]
+[2025-08-28 11:01:33,157][trainers.trainer][INFO] - Rank: 0, Progress: [259/6381]
+[2025-08-28 11:01:33,262][trainers.trainer][INFO] - Rank: 0, Progress: [260/6381]
+[2025-08-28 11:01:33,385][trainers.trainer][INFO] - Rank: 0, Progress: [261/6381]
+[2025-08-28 11:01:33,506][trainers.trainer][INFO] - Rank: 0, Progress: [262/6381]
+[2025-08-28 11:01:33,626][trainers.trainer][INFO] - Rank: 0, Progress: [263/6381]
+[2025-08-28 11:01:33,836][trainers.trainer][INFO] - Rank: 0, Progress: [264/6381]
+[2025-08-28 11:01:33,942][trainers.trainer][INFO] - Rank: 0, Progress: [265/6381]
+[2025-08-28 11:01:34,035][trainers.trainer][INFO] - Rank: 0, Progress: [266/6381]
+[2025-08-28 11:01:34,114][trainers.trainer][INFO] - Rank: 0, Progress: [267/6381]
+[2025-08-28 11:01:34,236][trainers.trainer][INFO] - Rank: 0, Progress: [268/6381]
+[2025-08-28 11:01:34,391][trainers.trainer][INFO] - Rank: 0, Progress: [269/6381]
+[2025-08-28 11:01:34,614][trainers.trainer][INFO] - Rank: 0, Progress: [270/6381]
+[2025-08-28 11:01:34,688][trainers.trainer][INFO] - Rank: 0, Progress: [271/6381]
+[2025-08-28 11:01:34,797][trainers.trainer][INFO] - Rank: 0, Progress: [272/6381]
+[2025-08-28 11:01:34,919][trainers.trainer][INFO] - Rank: 0, Progress: [273/6381]
+[2025-08-28 11:01:35,027][trainers.trainer][INFO] - Rank: 0, Progress: [274/6381]
+[2025-08-28 11:01:35,226][trainers.trainer][INFO] - Rank: 0, Progress: [275/6381]
+[2025-08-28 11:01:35,477][trainers.trainer][INFO] - Rank: 0, Progress: [276/6381]
+[2025-08-28 11:01:35,608][trainers.trainer][INFO] - Rank: 0, Progress: [277/6381]
+[2025-08-28 11:01:35,684][trainers.trainer][INFO] - Rank: 0, Progress: [278/6381]
+[2025-08-28 11:01:35,781][trainers.trainer][INFO] - Rank: 0, Progress: [279/6381]
+[2025-08-28 11:01:35,878][trainers.trainer][INFO] - Rank: 0, Progress: [280/6381]
+[2025-08-28 11:01:35,985][trainers.trainer][INFO] - Rank: 0, Progress: [281/6381]
+[2025-08-28 11:01:36,196][trainers.trainer][INFO] - Rank: 0, Progress: [282/6381]
+[2025-08-28 11:01:36,305][trainers.trainer][INFO] - Rank: 0, Progress: [283/6381]
+[2025-08-28 11:01:36,374][trainers.trainer][INFO] - Rank: 0, Progress: [284/6381]
+[2025-08-28 11:01:36,457][trainers.trainer][INFO] - Rank: 0, Progress: [285/6381]
+[2025-08-28 11:01:36,557][trainers.trainer][INFO] - Rank: 0, Progress: [286/6381]
+[2025-08-28 11:01:36,727][trainers.trainer][INFO] - Rank: 0, Progress: [287/6381]
+[2025-08-28 11:01:36,906][trainers.trainer][INFO] - Rank: 0, Progress: [288/6381]
+[2025-08-28 11:01:37,056][trainers.trainer][INFO] - Rank: 0, Progress: [289/6381]
+[2025-08-28 11:01:37,214][trainers.trainer][INFO] - Rank: 0, Progress: [290/6381]
+[2025-08-28 11:01:37,297][trainers.trainer][INFO] - Rank: 0, Progress: [291/6381]
+[2025-08-28 11:01:37,425][trainers.trainer][INFO] - Rank: 0, Progress: [292/6381]
+[2025-08-28 11:01:37,589][trainers.trainer][INFO] - Rank: 0, Progress: [293/6381]
+[2025-08-28 11:01:37,865][trainers.trainer][INFO] - Rank: 0, Progress: [294/6381]
+[2025-08-28 11:01:38,063][trainers.trainer][INFO] - Rank: 0, Progress: [295/6381]
+[2025-08-28 11:01:38,161][trainers.trainer][INFO] - Rank: 0, Progress: [296/6381]
+[2025-08-28 11:01:38,286][trainers.trainer][INFO] - Rank: 0, Progress: [297/6381]
+[2025-08-28 11:01:38,373][trainers.trainer][INFO] - Rank: 0, Progress: [298/6381]
+[2025-08-28 11:01:38,509][trainers.trainer][INFO] - Rank: 0, Progress: [299/6381]
+[2025-08-28 11:01:38,755][trainers.trainer][INFO] - Rank: 0, Progress: [300/6381]
+[2025-08-28 11:01:38,845][trainers.trainer][INFO] - Rank: 0, Progress: [301/6381]
+[2025-08-28 11:01:38,961][trainers.trainer][INFO] - Rank: 0, Progress: [302/6381]
+[2025-08-28 11:01:39,090][trainers.trainer][INFO] - Rank: 0, Progress: [303/6381]
+[2025-08-28 11:01:39,206][trainers.trainer][INFO] - Rank: 0, Progress: [304/6381]
+[2025-08-28 11:01:39,313][trainers.trainer][INFO] - Rank: 0, Progress: [305/6381]
+[2025-08-28 11:01:39,469][trainers.trainer][INFO] - Rank: 0, Progress: [306/6381]
+[2025-08-28 11:01:39,545][trainers.trainer][INFO] - Rank: 0, Progress: [307/6381]
+[2025-08-28 11:01:39,630][trainers.trainer][INFO] - Rank: 0, Progress: [308/6381]
+[2025-08-28 11:01:39,747][trainers.trainer][INFO] - Rank: 0, Progress: [309/6381]
+[2025-08-28 11:01:39,898][trainers.trainer][INFO] - Rank: 0, Progress: [310/6381]
+[2025-08-28 11:01:39,989][trainers.trainer][INFO] - Rank: 0, Progress: [311/6381]
+[2025-08-28 11:01:40,199][trainers.trainer][INFO] - Rank: 0, Progress: [312/6381]
+[2025-08-28 11:01:40,497][trainers.trainer][INFO] - Rank: 0, Progress: [313/6381]
+[2025-08-28 11:01:40,684][trainers.trainer][INFO] - Rank: 0, Progress: [314/6381]
+[2025-08-28 11:01:40,937][trainers.trainer][INFO] - Rank: 0, Progress: [315/6381]
+[2025-08-28 11:01:41,072][trainers.trainer][INFO] - Rank: 0, Progress: [316/6381]
+[2025-08-28 11:01:41,218][trainers.trainer][INFO] - Rank: 0, Progress: [317/6381]
+[2025-08-28 11:01:41,363][trainers.trainer][INFO] - Rank: 0, Progress: [318/6381]
+[2025-08-28 11:01:41,445][trainers.trainer][INFO] - Rank: 0, Progress: [319/6381]
+[2025-08-28 11:01:41,566][trainers.trainer][INFO] - Rank: 0, Progress: [320/6381]
+[2025-08-28 11:01:41,716][trainers.trainer][INFO] - Rank: 0, Progress: [321/6381]
+[2025-08-28 11:01:41,851][trainers.trainer][INFO] - Rank: 0, Progress: [322/6381]
+[2025-08-28 11:01:41,954][trainers.trainer][INFO] - Rank: 0, Progress: [323/6381]
+[2025-08-28 11:01:42,136][trainers.trainer][INFO] - Rank: 0, Progress: [324/6381]
+[2025-08-28 11:01:42,212][trainers.trainer][INFO] - Rank: 0, Progress: [325/6381]
+[2025-08-28 11:01:42,304][trainers.trainer][INFO] - Rank: 0, Progress: [326/6381]
+[2025-08-28 11:01:42,517][trainers.trainer][INFO] - Rank: 0, Progress: [327/6381]
+[2025-08-28 11:01:42,636][trainers.trainer][INFO] - Rank: 0, Progress: [328/6381]
+[2025-08-28 11:01:42,758][trainers.trainer][INFO] - Rank: 0, Progress: [329/6381]
+[2025-08-28 11:01:42,958][trainers.trainer][INFO] - Rank: 0, Progress: [330/6381]
+[2025-08-28 11:01:43,056][trainers.trainer][INFO] - Rank: 0, Progress: [331/6381]
+[2025-08-28 11:01:43,177][trainers.trainer][INFO] - Rank: 0, Progress: [332/6381]
+[2025-08-28 11:01:43,293][trainers.trainer][INFO] - Rank: 0, Progress: [333/6381]
+[2025-08-28 11:01:43,412][trainers.trainer][INFO] - Rank: 0, Progress: [334/6381]
+[2025-08-28 11:01:43,552][trainers.trainer][INFO] - Rank: 0, Progress: [335/6381]
+[2025-08-28 11:01:43,734][trainers.trainer][INFO] - Rank: 0, Progress: [336/6381]
+[2025-08-28 11:01:43,876][trainers.trainer][INFO] - Rank: 0, Progress: [337/6381]
+[2025-08-28 11:01:44,006][trainers.trainer][INFO] - Rank: 0, Progress: [338/6381]
+[2025-08-28 11:01:44,128][trainers.trainer][INFO] - Rank: 0, Progress: [339/6381]
+[2025-08-28 11:01:44,204][trainers.trainer][INFO] - Rank: 0, Progress: [340/6381]
+[2025-08-28 11:01:44,340][trainers.trainer][INFO] - Rank: 0, Progress: [341/6381]
+[2025-08-28 11:01:44,493][trainers.trainer][INFO] - Rank: 0, Progress: [342/6381]
+[2025-08-28 11:01:44,773][trainers.trainer][INFO] - Rank: 0, Progress: [343/6381]
+[2025-08-28 11:01:44,909][trainers.trainer][INFO] - Rank: 0, Progress: [344/6381]
+[2025-08-28 11:01:45,042][trainers.trainer][INFO] - Rank: 0, Progress: [345/6381]
+[2025-08-28 11:01:45,252][trainers.trainer][INFO] - Rank: 0, Progress: [346/6381]
+[2025-08-28 11:01:45,389][trainers.trainer][INFO] - Rank: 0, Progress: [347/6381]
+[2025-08-28 11:01:45,597][trainers.trainer][INFO] - Rank: 0, Progress: [348/6381]
+[2025-08-28 11:01:45,762][trainers.trainer][INFO] - Rank: 0, Progress: [349/6381]
+[2025-08-28 11:01:45,878][trainers.trainer][INFO] - Rank: 0, Progress: [350/6381]
+[2025-08-28 11:01:46,035][trainers.trainer][INFO] - Rank: 0, Progress: [351/6381]
+[2025-08-28 11:01:46,153][trainers.trainer][INFO] - Rank: 0, Progress: [352/6381]
+[2025-08-28 11:01:46,307][trainers.trainer][INFO] - Rank: 0, Progress: [353/6381]
+[2025-08-28 11:01:46,666][trainers.trainer][INFO] - Rank: 0, Progress: [354/6381]
+[2025-08-28 11:01:46,814][trainers.trainer][INFO] - Rank: 0, Progress: [355/6381]
+[2025-08-28 11:01:46,961][trainers.trainer][INFO] - Rank: 0, Progress: [356/6381]
+[2025-08-28 11:01:47,086][trainers.trainer][INFO] - Rank: 0, Progress: [357/6381]
+[2025-08-28 11:01:47,206][trainers.trainer][INFO] - Rank: 0, Progress: [358/6381]
+[2025-08-28 11:01:47,349][trainers.trainer][INFO] - Rank: 0, Progress: [359/6381]
+[2025-08-28 11:01:47,566][trainers.trainer][INFO] - Rank: 0, Progress: [360/6381]
+[2025-08-28 11:01:47,727][trainers.trainer][INFO] - Rank: 0, Progress: [361/6381]
+[2025-08-28 11:01:47,940][trainers.trainer][INFO] - Rank: 0, Progress: [362/6381]
+[2025-08-28 11:01:48,059][trainers.trainer][INFO] - Rank: 0, Progress: [363/6381]
+[2025-08-28 11:01:48,215][trainers.trainer][INFO] - Rank: 0, Progress: [364/6381]
+[2025-08-28 11:01:48,285][trainers.trainer][INFO] - Rank: 0, Progress: [365/6381]
+[2025-08-28 11:01:48,555][trainers.trainer][INFO] - Rank: 0, Progress: [366/6381]
+[2025-08-28 11:01:48,794][trainers.trainer][INFO] - Rank: 0, Progress: [367/6381]
+[2025-08-28 11:01:49,134][trainers.trainer][INFO] - Rank: 0, Progress: [368/6381]
+[2025-08-28 11:01:49,275][trainers.trainer][INFO] - Rank: 0, Progress: [369/6381]
+[2025-08-28 11:01:49,364][trainers.trainer][INFO] - Rank: 0, Progress: [370/6381]
+[2025-08-28 11:01:49,457][trainers.trainer][INFO] - Rank: 0, Progress: [371/6381]
+[2025-08-28 11:01:49,692][trainers.trainer][INFO] - Rank: 0, Progress: [372/6381]
+[2025-08-28 11:01:49,784][trainers.trainer][INFO] - Rank: 0, Progress: [373/6381]
+[2025-08-28 11:01:49,957][trainers.trainer][INFO] - Rank: 0, Progress: [374/6381]
+[2025-08-28 11:01:50,106][trainers.trainer][INFO] - Rank: 0, Progress: [375/6381]
+[2025-08-28 11:01:50,190][trainers.trainer][INFO] - Rank: 0, Progress: [376/6381]
+[2025-08-28 11:01:50,276][trainers.trainer][INFO] - Rank: 0, Progress: [377/6381]
+[2025-08-28 11:01:50,457][trainers.trainer][INFO] - Rank: 0, Progress: [378/6381]
+[2025-08-28 11:01:50,571][trainers.trainer][INFO] - Rank: 0, Progress: [379/6381]
+[2025-08-28 11:01:50,756][trainers.trainer][INFO] - Rank: 0, Progress: [380/6381]
+[2025-08-28 11:01:50,931][trainers.trainer][INFO] - Rank: 0, Progress: [381/6381]
+[2025-08-28 11:01:51,080][trainers.trainer][INFO] - Rank: 0, Progress: [382/6381]
+[2025-08-28 11:01:51,228][trainers.trainer][INFO] - Rank: 0, Progress: [383/6381]
+[2025-08-28 11:01:51,444][trainers.trainer][INFO] - Rank: 0, Progress: [384/6381]
+[2025-08-28 11:01:51,523][trainers.trainer][INFO] - Rank: 0, Progress: [385/6381]
+[2025-08-28 11:01:51,629][trainers.trainer][INFO] - Rank: 0, Progress: [386/6381]
+[2025-08-28 11:01:51,858][trainers.trainer][INFO] - Rank: 0, Progress: [387/6381]
+[2025-08-28 11:01:52,092][trainers.trainer][INFO] - Rank: 0, Progress: [388/6381]
+[2025-08-28 11:01:52,242][trainers.trainer][INFO] - Rank: 0, Progress: [389/6381]
+[2025-08-28 11:01:52,374][trainers.trainer][INFO] - Rank: 0, Progress: [390/6381]
+[2025-08-28 11:01:52,446][trainers.trainer][INFO] - Rank: 0, Progress: [391/6381]
+[2025-08-28 11:01:52,582][trainers.trainer][INFO] - Rank: 0, Progress: [392/6381]
+[2025-08-28 11:01:52,772][trainers.trainer][INFO] - Rank: 0, Progress: [393/6381]
+[2025-08-28 11:01:52,909][trainers.trainer][INFO] - Rank: 0, Progress: [394/6381]
+[2025-08-28 11:01:53,004][trainers.trainer][INFO] - Rank: 0, Progress: [395/6381]
+[2025-08-28 11:01:53,237][trainers.trainer][INFO] - Rank: 0, Progress: [396/6381]
+[2025-08-28 11:01:53,369][trainers.trainer][INFO] - Rank: 0, Progress: [397/6381]
+[2025-08-28 11:01:53,442][trainers.trainer][INFO] - Rank: 0, Progress: [398/6381]
+[2025-08-28 11:01:53,528][trainers.trainer][INFO] - Rank: 0, Progress: [399/6381]
+[2025-08-28 11:01:53,715][trainers.trainer][INFO] - Rank: 0, Progress: [400/6381]
+[2025-08-28 11:01:53,930][trainers.trainer][INFO] - Rank: 0, Progress: [401/6381]
+[2025-08-28 11:01:54,060][trainers.trainer][INFO] - Rank: 0, Progress: [402/6381]
+[2025-08-28 11:01:54,206][trainers.trainer][INFO] - Rank: 0, Progress: [403/6381]
+[2025-08-28 11:01:54,314][trainers.trainer][INFO] - Rank: 0, Progress: [404/6381]
+[2025-08-28 11:01:54,492][trainers.trainer][INFO] - Rank: 0, Progress: [405/6381]
+[2025-08-28 11:01:54,573][trainers.trainer][INFO] - Rank: 0, Progress: [406/6381]
+[2025-08-28 11:01:54,701][trainers.trainer][INFO] - Rank: 0, Progress: [407/6381]
+[2025-08-28 11:01:55,002][trainers.trainer][INFO] - Rank: 0, Progress: [408/6381]
+[2025-08-28 11:01:55,082][trainers.trainer][INFO] - Rank: 0, Progress: [409/6381]
+[2025-08-28 11:01:55,231][trainers.trainer][INFO] - Rank: 0, Progress: [410/6381]
+[2025-08-28 11:01:55,378][trainers.trainer][INFO] - Rank: 0, Progress: [411/6381]
+[2025-08-28 11:01:55,455][trainers.trainer][INFO] - Rank: 0, Progress: [412/6381]
+[2025-08-28 11:01:55,623][trainers.trainer][INFO] - Rank: 0, Progress: [413/6381]
+[2025-08-28 11:02:52,106][__main__][INFO] - Working directory : /root/workspace/SwiftTailor3
+[2025-08-28 11:02:52,106][__main__][INFO] - Output directory : /root/workspace/SwiftTailor3/output/train
+[2025-08-28 11:02:52,114][data.data_wrappers.data_wrapper][INFO] - Loading data split from assets/data_configs/garmentcodedata_datasplit.json
+[2025-08-28 11:03:16,792][data.data_wrappers.data_wrapper][INFO] - Dataset split: 114869 / 6381 / 6381
+[2025-08-28 11:03:16,801][__main__][INFO] - Dataset created.
+[2025-08-28 11:03:17,090][__main__][INFO] - Added 122 tokens to the tokenizer.
+[2025-08-28 11:03:17,111][__main__][INFO] - Token name to index dictionary: {'<pattern_cmd_MOVE>': 151674, '<pattern_cmd_LINE>': 151675, '<pattern_cmd_CLINE>': 151676, '<pattern_cmd_CURVE>': 151677, '<pattern_cmd_CCURVE>': 151678, '<pattern_cmd_CUBIC>': 151679, '<pattern_cmd_CCUBIC>': 151680, '<pattern_cmd_ARC>': 151681, '<pattern_cmd_CARC>': 151682, '<panel_start>': 151683, '<panel_end>': 151684, '<pattern_start>': 151685, '<pattern_end>': 151686, '<stitch_tag_0>': 151687, '<stitch_tag_1>': 151688, '<stitch_tag_2>': 151689, '<stitch_tag_3>': 151690, '<stitch_tag_4>': 151691, '<stitch_tag_5>': 151692, '<stitch_tag_6>': 151693, '<stitch_tag_7>': 151694, '<stitch_tag_8>': 151695, '<stitch_tag_9>': 151696, '<stitch_tag_10>': 151697, '<stitch_tag_11>': 151698, '<stitch_tag_12>': 151699, '<stitch_tag_13>': 151700, '<stitch_tag_14>': 151701, '<stitch_tag_15>': 151702, '<stitch_tag_16>': 151703, '<stitch_tag_17>': 151704, '<stitch_tag_18>': 151705, '<stitch_tag_19>': 151706, '<stitch_tag_20>': 151707, '<stitch_tag_21>': 151708, '<stitch_tag_22>': 151709, '<stitch_tag_23>': 151710, '<stitch_tag_24>': 151711, '<stitch_tag_25>': 151712, '<stitch_tag_26>': 151713, '<stitch_tag_27>': 151714, '<stitch_tag_28>': 151715, '<stitch_tag_29>': 151716, '<stitch_tag_30>': 151717, '<stitch_tag_31>': 151718, '<stitch_tag_32>': 151719, '<stitch_tag_33>': 151720, '<stitch_tag_34>': 151721, '<stitch_tag_35>': 151722, '<stitch_tag_36>': 151723, '<stitch_tag_37>': 151724, '<stitch_tag_38>': 151725, '<stitch_tag_39>': 151726, '<stitch_tag_40>': 151727, '<stitch_tag_41>': 151728, '<stitch_tag_42>': 151729, '<stitch_tag_43>': 151730, '<stitch_tag_44>': 151731, '<stitch_tag_45>': 151732, '<stitch_tag_46>': 151733, '<stitch_tag_47>': 151734, '<stitch_tag_48>': 151735, '<stitch_tag_49>': 151736, '<stitch_tag_50>': 151737, '<stitch_tag_51>': 151738, '<stitch_tag_52>': 151739, '<stitch_tag_53>': 151740, '<stitch_tag_54>': 151741, '<stitch_tag_55>': 151742, '<stitch_tag_56>': 151743, '<stitch_tag_57>': 151744, '<stitch_tag_58>': 151745, '<stitch_tag_59>': 151746, '<stitch_tag_60>': 151747, '<stitch_tag_61>': 151748, '<stitch_tag_62>': 151749, '<stitch_tag_63>': 151750, '<stitch_tag_64>': 151751, '<stitch_tag_65>': 151752, '<stitch_tag_66>': 151753, '<stitch_tag_67>': 151754, '<stitch_tag_68>': 151755, '<stitch_tag_69>': 151756, '<stitch_tag_70>': 151757, '<stitch_tag_71>': 151758, '<stitch_tag_72>': 151759, '<stitch_tag_73>': 151760, '<stitch_tag_74>': 151761, '<stitch_tag_75>': 151762, '<stitch_tag_76>': 151763, '<stitch_tag_77>': 151764, '<stitch_tag_78>': 151765, '<stitch_tag_79>': 151766, '<stitch_tag_80>': 151767, '<stitch_tag_81>': 151768, '<stitch_tag_82>': 151769, '<stitch_tag_83>': 151770, '<stitch_tag_84>': 151771, '<stitch_tag_85>': 151772, '<stitch_tag_86>': 151773, '<stitch_tag_87>': 151774, '<stitch_tag_88>': 151775, '<stitch_tag_89>': 151776, '<stitch_tag_90>': 151777, '<stitch_tag_91>': 151778, '<stitch_tag_92>': 151779, '<stitch_tag_93>': 151780, '<stitch_tag_94>': 151781, '<stitch_tag_95>': 151782, '<stitch_tag_96>': 151783, '<stitch_tag_97>': 151784, '<stitch_tag_98>': 151785, '<stitch_tag_99>': 151786, '<stitch_tag_100>': 151787, '<stitch_tag_101>': 151788, '<stitch_tag_102>': 151789, '<stitch_tag_103>': 151790, '<stitch_tag_104>': 151791, '<stitch_tag_105>': 151792, '<stitch_tag_106>': 151793, '<stitch_tag_107>': 151794, '<stitch_tag_null>': 151795}
+[2025-08-28 11:03:17,114][models.internvl.configuration_internvl_chat][INFO] - vision_select_layer: -1
+[2025-08-28 11:03:17,114][models.internvl.configuration_internvl_chat][INFO] - ps_version: v2
+[2025-08-28 11:03:17,115][models.internvl.configuration_internvl_chat][INFO] - min_dynamic_patch: 1
+[2025-08-28 11:03:17,115][models.internvl.configuration_internvl_chat][INFO] - max_dynamic_patch: 12
+[2025-08-28 11:03:17,115][__main__][INFO] - Loading model from output/train/ckpt_29999/output_dir
+[2025-08-28 11:03:17,129][models.st_model][INFO] - num_image_token: 256
+[2025-08-28 11:03:17,129][models.st_model][INFO] - ps_version: v2
+[2025-08-28 11:04:11,409][models.internvl.configuration_internvl_chat][INFO] - vision_config is None. Initializing the InternVisionConfig with default values.
+[2025-08-28 11:04:11,410][models.internvl.configuration_internvl_chat][INFO] - llm_config is None. Initializing the LlamaConfig config with default values (`LlamaConfig`).
+[2025-08-28 11:04:11,411][models.internvl.configuration_internvl_chat][INFO] - vision_select_layer: -1
+[2025-08-28 11:04:11,411][models.internvl.configuration_internvl_chat][INFO] - ps_version: v1
+[2025-08-28 11:04:11,411][models.internvl.configuration_internvl_chat][INFO] - min_dynamic_patch: 1
+[2025-08-28 11:04:11,411][models.internvl.configuration_internvl_chat][INFO] - max_dynamic_patch: 6
+[2025-08-28 11:18:45,123][trainers.trainer][INFO] - Rank: 0, Progress: [0/6381]
+[2025-08-28 11:18:45,141][models.internvl.configuration_internvl_chat][INFO] - vision_config is None. Initializing the InternVisionConfig with default values.
+[2025-08-28 11:18:45,141][models.internvl.configuration_internvl_chat][INFO] - llm_config is None. Initializing the LlamaConfig config with default values (`LlamaConfig`).
+[2025-08-28 11:18:45,143][models.internvl.configuration_internvl_chat][INFO] - vision_select_layer: -1