The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
status: string
source_files: int64
relocated_config: string
clean_action_ratio: double
data_windows: int64
changed_window_fields: list<item: string>
child 0, item: string
target_training_launched: bool
local_fixture_uses_read_only_existing_data: bool
transfer_tests: int64
ignored_payload_fields: list<item: string>
child 0, item: string
component_sources: struct<vae: string, text: string>
child 0, vae: string
child 1, text: string
new_initialization: struct<depth_input_proj.weight: string, wrist_mask_proj.weight: string>
child 0, depth_input_proj.weight: string
child 1, wrist_mask_proj.weight: string
optimizer_restored: bool
projection_initialization: string
strict_full_state: bool
initialized: list<item: string>
child 0, item: string
reference_source_commit: string
semantic_limit: string
elapsed_seconds: double
main_layers: int64
new_training_step: int64
shape_conflicts: list<item: null>
child 0, item: null
real_official_checkpoint_loaded: bool
total_parameters: int64
production_training_run: bool
source_format: string
type_conflicts: list<item: null>
child 0, item: null
extra: list<item: null>
child 0, item: null
reused: list<item: string>
child 0, item: string
source_sha256: string
excluded_non_dit: list<item: string>
child 0, item: string
scheduler_restored: bool
depth_head_source: string
checkpoint_training_commit: null
allowed_missing: list<item: string>
child 0, item: string
reset: list<item: string>
child 0, item: string
inference_run: bool
...
3, total: struct<tensors: int64, elements: int64>
child 0, tensors: int64
child 1, elements: int64
missing: list<item: null>
child 0, item: null
vae_text_loaded_in_this_dit_only_audit: bool
validation_device: string
target_layout: struct<rgb: list<item: string>, condition: list<item: string>, depth: list<item: string>, wrist_mask (... 274 chars omitted)
child 0, rgb: list<item: string>
child 0, item: string
child 1, condition: list<item: string>
child 0, item: string
child 2, depth: list<item: string>
child 0, item: string
child 3, wrist_mask: struct<version: string, camera_keys: list<item: string>, source: string, threshold_m: double, backgr (... 171 chars omitted)
child 0, version: string
child 1, camera_keys: list<item: string>
child 0, item: string
child 2, source: string
child 3, threshold_m: double
child 4, background_m: double
child 5, invalid_zero_is_robot: bool
child 6, coordinate_system: string
child 7, resize: string
child 8, pooling: string
child 9, injection: string
child 10, projection: string
child 11, visible_segmentation: bool
velocity_head_initialization: string
dtype_conversions: list<item: struct<name: string, source: string, target: string>>
child 0, item: struct<name: string, source: string, target: string>
child 0, name: string
child 1, source: string
child 2, target: string
new_wrist_projection_parameters: int64
to
{'status': Value('string'), 'source_format': Value('string'), 'reference_source_commit': Value('string'), 'optimizer_restored': Value('bool'), 'scheduler_restored': Value('bool'), 'new_training_step': Value('int64'), 'missing': List(Value('null')), 'extra': List(Value('null')), 'shape_conflicts': List(Value('null')), 'type_conflicts': List(Value('null')), 'reused': List(Value('string')), 'initialized': List(Value('string')), 'reset': List(Value('string')), 'dtype_conversions': List({'name': Value('string'), 'source': Value('string'), 'target': Value('string')}), 'source_path': Value('string'), 'source_sha256': Value('string'), 'expected_sha256': Value('string'), 'target_layout': {'rgb': List(Value('string')), 'condition': List(Value('string')), 'depth': List(Value('string')), 'wrist_mask': {'version': Value('string'), 'camera_keys': List(Value('string')), 'source': Value('string'), 'threshold_m': Value('float64'), 'background_m': Value('float64'), 'invalid_zero_is_robot': Value('bool'), 'coordinate_system': Value('string'), 'resize': Value('string'), 'pooling': Value('string'), 'injection': Value('string'), 'projection': Value('string'), 'visible_segmentation': Value('bool')}}, 'ignored_payload_fields': List(Value('string')), 'allowed_missing': List(Value('string')), 'coverage': {'reused': {'tensors': Value('int64'), 'elements': Value('int64')}, 'new': {'tensors': Value('int64'), 'elements': Value('int64')}, 'reset': {'tensors': Value('int64'), 'elements': Value('int64')}, 'total': {'tensors': Value('int64'), 'elements': Value('int64')}}, 'new_initialization': {'depth_input_proj.weight': Value('string'), 'wrist_mask_proj.weight': Value('string')}, 'excluded_non_dit': List(Value('string')), 'projection_initialization': Value('string'), 'velocity_head_initialization': Value('string'), 'depth_head_source': Value('string'), 'checkpoint_training_commit': Value('null'), 'semantic_limit': Value('string'), 'validation_device': Value('string'), 'real_official_checkpoint_loaded': Value('bool'), 'strict_full_state': Value('bool'), 'total_parameters': Value('int64'), 'new_wrist_projection_parameters': Value('int64'), 'main_layers': Value('int64'), 'depth_layers': Value('int64'), 'elapsed_seconds': Value('float64'), 'inference_run': Value('bool'), 'production_training_run': Value('bool'), 'vae_text_loaded_in_this_dit_only_audit': Value('bool'), 'component_sources': {'vae': Value('string'), 'text': Value('string')}}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
status: string
source_files: int64
relocated_config: string
clean_action_ratio: double
data_windows: int64
changed_window_fields: list<item: string>
child 0, item: string
target_training_launched: bool
local_fixture_uses_read_only_existing_data: bool
transfer_tests: int64
ignored_payload_fields: list<item: string>
child 0, item: string
component_sources: struct<vae: string, text: string>
child 0, vae: string
child 1, text: string
new_initialization: struct<depth_input_proj.weight: string, wrist_mask_proj.weight: string>
child 0, depth_input_proj.weight: string
child 1, wrist_mask_proj.weight: string
optimizer_restored: bool
projection_initialization: string
strict_full_state: bool
initialized: list<item: string>
child 0, item: string
reference_source_commit: string
semantic_limit: string
elapsed_seconds: double
main_layers: int64
new_training_step: int64
shape_conflicts: list<item: null>
child 0, item: null
real_official_checkpoint_loaded: bool
total_parameters: int64
production_training_run: bool
source_format: string
type_conflicts: list<item: null>
child 0, item: null
extra: list<item: null>
child 0, item: null
reused: list<item: string>
child 0, item: string
source_sha256: string
excluded_non_dit: list<item: string>
child 0, item: string
scheduler_restored: bool
depth_head_source: string
checkpoint_training_commit: null
allowed_missing: list<item: string>
child 0, item: string
reset: list<item: string>
child 0, item: string
inference_run: bool
...
3, total: struct<tensors: int64, elements: int64>
child 0, tensors: int64
child 1, elements: int64
missing: list<item: null>
child 0, item: null
vae_text_loaded_in_this_dit_only_audit: bool
validation_device: string
target_layout: struct<rgb: list<item: string>, condition: list<item: string>, depth: list<item: string>, wrist_mask (... 274 chars omitted)
child 0, rgb: list<item: string>
child 0, item: string
child 1, condition: list<item: string>
child 0, item: string
child 2, depth: list<item: string>
child 0, item: string
child 3, wrist_mask: struct<version: string, camera_keys: list<item: string>, source: string, threshold_m: double, backgr (... 171 chars omitted)
child 0, version: string
child 1, camera_keys: list<item: string>
child 0, item: string
child 2, source: string
child 3, threshold_m: double
child 4, background_m: double
child 5, invalid_zero_is_robot: bool
child 6, coordinate_system: string
child 7, resize: string
child 8, pooling: string
child 9, injection: string
child 10, projection: string
child 11, visible_segmentation: bool
velocity_head_initialization: string
dtype_conversions: list<item: struct<name: string, source: string, target: string>>
child 0, item: struct<name: string, source: string, target: string>
child 0, name: string
child 1, source: string
child 2, target: string
new_wrist_projection_parameters: int64
to
{'status': Value('string'), 'source_format': Value('string'), 'reference_source_commit': Value('string'), 'optimizer_restored': Value('bool'), 'scheduler_restored': Value('bool'), 'new_training_step': Value('int64'), 'missing': List(Value('null')), 'extra': List(Value('null')), 'shape_conflicts': List(Value('null')), 'type_conflicts': List(Value('null')), 'reused': List(Value('string')), 'initialized': List(Value('string')), 'reset': List(Value('string')), 'dtype_conversions': List({'name': Value('string'), 'source': Value('string'), 'target': Value('string')}), 'source_path': Value('string'), 'source_sha256': Value('string'), 'expected_sha256': Value('string'), 'target_layout': {'rgb': List(Value('string')), 'condition': List(Value('string')), 'depth': List(Value('string')), 'wrist_mask': {'version': Value('string'), 'camera_keys': List(Value('string')), 'source': Value('string'), 'threshold_m': Value('float64'), 'background_m': Value('float64'), 'invalid_zero_is_robot': Value('bool'), 'coordinate_system': Value('string'), 'resize': Value('string'), 'pooling': Value('string'), 'injection': Value('string'), 'projection': Value('string'), 'visible_segmentation': Value('bool')}}, 'ignored_payload_fields': List(Value('string')), 'allowed_missing': List(Value('string')), 'coverage': {'reused': {'tensors': Value('int64'), 'elements': Value('int64')}, 'new': {'tensors': Value('int64'), 'elements': Value('int64')}, 'reset': {'tensors': Value('int64'), 'elements': Value('int64')}, 'total': {'tensors': Value('int64'), 'elements': Value('int64')}}, 'new_initialization': {'depth_input_proj.weight': Value('string'), 'wrist_mask_proj.weight': Value('string')}, 'excluded_non_dit': List(Value('string')), 'projection_initialization': Value('string'), 'velocity_head_initialization': Value('string'), 'depth_head_source': Value('string'), 'checkpoint_training_commit': Value('null'), 'semantic_limit': Value('string'), 'validation_device': Value('string'), 'real_official_checkpoint_loaded': Value('bool'), 'strict_full_state': Value('bool'), 'total_parameters': Value('int64'), 'new_wrist_projection_parameters': Value('int64'), 'main_layers': Value('int64'), 'depth_layers': Value('int64'), 'elapsed_seconds': Value('float64'), 'inference_run': Value('bool'), 'production_training_run': Value('bool'), 'vae_text_loaded_in_this_dit_only_audit': Value('bool'), 'component_sources': {'vae': Value('string'), 'text': Value('string')}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
DepthWAM V2:接收方部署、训练与 RoboTwin 评测说明
交付日期:2026-10-13。本文供协作者在另一台服务器使用已交付的代码、数据和权重。两台服务器不需要互通。
本文命令是目标服务器操作说明,没有在目标服务器执行过。源端文件上传、校验已完成;不代表接收方环境、正式训练或闭环评测已验证。
如果由接收方 AI 负责部署和代跑,请先把迁移目录内 ai_handoff_v1/START_AI_PROMPT.md 的提示词交给它,并指定本机下载/部署目录、4 张已分配 GPU。SwanLab 目标已固定为项目所有者的 mambaou/depthwam,不使用代跑者的个人项目。AI 应先读 ai_handoff_v1/AI_HANDOFF.md 和 experiment_contract.json,再按本文命令执行。机器信息可填入 receiver_settings.example.json 的新副本。任务书已明确连续执行范围、验收、失败处理、训练后评测和回传要求;它不是可以直接启动 GPU 的自动化程序。
给接收方 AI 的最新补充(可直接复制):
SwanLab 登录已由我委托协作者提前处理,请复用现有凭据,不再安排登录或索要 API Key。用同一 Linux 用户按迁移包的 README 和 AI_HANDOFF.md 创建训练、评测环境;若人工登录使用自定义 SWANLAB_ROOT,沿用实际目录并写入注册、训练、上传器和 tmux 启动脚本;若使用项目级凭据,保持正确的工作目录。通过现有 SDK/检查脚本核验账号为 mambaou、目标为 mambaou/depthwam,为 V2 新建实验,不追加旧曲线。不要读取或输出密钥明文。凭据不可见时报告具体问题并继续环境准备,不自行换账号或重新登录。ModelScope 继续使用接收方已获下载权限的账号。完整任务见 ai_handoff_v1/START_AI_PROMPT.md。
材料完整性提醒:原 download_transfer.py 不会自动下载后来补充的完整源码或 AI 文档;从网页/SDK 拉取迁移目录时须包括 source_complete_v1/ 与 ai_handoff_v1/。交接检查命令为 python <迁移目录>/ai_handoff_v1/check_handoff.py --transfer-root <迁移目录>;其默认结果仅证明文件齐全及小文件哈希正确,不代表大数据包已全部哈希验证。完整校验仍由第 3 节工具或 --verify-bulk 完成。
1. 交付范围和实验约定
| 项目 | 本次约定 |
|---|---|
| 模型 | mixed_depth_flow_fork,主干 R/C/A/P,保留 RGB 中间分叉和后 10 层配对深度分支 |
| V2 编码 | body_piecewise_10_80_10_v1,仅 cam_high 本体伪彩采用新映射 |
| V2 推理日程 | depth_progress_spread_v1,RGB/深度各 50 步,动作/未来状态各 10 步,H=5 |
| 训练时间策略 | schedule_matched_v1,与实际推理事件/采样器时钟匹配 |
| 数据 | click_bell、pick_dual_bottles、lift_pot,均含 clean/randomized |
| 动作/状态 | 14 维相对 EE 动作、16 维 EE 状态,action_num=4,预测/默认执行 32 动作 |
| 初始化 | 官方 X-WAM weights-only 初始化,resume=null |
| 训练 | 4 GPU × 每卡 batch=4,梯度累积 1,全局 batch=16,20,000 次优化器更新 |
| 优化器/LR | DeepSpeed ZeRO-2 / FusedAdam,AdamW 语义,LR=1e-5,betas=(0.9,0.999),eps=1e-8,WD=0.01;200 步 warmup 后 constant |
| 其他 | bf16,冻结原 VAE,均匀窗口采样,不启用接触采样,保留 SwanLab |
| 保存 | 每 2,000 步保存权重及完整状态,共 10 组;保留 20k 最终组 |
| 初筛 | 8k/10k/12k/14k,每任务 clean 10 + randomized 10,每 checkpoint 60 回合 |
| 完整评测 | 每个入选 checkpoint:3 任务 × 2 场景 × 100 = 600 回合 |
本次 V2 预算按最新要求调整为 20k。已封存代码包中的原模板仍是 14k;本文生成的 _local 配置和正式启动命令均显式覆盖到 20k。原服务器 V1 续训属于独立实验。本轮仅调整 V2 训练预算,初筛 checkpoint 仍为 8k/10k/12k/14k;16k/18k/20k 权重正常保存,暂不自动纳入筛选。
混合监督保持 lambda_depth_prop=0、lambda_depth_mixed=1;不能退回 body-only x0 监督。保留当前干净 body-depth 条件 C。推理不读取完整场景深度真值。
默认动作更新在第 6、11、16、21、26、31、36、41、46、51 次前向。最后一次读取深度已完成的输入。共 51 次完整前向;相对旧完整 50 轮日程多 1 次前向,不代表已测得性能收益。
2. 接收方先准备什么
- Linux、Bash、Git、tmux、FFmpeg/ffprobe、编译工具和与目标 GPU 兼容的 NVIDIA 驱动/CUDA Toolkit。PyTorch 自带 CUDA runtime 不等于已经有编译扩展需要的
nvcc。 - 可分配的 4 张 GPU。源端使用 A100 80GB;其他型号/显存能否维持 batch=4,须由目标服务器冒烟确认。源端的 GPU 编号不能照搬。
- 足够 CPU、主机内存和本地磁盘。下载包约 105.72 GiB,解压再需约同等容量,公共模型约 12 GiB。保留 10 组完整训练状态和运行余量,部署规划按约 2 TiB 可用空间准备;这是预留估算,不是精确最低值。源端 V1 的单组完整状态约 90 GiB,另有约 12.65 GiB 的权重文件。
- ModelScope 该申请制数据集的下载权限;SwanLab 项目所有者
mambaou的授权登录,上传到既有项目depthwam。ModelScope 仍使用接收方自己获批的账号;两项登录不要混淆。完整项目源码已单独打包,按本文默认流程部署不需要私有 GitHub 仓库权限。Git 工具仍用于安装公共依赖;公共组件下载仍需网络或另行离线交付。
仓库当前采用申请制下载(2026-10-13 核实)。接收方先登录自己的 ModelScope 账号,打开 https://modelscope.cn/datasets/linyinyu/depthwam-v2-transfer 提交下载申请,并把账号名告诉仓库所有者。所有者在仓库「设置 → 申请信息」中处理对应申请;批准后,接收方在目标服务器使用自己的 Token 登录,再执行下面的下载命令。只有需要另行访问私有 GitHub 或后续协作时,才需要另外配置 GitHub 权限。
不需要公开仓库或共用源端 ModelScope 账号。不要把密码、ModelScope Token、SwanLab Key 发到 AI 对话或写入 README。SwanLab 授权由所有者与协作者在文档之外处理,本包不包含凭据。仓库所有者登录成功不等于接收方账户已有读取权限。若无法给接收方提供适当的仓库读取权限,由所有者下载完整迁移目录后,通过双方认可的方式交付,使用下面的离线校验路径。
以下命令在 Bash 终端运行,不要粘贴到 Python 的 >>> 提示符中。如果在 >>> 中,先输入 exit()。
3. 下载、校验和解压
仓库:https://modelscope.cn/datasets/linyinyu/depthwam-v2-transfer
目录前缀:v2_migration_20261013/
原始迁移包上传完成时间:2026-10-13 16:57(北京时间)。共 81 个文件,113,511,008,414 bytes,含最后上传的 TRANSFER_COMPLETE.json。这一数量仅指已封存迁移包,不包含后来新增的 README 等说明文件。原始文件及其 manifest 保持不变。
完整操作说明同时放在仓库根目录 README.md 和 v2_migration_20261013/README.md;CPU 检查记录在 v2_migration_20261013/README.validation.json。完整源码补充包位于 v2_migration_20261013/source_complete_v1/。旧下载工具仍按封存 manifest 获取原始数据包,不会自动下载这些补充文件;源码按第 4 节另行下载,本文也需单独保存。
3.1 设置目标目录
先把 /YOUR/ABSOLUTE/PATH/depthwam-v2 改为接收方可写的大容量目录。路径建议不含空格。保存这些变量,后续各终端都需重新载入。
export V2_DEPLOY=/YOUR/ABSOLUTE/PATH/depthwam-v2
export V2_PROJECT="$V2_DEPLOY/project"
export V2_UNPACK="$V2_DEPLOY/unpacked"
export V2_TRANSFER="$V2_DEPLOY/download/v2_migration_20261013"
export V2_MODELS="$V2_UNPACK/checkpoints"
mkdir -p "$V2_DEPLOY"
首次下载可使用独立 Python 3.10 环境,安装源端使用的 SDK:
conda create -n depthwam-transfer python=3.10 pip -y
conda activate depthwam-transfer
python -m pip install modelscope==1.34.0
3.2 在线下载:只在接收方已有读取权限时使用
交互登录,不把 Token 放进命令参数或命令历史:
python -c 'import getpass; from modelscope.hub.api import HubApi; HubApi().login(getpass.getpass("ModelScope token (hidden): "))'
如果采用旧下载脚本,先同时获取后来新增的完整源码和 AI 交接文件。下面只下载约 1 MiB 的补充材料,不下载大分包、不启动训练;已通过整个目录下载取得这些文件时可跳过:
python - <<'PY'
import os
from pathlib import Path
from modelscope.hub.file_download import dataset_file_download
root=Path(os.environ['V2_DEPLOY'])
files=['README.md','README.validation.json',
'source_complete_v1/depthwam_v2_source.tar.gz','source_complete_v1/source_bundle.json',
'source_complete_v1/SHA256SUMS','source_complete_v1/validation.json',
'ai_handoff_v1/START_AI_PROMPT.md','ai_handoff_v1/AI_HANDOFF.md',
'ai_handoff_v1/experiment_contract.json','ai_handoff_v1/receiver_settings.example.json',
'ai_handoff_v1/check_handoff.py','ai_handoff_v1/check_swanlab_target.py',
'ai_handoff_v1/validation.json','ai_handoff_v1/HANDOFF_MANIFEST.json']
for name in files:
dataset_file_download(dataset_id='linyinyu/depthwam-v2-transfer',
file_path='v2_migration_20261013/'+name,revision='master',local_dir=str(root/'download'))
print('SOURCE_AND_AI_HANDOFF_DOWNLOADED; verify hashes before using the source')
PY
先下载引导脚本,并核对固定 SHA256:
python - <<'PY'
import hashlib, os, shutil
from pathlib import Path
from modelscope.hub.file_download import dataset_file_download
root = Path(os.environ['V2_DEPLOY'])
p = Path(dataset_file_download(
dataset_id='linyinyu/depthwam-v2-transfer',
file_path='v2_migration_20261013/download_transfer.py',
revision='master', local_dir=str(root/'download')))
assert hashlib.sha256(p.read_bytes()).hexdigest() == '7dc7e571148da85ee1839405532b760498672def7ac80f28025408b1509c6c4f'
shutil.copy2(p, root/'download_transfer.py')
print(root/'download_transfer.py')
PY
在 tmux 里执行下载和解压,断开 SSH 不影响它:
tmux new-session -s depthwam-v2-download
python "$V2_DEPLOY/download_transfer.py" \
--download-to "$V2_DEPLOY/download" \
--extract-to "$V2_UNPACK" > "$V2_DEPLOY/download.log" 2>&1
按 Ctrl-b 后按 d 离开 tmux。查看:tail -f "$V2_DEPLOY/download.log"。
脚本会检查私有仓库、完成标记、逐文件 SHA256/大小、各分包 SHA256、整组 TAR 哈希和成员数量。--extract-to 必须是不存在的新目录;失败目录保留,重试解压换一个新目录,不能覆盖已有项目。下载可复用已下载文件,但校验仍会重做。
3.3 离线交付:接收方拿到了完整目录时使用
收到的 v2_migration_20261013 目录放到 $V2_TRANSFER。不要只拿 *.part*;manifest、小代码包和下载脚本都需要。
下面只做本地校验/解压,不登录、不访问 ModelScope:
python - <<'PY'
import hashlib, importlib.util, json, os, tarfile
from pathlib import Path
root = Path(os.environ['V2_TRANSFER'])
dest = Path(os.environ['V2_UNPACK'])
script = root/'download_transfer.py'
assert hashlib.sha256(script.read_bytes()).hexdigest() == '7dc7e571148da85ee1839405532b760498672def7ac80f28025408b1509c6c4f'
spec = importlib.util.spec_from_file_location('transfer', script)
m = importlib.util.module_from_spec(spec); spec.loader.exec_module(m)
receipt = json.loads((root/'TRANSFER_COMPLETE.json').read_text())
assert receipt['state'] == 'TRANSFER_COMPLETE'
assert m.digest(root/'transfer_manifest.json') == receipt['manifest_sha256']
manifest = json.loads((root/'transfer_manifest.json').read_text())
assert (manifest['repo_id'], manifest['prefix']) == ('linyinyu/depthwam-v2-transfer', 'v2_migration_20261013')
for row in manifest['files']:
m.check_relative(row['file'])
p = root/row['file']
assert p.stat().st_size == row['bytes'] and m.digest(p) == row['sha256'], row['file']
dest.mkdir(parents=True, exist_ok=False)
for group in manifest['groups']:
m.extract_group(root, group, dest)
for name, subdir in [('v2_code_overlay.tar.gz','code_overlay'),
('validated_v2_source_reference.tar.gz','validated_source_reference'),
('robotwin_eval_code_and_cases.tar.gz','evaluation')]:
with tarfile.open(root/'prepared'/name, 'r:gz') as archive:
m.extract_files(archive, dest/subdir)
print('OFFLINE_VERIFY_AND_EXTRACT_OK', dest)
PY
各 *.tar.partNNNN 是同一 TAR 的连续分片,不能逐个当成独立 TAR 解压。上述脚本流式拼接读取,不额外生成一份 106 GiB 的拼接 TAR。
3.4 解压后应有的目录
unpacked/
workspace/data/robotwin_three_ee_clock15_v1/ # dataset、stats、provenance
workspace/data/subsets/ # 冻结窗口清单
workspace/data/text_embeds_cache/robotwin/ # 8,879 个所需文本缓存
checkpoints/X-WAM-official-bb6fd164/ # 官方预训练权重
robotwin_assets/ # 物体、纹理、机器人等大资产
code_overlay/ # V2 当前修改 + v2_transfer_provenance.json
validated_source_reference/ # 历史验证源码,仅供对照
evaluation/RoboTwin_eval_private/ # 当前冻结评测后端代码及机器人配置
evaluation/preparation/ # 冻结案例、原协议、训练 seed 证据
迁移包不含 V1 完整训练状态、LIBERO、原始六任务采集全集、编译缓存或凭据。prepared/preparation.json 中的 PREPARED_NOT_TRANSFERRED 是打包前记录;传输是否完成以 TRANSFER_COMPLETE.json 和校验为准。
4. 恢复完整源码(默认不访问私有 GitHub)
完整包已经合成基准提交 c99da64cc438493e185e575c57eecbec55a3ee03 的文件和原封存的 36 个 V2 修改/新增文件;不再要求 clone 或二次应用 overlay。它不是打包时随意复制的当前脏工作树,也不包含 Git 历史、凭据、数据、权重、日志或编译产物。
包名:source_complete_v1/depthwam_v2_source.tar.gz,752,376 bytes(约 735 KiB),328 个文件。SOURCE_SNAPSHOT.json 保存基准、overlay 身份和逐文件校验。原始任务模板仍继承 14k,第 6 节生成的 _local 配置显式设置为 20k。
4.1 下载完整源码补充包
使用第 3 节已登录的 ModelScope 账号,下载源码包及其清单并核对固定哈希:
python - <<'PY'
import hashlib, os
from pathlib import Path
from modelscope.hub.file_download import dataset_file_download
root=Path(os.environ['V2_DEPLOY'])
files={
'depthwam_v2_source.tar.gz':'adff12cc089e4e95e31111b0fd0b618bcd49cd9169f46483c73bebbc4db3ccfc',
'source_bundle.json':'a2410a028a3ade330cf4a31827498f2a4aee8d3fd5456be9e68c57d72083735f',
}
for name,sha in files.items():
p=Path(dataset_file_download(dataset_id='linyinyu/depthwam-v2-transfer',
file_path='v2_migration_20261013/source_complete_v1/'+name,
revision='master',local_dir=str(root/'download')))
assert hashlib.sha256(p.read_bytes()).hexdigest()==sha,name
print('FULL_SOURCE_DOWNLOAD_VERIFIED')
PY
离线交付时,把这两个补充文件放入 $V2_TRANSFER/source_complete_v1/,再执行下面的同一套本地校验。旧下载脚本不会自动获取补充包。
4.2 校验、解压并绑定数据
$V2_PROJECT 必须是不存在的新目录;已经按旧 GitHub 流程部署过的接收方不重复解压、不覆盖原项目。数据包须先按第 3 节解压到 $V2_UNPACK。
python - <<'PY'
import hashlib, json, os, shutil, tarfile
from pathlib import Path, PurePosixPath
project=Path(os.environ['V2_PROJECT'])
source=Path(os.environ['V2_TRANSFER'])/'source_complete_v1'
bundle_path=source/'source_bundle.json'
assert hashlib.sha256(bundle_path.read_bytes()).hexdigest()=='a2410a028a3ade330cf4a31827498f2a4aee8d3fd5456be9e68c57d72083735f'
bundle=json.loads(bundle_path.read_text())
assert bundle['archive']=='depthwam_v2_source.tar.gz'
archive_path=source/bundle['archive']
assert archive_path.stat().st_size==bundle['bytes']
assert hashlib.sha256(archive_path.read_bytes()).hexdigest()==bundle['sha256']
project.mkdir(parents=True,exist_ok=False)
with tarfile.open(archive_path,'r:gz') as archive:
members=archive.getmembers(); seen=set()
for member in members:
rel=PurePosixPath(member.name)
assert member.isfile() and not rel.is_absolute() and '..' not in rel.parts
assert '.git' not in rel.parts and member.name not in seen
seen.add(member.name)
for member in members:
dst=project/member.name; dst.parent.mkdir(parents=True,exist_ok=True)
with archive.extractfile(member) as src,dst.open('xb') as out:
shutil.copyfileobj(src,out)
dst.chmod(member.mode & 0o777)
manifest_path=project/'SOURCE_SNAPSHOT.json'
assert hashlib.sha256(manifest_path.read_bytes()).hexdigest()==bundle['source_manifest_sha256']
manifest=json.loads(manifest_path.read_text())
assert manifest['base_commit']==bundle['base_commit']
assert manifest['overlay_sha256']==bundle['overlay_sha256']
assert seen==set(manifest['files'])|{'SOURCE_SNAPSHOT.json'}
assert len(seen)==bundle['source_files']
for rel,record in manifest['files'].items():
p=project/rel
assert p.stat().st_size==record['bytes'] and hashlib.sha256(p.read_bytes()).hexdigest()==record['sha256'],rel
assert not (project/'.git').exists()
assert not (project/'data').exists(), '项目 data 已存在,先人工核对,不能覆盖'
(project/'data').symlink_to(Path(os.environ['V2_UNPACK'])/'workspace/data', target_is_directory=True)
report_rel=Path('runs/robotwin_three_ee_prepare_20261005/full_check/report.json')
report_dst=project/report_rel
assert not report_dst.exists()
report_dst.parent.mkdir(parents=True,exist_ok=True)
shutil.copy2(Path(os.environ['V2_UNPACK'])/'workspace'/report_rel,report_dst)
print('FULL_SOURCE_AND_DATA_BINDING_OK',len(seen))
PY
解压项目没有 .git,版本证据使用 SOURCE_SNAPSHOT.json、v2_transfer_provenance.json 和压缩包 SHA256,不伪造 Git commit。当前训练入口使用已提供的 EE 统计量,不调用需要 Git 仓库的重新统计缓存分支。若以后启用该分支,需要单独处理其 Git 依赖。
不要再应用旧 code_overlay 或 validated_source_reference 覆盖此代码;前者已合入,后者是较早的验证快照。不要复制源端 worktree 的 .git 文件。不要启动包内原服务器专用的 run_reboot_training.py、run_mixed_v2_after_training.py 等调度器,它们含旧路径/旧任务安排。
5. 安装训练环境和公共模型
5.1 训练环境
源端实际版本:Python 3.10.20、PyTorch 2.7.1+cu128、torchvision 0.22.1+cu128、torchcodec 0.5+cu128、DeepSpeed 0.18.5、SwanLab 0.8.2。完整清单在 train_installed_versions.json。
下面是新建环境的安装入口。驱动、CUDA Toolkit、编译器和 wheel 可用性须在目标服务器确认;不要把“有版本清单”当成环境已复现。
conda create -n depthwam-v2-train python=3.10 pip -y
conda activate depthwam-v2-train
python -m pip install torch==2.7.1 torchvision==0.22.1 \
--index-url https://download.pytorch.org/whl/cu128
python -m pip install torchcodec==0.5 --index-url https://download.pytorch.org/whl/cu128
cd "$V2_PROJECT"
python -m pip install -e . --extra-index-url https://download.pytorch.org/whl/cu128
python -m pip install pytest matplotlib scipy==1.15.3 opencv-python==4.11.0.86
python -m pip check
export V2_TRAIN_PY="$(command -v python)"
如果 DeepSpeed 构建失败,先核对本机 nvcc、编译器及 CUDA 路径;不要把其他机器的 .so 或 torch extension cache 复制过来。正常训练使用 FusedAdam,不能为绕过安装问题静默换成另一优化器。
5.2 公共模型:只下载所需文件
已交付的官方 X-WAM 文件:
$V2_MODELS/X-WAM-official-bb6fd164/pretrained/checkpoints/last.ckpt/checkpoint/mp_rank_00_model_states.pt
SHA256: e31a909cdd31316386c582f9ff22c9cea5d34891edbd9c7fe75bdd5d7e737880
来源: sharinka0715/X-WAM-checkpoints
revision: bb6fd1643cfa8bdc751612a7ace0bd8062a8917a
还需 VAE(约 1.31 GiB)、T5(约 10.58 GiB)、tokenizer。训练使用文本缓存,评测新指令仍需要 T5/tokenizer。无需下载整个 Wan DiT 分片,官方初始化路径会跳过它。
python - <<'PY'
import os
from pathlib import Path
from modelscope import snapshot_download
base = Path(os.environ['V2_MODELS'])
repo = 'DiffSynth-Studio/Wan-Series-Converted-Safetensors'
snapshot_download(repo, local_dir=str(base/repo), allow_file_pattern=[
'Wan2.2_VAE.safetensors', 'models_t5_umt5-xxl-enc-bf16.safetensors'])
repo = 'Wan-AI/Wan2.1-T2V-1.3B'
snapshot_download(repo, local_dir=str(base/repo), allow_file_pattern='google/umt5-xxl/*')
print('PUBLIC_MODELS_DOWNLOADED', base)
PY
下载后记录文件 SHA256 和实际 revision,纳入接收方运行记录。公共仓库未在本迁移包中冻结 revision;如果下载内容与项目加载器的模型校验不兼容,停止并核对来源,不绕过校验。离线目标服务器需由有网络的一端先下载这三个组件后再交付。
6. 生成目标机器配置:只改路径,不改数据窗口
原窗口清单绑定源端绝对路径。下面创建新清单、新 task 和新 sim 配置;原清单、stats、provenance、视频及 EE parquet 均不改写。
cd "$V2_PROJECT"
python - <<'PY'
import copy, hashlib, json, os
from pathlib import Path
from omegaconf import OmegaConf
root = Path(os.environ['V2_PROJECT']).resolve()
models = Path(os.environ['V2_MODELS']).resolve()
local = root/'deployment_local'; local.mkdir(exist_ok=False)
original = root/'data/subsets/robotwin_three_mixed_ee_nohit_windows_v2.json'
raw = original.read_bytes()
assert hashlib.sha256(raw).hexdigest() == 'ce7717a838da27a40427a4f5fc7244da9299b2bb11087e1e3e8dbd8e0fbbd97b'
old = json.loads(raw); new = copy.deepcopy(old)
data = (root/'data/robotwin_three_ee_clock15_v1').resolve()
new.update(dataset_root=str(data/'dataset'), source_provenance=str(data/'provenance.json'),
source_stats=str(data/'stats.json'))
assert {k:v for k,v in new.items() if k not in ('dataset_root','source_provenance','source_stats')} == {
k:v for k,v in old.items() if k not in ('dataset_root','source_provenance','source_stats')}
for p, digest in [(data/'provenance.json', new['source_provenance_sha256']),
(data/'stats.json', new['source_stats_sha256']),
(data/'dataset/meta/info.json', new['dataset_info_sha256'])]:
assert hashlib.sha256(p.read_bytes()).hexdigest() == digest, p
manifest = local/'windows.json'
manifest.write_text(json.dumps(new, ensure_ascii=False, indent=2)+'\n')
sha = hashlib.sha256(manifest.read_bytes()).hexdigest()
xwam = models/'X-WAM-official-bb6fd164/pretrained/checkpoints/last.ckpt/checkpoint/mp_rank_00_model_states.pt'
assert xwam.is_file()
config = dict(defaults=['robotwin_three_mixed_ee_v2','_self_'], resume=None,
max_steps=20000, lr_scheduler_total_steps=20000, save_every=2000,
data=dict(train=dict(dataset_dirs=[str(data/'dataset')],
pretrained_norm_stats=str(data/'stats.json'), window_manifest=str(manifest),
window_manifest_sha256=sha,
text_embedding_cache_dir=str((root/'data/text_embeds_cache/robotwin').resolve()))),
model=dict(xwam_pretrained_checkpoint=str(xwam),
robotwin_data_interface=dict(window_manifest_sha256=sha)))
def save_task(name, value):
p = root/'configs/task'/name
with p.open('x') as f:
f.write('# @package _global_\n'+OmegaConf.to_yaml(OmegaConf.create(value)))
save_task('robotwin_three_mixed_ee_v2_local.yaml', config)
save_task('robotwin_three_mixed_ee_v2_local_eval.yaml', dict(
defaults=['robotwin_three_mixed_ee_v2_local','_self_'],
model=dict(xwam_pretrained_checkpoint=None, skip_dit_load_from_pretrain=True,
load_text_encoder=True)))
sim = dict(defaults=['sim_robotwin_three_mixed_ee_v2',
{'override /task':'robotwin_three_mixed_ee_v2_local_eval'},'_self_'],
EVALUATION=dict(sim_cfg_path=str(root/'configs/sim_robotwin_three_mixed_ee_v2_local.yaml'),
dataset_stats_path=str(data/'stats.json')))
with (root/'configs/sim_robotwin_three_mixed_ee_v2_local.yaml').open('x') as f:
f.write('# @package _global_\n'+OmegaConf.to_yaml(OmegaConf.create(sim)))
(local/'path_rebinding.json').write_text(json.dumps(dict(
original_manifest_sha256=hashlib.sha256(raw).hexdigest(), local_manifest_sha256=sha,
windows_unchanged=True, stats_unchanged=True, provenance_unchanged=True,
changed_fields=['dataset_root','source_provenance','source_stats'],
windows=sum(len(e['starts']) for e in new['episodes'])), indent=2)+'\n')
print('LOCAL_CONFIG_READY', sha)
PY
生成后训练与评测必须都使用 _local 入口。若此前已按 14k 版说明生成过本地配置,只在接收方的 configs/task/robotwin_three_mixed_ee_v2_local.yaml 中显式设置 max_steps: 20000、lr_scheduler_total_steps: 20000、save_every: 2000,并重新生成尚未启动的新实验运行配置;不重复执行会覆盖原目录的初始化步骤。新 V2 checkpoint 会记录新的路径绑定清单哈希;不能只改训练配置、评测却仍使用原哈希。不能修改旧 checkpoint 的接口标记来“通过”检查。
数据应保持以下数量:
| 任务/场景 | 交付轨迹 | 有窗口轨迹 | 有效窗口 |
|---|---|---|---|
| click_bell / clean | 50 | 47 | 1,633 |
| click_bell / randomized | 468 | 458 | 15,716 |
| pick_dual_bottles / clean | 50 | 50 | 3,112 |
| pick_dual_bottles / randomized | 467 | 467 | 29,341 |
| lift_pot / clean | 50 | 50 | 2,894 |
| lift_pot / randomized | 470 | 470 | 28,759 |
| 合计 | 1,555 | 1,542 | 81,455 |
窗口含 33 个同步观测、32 个动作,图像取 [0,4,8,...,32] 共 9 帧;原时间间隔筛选为 0.03–0.09 秒,保留 physical_time_s/simulation_step/clock_error_s/source_frame_index,没有独立重采样动作。32 次环境动作不能表述为恒定物理时长。
深度语义保持:body 原裁剪 [0,4.095]m,背景 4.095m,body mask 阈值 3.5m;scene 灰度 [0.1,4.1]m,已验证 no-hit 零值按 4.1m 处理。可见本体区域使用 scene depth 伪彩,不能用 body-only 轮廓覆盖遮挡物。左右腕保持旧编码。
7. CPU 检查和目标服务器短程训练验证
统一运行环境变量(以后启动 tmux 也需继承):
export PYTHONPATH="$V2_PROJECT/src:$V2_PROJECT"
export DIFFSYNTH_MODEL_BASE_PATH="$V2_MODELS"
export DIFFSYNTH_SKIP_DOWNLOAD=true
export HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1
export HF_DATASETS_CACHE="$V2_DEPLOY/hf_datasets_cache"
export OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export PYTHONUNBUFFERED=1
CPU 单元测试不需要 GPU:
cd "$V2_PROJECT"
CUDA_VISIBLE_DEVICES='' OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 \
"$V2_TRAIN_PY" -m pytest -q tests/test_mixed_depth_v2.py \
tests/test_mixed_depth_flow.py
CUDA_VISIBLE_DEVICES='' OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 \
"$V2_TRAIN_PY" -m pytest -q tests/test_robotwin_mixed_ee.py \
-k 'reject_before_mutation or synthetic_migration or factory_never_overwrites or ee32_mixed_forward or mock_ee_execution'
第二条只选择可移植的小模型/接口测试。该文件的其他测试依赖旧窗口清单或源端绝对路径下的官方源码;本迁移包未包含这些测试夹具,不能把它们在目标服务器缺文件当成新模型失败,也不能声称全部测试已通过。目标 _local 配置由下一条专门验证。
已有 check_mixed_depth_v2.py CLI 写死原 task,不支持 --task 参数。迁移后用下面方式调用其真实数据/小模型检查函数,并显式传 _local 配置:
CUDA_VISIBLE_DEVICES='' OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 \
"$V2_TRAIN_PY" - <<'PY'
import json, os
from pathlib import Path
import torch
from omegaconf import OmegaConf
from scripts.check_mixed_depth_v2 import compose, plots_and_trace, real_data, tiny_model_smoke
from fastwam.robotwin_mixed_ee import validate_ee_run_config
root = Path(os.environ['V2_PROJECT']); os.chdir(root); torch.set_num_threads(1)
out = root/'deployment_local/cpu_check'; out.mkdir(exist_ok=False)
os.environ['HF_DATASETS_CACHE'] = str(out/'hf_cache')
cfg = compose('robotwin_three_mixed_ee_v2_local')
validate_ee_run_config(cfg)
assert (cfg.batch_size,cfg.gradient_accumulation_steps,cfg.max_steps,cfg.save_every)==(4,1,20000,2000)
assert cfg.lr_scheduler_total_steps==20000
assert cfg.model.flow.depth_time_policy.mode=='schedule_matched_v1'
assert cfg.model.inference.mixed_sampling_schedule=='depth_progress_spread_v1'
assert cfg.data.train.body_depth_encoding.version=='body_piecewise_10_80_10_v1'
OmegaConf.save(OmegaConf.to_container(cfg,resolve=True),out/'resolved_train.yaml')
report=plots_and_trace(cfg,out)
report['real_data']=real_data(cfg,out)
report['cpu_smoke']=tiny_model_smoke(cfg,out)
(out/'report.json').write_text(json.dumps(report,indent=2)+'\n')
print('CPU_CHECK_PASS',out)
PY
应检查六个 task/source 分组的真实样本:RGB/body/mixed 均为 [3,9,3,256,320]、action [32,14]、proprio [9,16],动作有效、没有终端 padding,在线 C 与训练 C 一致;输出还包含新旧伪彩对照、51 轮时钟 trace。小模型的 TinyVAE 不是生产 VAE 验证。
下面开始占用目标服务器 GPU,须由接收方确认卡已分配。没有自动显存门禁,也不停止别人的进程。将 GPU 字符串替换成目标机器允许的四张物理卡:
export V2_GPUS=GPU_A,GPU_B,GPU_C,GPU_D
unset CUDA_VISIBLE_DEVICES RANK LOCAL_RANK WORLD_SIZE
export V2_SMOKE="$V2_DEPLOY/smoke_v2"
"$(dirname "$V2_TRAIN_PY")/accelerate" launch \
--config_file scripts/accelerate_configs/accelerate_zero2_ds.yaml \
--num_processes 4 --gpu_ids "$V2_GPUS" --main_process_port 29671 \
scripts/train.py task=robotwin_three_mixed_ee_v2_local \
output_dir="$V2_SMOKE" resume=null max_steps=2 save_every=2 log_every=1 \
swanlab.enabled=true swanlab.mode=offline swanlab.transport=sdk \
numerics_debug.enabled=true numerics_debug.max_updates=2 \
numerics_debug.skip_final_checkpoint=false > "$V2_DEPLOY/smoke.log" 2>&1
检查四个 rank 的有限 loss/gradient、显存峰值、退出码、xwam_migration.rank*.json、mixed_precision.rank*.json、training_audit_rank*.jsonl 和 step 2 的权重/完整状态。不得因为 OOM 静默减 batch 或分辨率。
还应验证两种恢复,输出目录分别独立,仍用上面相同的 4 卡 launch 前缀:
- 完整状态恢复:
output_dir=$V2_DEPLOY/smoke_state_resume resume=$V2_SMOKE/checkpoints/state/step_000002 max_steps=3 model.xwam_pretrained_checkpoint=null model.skip_dit_load_from_pretrain=true;应从第 2 步和原 Adam 状态续至 3。 - 权重恢复:
output_dir=$V2_DEPLOY/smoke_weights_resume resume=$V2_SMOKE/checkpoints/weights/step_000002.pt max_steps=1 model.xwam_pretrained_checkpoint=null model.skip_dit_load_from_pretrain=true;是新的优化器/步数。
恢复测试均保留 lr_scheduler_total_steps=20000、save_every=0、numerics_debug.skip_final_checkpoint=true、SwanLab offline;不能把 smoke 的 max_steps 当成 LR 总时钟。正式训练之后重新从官方初始化开始,不从 smoke 续训。
8. 正式训练:tmux、SwanLab 和完整状态
只在前述目标机器验证通过后执行。正式 task 为 robotwin_three_mixed_ee_v2_local。训练停止点和 scheduler horizon 均为 20,000;200 步 warmup 后仍使用原 constant LR,不改变学习率、batch 或模型设计。
保持已有的“训练写本地 scalar journal,独立 CPU 进程上传 SwanLab”方式,避免上传重试影响训练。曲线必须上传到项目所有者当前的 mambaou/depthwam(https://swanlab.cn/@mambaou/depthwam),在该项目下创建新的 V2 实验身份;不可使用代跑者账号的项目,不可复用源端 V1/V2 run ID,更不能追加到当前 V1 的 r8c0fb1e。
2026-10-13 源端只读核实:当前登录用户 mambaou、空间类型 PERSON、角色 OWNER。按所有者最新交接,目标端登录已委托协作者提前处理,AI 不再安排登录或索要 Key。使用完成登录的同一 Linux 用户复用现有凭据,更换 Conda 环境本身不要求重新登录。若使用自定义 SWANLAB_ROOT,沿用实际目录并将其写入注册、训练、上传器及 tmux 启动脚本;若使用项目级 .swanlab,保持可读取凭据的工作目录。不要假定具体目录,不读取或输出密钥明文,不覆盖其他作业凭据。人工已处理是交接信息,不等于目标服务器身份已经由本包核验;仅设置 workspace 也不能代替有效授权。
注册前运行 ai_handoff_v1/check_swanlab_target.py(见下面命令),验证实际登录为 mambaou、目标路径属于 mambaou/depthwam。缺授权可以继续 CPU 准备与离线 smoke,但不能把正式训练静默改传到代跑者账户。正式启动前必须完成正确项目下的新 run 注册;训练过程中若暂时断网,继续保留本地 journal 并由独立上传器重试、报告云端落后,不因上传等待阻塞 GPU。
- 对接收方新 run 使用
scripts/bootstrap_v2_swanlab.py --root <新运行目录> --create注册一次。该目录需有request.json(swanlab_run=工作区/项目/新run_id、name/group、scalar_columns)和formal_config.json。scalar_columns包含train/loss、train/grad_norm、train/lr、performance/steps_per_sec、performance/samples_per_sec,再加 smokeloss_components的所有train/<key>。不要直接运行不带--create的旧布局调度入口。 - 正式训练传入
swanlab.transport=external swanlab.run_path=<新run路径>,训练只在formal/swanlog/scalars.jsonl写本地记录。 - 在另一个 tmux 中启动
upload_training_scalars.py;这个进程不占 GPU。它的--worker-statusJSON 需记录本次 accelerate 进程的真实train_pid;--exit-file需记录本次真实退出码和 UTCfinished_at。不能拿源端状态文件代替。
本次默认使用下面的 external 路径。未经项目所有者同意,不把正式训练替换成其他账户、项目或仅离线交付;网络临时中断不改变既定上传目标。新 run ID 和状态文件由接收方 AI 生成并持久保存,重试沿用同一身份。
export V2_RUN="$V2_DEPLOY/run_v2_20k"
export V2_SWAN_RUN=mambaou/depthwam/NEW_RUN_ID
mkdir -p "$V2_RUN"
把 NEW_RUN_ID 替换为未使用的新 run ID,其余 mambaou/depthwam 保持固定;复用人工提前完成的登录,确保下面各进程都能访问同一份既有授权。下面检查授权、生成注册材料并注册一次;身份不符或凭据不可见时报告具体问题,不自行重新登录。网络失败先检查注册状态和相同 run ID,不能换 ID 无限制创建实验。
cd "$V2_PROJECT"
CUDA_VISIBLE_DEVICES='' "$V2_TRAIN_PY" "$V2_TRANSFER/ai_handoff_v1/check_swanlab_target.py" --run "$V2_SWAN_RUN"
"$V2_TRAIN_PY" - <<'PY'
import json, os
from pathlib import Path
from omegaconf import OmegaConf
from scripts.check_mixed_depth_v2 import compose
root=Path(os.environ['V2_RUN'])
run=os.environ['V2_SWAN_RUN']
assert len(run.split('/'))==3 and run.split('/')[:2]==['mambaou','depthwam']
assert run.split('/')[-1] not in ('NEW_RUN_ID','r8c0fb1e')
cfg=compose('robotwin_three_mixed_ee_v2_local')
assert (cfg.max_steps,cfg.lr_scheduler_total_steps,cfg.save_every)==(20000,20000,2000)
cfg.output_dir=str(root/'formal')
cfg.swanlab.enabled=True; cfg.swanlab.transport='external'
cfg.swanlab.mode='cloud'; cfg.swanlab.run_path=run
cfg.swanlab.workspace='mambaou'; cfg.swanlab.project='depthwam'
cfg.performance.profile_updates=0; cfg.performance.trace_update=0
rows=[json.loads(s) for s in (Path(os.environ['V2_SMOKE'])/'training_audit_rank0.jsonl').read_text().splitlines()]
assert rows[-1]['step']==2
columns=['train/loss','train/grad_norm','train/lr','performance/steps_per_sec','performance/samples_per_sec']
columns += ['train/'+k for k in rows[-1]['loss_components']]
request=dict(swanlab_run=run,swanlab_name='robotwin-three-mixed-ee-v2-20k',
swanlab_group='robotwin-three-mixed-ee-v2',scalar_columns=sorted(set(columns)),
swanlab_description='V2 piecewise body encoding + spread depth schedule; fresh official weights, 20k updates.')
for name,value in [('request.json',request),('formal_config.json',OmegaConf.to_container(cfg,resolve=True))]:
with (root/name).open('x') as f: json.dump(value,f,ensure_ascii=False,indent=2)
print('SWANLAB_REGISTRATION_FILES_READY')
PY
CUDA_VISIBLE_DEVICES='' "$V2_TRAIN_PY" scripts/bootstrap_v2_swanlab.py --root "$V2_RUN" --create
将下面保存为新服务器的 launch_v2.sh。脚本读取前面 export 的路径/卡号/新 SwanLab 身份,所有路径均属于新服务器:
#!/usr/bin/env bash
set -euo pipefail
: "${V2_PROJECT:?}" "${V2_TRAIN_PY:?}" "${V2_GPUS:?}" "${V2_RUN:?}" "${V2_SWAN_RUN:?}" "${V2_TRANSFER:?}"
cd "$V2_PROJECT"
test ! -e "$V2_RUN/formal"
test -f "$V2_RUN/swanlab_registered.json"
CUDA_VISIBLE_DEVICES='' "$V2_TRAIN_PY" "$V2_TRANSFER/ai_handoff_v1/check_swanlab_target.py" --run "$V2_SWAN_RUN"
unset CUDA_VISIBLE_DEVICES RANK LOCAL_RANK WORLD_SIZE SWANLAB_MODE
mkdir -p "$V2_RUN/formal"
"$(dirname "$V2_TRAIN_PY")/accelerate" launch \
--config_file scripts/accelerate_configs/accelerate_zero2_ds.yaml \
--num_processes 4 --gpu_ids "$V2_GPUS" --main_process_port 29671 \
scripts/train.py task=robotwin_three_mixed_ee_v2_local \
output_dir="$V2_RUN/formal" resume=null max_steps=20000 save_every=2000 \
batch_size=4 gradient_accumulation_steps=1 lr_scheduler_total_steps=20000 \
swanlab.enabled=true swanlab.transport=external swanlab.mode=cloud \
swanlab.workspace=mambaou swanlab.project=depthwam \
swanlab.run_path="$V2_SWAN_RUN" \
performance.profile_updates=0 performance.trace_update=0 \
> "$V2_RUN/formal/train.log" 2>&1 &
v2_train_pid=$!
printf '{"train_pid":%s}\n' "$v2_train_pid" > "$V2_RUN/status.json"
set +e
wait "$v2_train_pid"
v2_exit_code=$?
set -e
export V2_EXIT_CODE="$v2_exit_code"
"$V2_TRAIN_PY" - <<'PY'
import datetime, json, os
from pathlib import Path
p=Path(os.environ['V2_RUN'])/'formal/exit_status.json'
p.write_text(json.dumps(dict(exit_code=int(os.environ['V2_EXIT_CODE']),
finished_at=datetime.datetime.now(datetime.timezone.utc).isoformat()))+'\n')
PY
exit "$v2_exit_code"
在已加载这些环境变量的终端,分别进入两个 tmux 会话执行(先 --create 注册成功):
tmux new-session -s depthwam-v2-train
bash "$V2_DEPLOY/launch_v2.sh"
tmux new-session -s depthwam-v2-swanlab
CUDA_VISIBLE_DEVICES='' "$V2_TRAIN_PY" scripts/upload_training_scalars.py \
--journal "$V2_RUN/formal/swanlog/scalars.jsonl" --run "$V2_SWAN_RUN" \
--worker-status "$V2_RUN/status.json" --output "$V2_RUN/upload" \
--exit-file "$V2_RUN/formal/exit_status.json" --watch \
> "$V2_RUN/upload.log" 2>&1
训练实际配置在 formal/config.yaml。另保留 Git base、overlay provenance、接收方路径配置、环境版本、公共模型哈希、GPU 型号和上述运行脚本。训练期间不要修改使用中的源码;后续代码修改应另开副本。
预期产物:
formal/checkpoints/weights/step_002000.pt ... step_020000.pt
formal/checkpoints/weights/step_002000.pt.interface.json
formal/checkpoints/state/step_002000/ ... step_020000/
formal/training_audit_rank0.jsonl # 其他 rank 同样保留
formal/swanlog/scalars.jsonl
formal/xwam_migration.rank0.json # 其他 rank 同样保留
formal/config.yaml
formal/train.log
formal/exit_status.json
upload/status.json
SwanLab 曲线、tmux 会话消失、日志最后出现 20k 都不能单独证明成功;需同时确认真实退出码、最终权重、.interface.json、完整状态和 trainer step=20000。完整状态恢复使用 resume=.../state/step_XXXXXX,同时 model.xwam_pretrained_checkpoint=null model.skip_dit_load_from_pretrain=true,保持同一 V2 分布和统计/窗口版本。.pt 仅是 weights-only,不含完整续训进度。
tmux 支持 SSH 断线后继续,不保证服务器重启后自动恢复。新服务器若要求开机恢复,需要另建对应机器的服务并核对完整 checkpoint,不能照搬源端服务文件。
9. 评测环境与资产装配
训练可先部署,评测须用独立环境。源端评测关键版本为 Python 3.10、PyTorch 2.4.1/CUDA 12.1、torchvision 0.19.1、SAPIEN 3.0.0b1、mplib 0.2.1、scipy 1.10.1、transforms3d 0.4.2、trimesh 4.4.3、open3d 0.18.0、PyAV 17.1.0;详见 eval_installed_versions.json。不要在评测环境执行会把 torch 升到训练版本的 pip install -e .。
建议先按关键版本建立环境,再按版本清单补齐模型侧纯 Python 依赖:
conda create -n depthwam-v2-eval python=3.10 pip -y
conda activate depthwam-v2-eval
python -m pip install torch==2.4.1 torchvision==0.19.1 \
--index-url https://download.pytorch.org/whl/cu121
python -m pip install numpy==1.26.4 scipy==1.10.1 sapien==3.0.0b1 mplib==0.2.1 \
transforms3d==0.4.2 trimesh==4.4.3 open3d==0.18.0 gymnasium==0.29.1 av==17.1.0 \
hydra-core==1.3.2 omegaconf==2.3.0 einops==0.8.1 transformers==4.49.0 \
safetensors==0.5.3 accelerate==1.12.0 rich==14.2.0 ftfy \
huggingface-hub==0.29.2 boto3==1.35.99 datasets==3.6.0 pyarrow==23.0.0 \
jsonlines==4.0.0 gitpython==3.1.45 opencv-python==4.11.0.86 \
toppra==0.6.10 imageio==2.34.2 imageio-ffmpeg==0.6.0 h5py==3.16.0 \
matplotlib==3.10.9 termcolor==3.3.0 pyglet==1.5.31 moviepy==2.2.1
export V2_EVAL_PY="$(command -v python)"
这是核心安装集合,不是已在新服务器验证过的完整锁文件;按实际 import 错误和交付版本清单补齐其他依赖,执行 pip check。如果本机无可用 Vulkan/EGL/NVIDIA 渲染后端,先解决渲染环境再跑模型。
Curobo 使用原版,不带此前 expert-fix 补丁:
git clone https://github.com/NVlabs/curobo.git "$V2_DEPLOY/curobo"
git -C "$V2_DEPLOY/curobo" checkout cca894de9ec74e77a0a4071319c81958991a9108
python -m pip install setuptools_scm wheel ninja
python -m pip install pybind11==3.1.0 numpy-quaternion==2024.0.13 \
yourdfpy==0.0.60 warp-lang==1.2.1 scikit-image==0.21.0 networkx pyyaml importlib_resources
# 确認 CUDA_HOME/nvcc 与评测 torch 的 CUDA 构建兼容后执行。
python -m pip install --no-deps --no-build-isolation -e "$V2_DEPLOY/curobo"
python -m pip check
在新目录装配资产:先大资产,再覆盖本次冻结后端的机器人 YAML,不能反过来。拷贝或 CoW 均可,不使用指向旧服务器的软链。
export V2_ROBOTWIN="$V2_DEPLOY/RoboTwin_eval_v2"
mkdir "$V2_ROBOTWIN"
mkdir "$V2_ROBOTWIN/assets"
cp -a --reflink=auto "$V2_UNPACK/robotwin_assets/." "$V2_ROBOTWIN/assets/"
cp -a "$V2_UNPACK/evaluation/RoboTwin_eval_private/." "$V2_ROBOTWIN/"
mkdir -p "$V2_ROBOTWIN/policy"
ln -s "$V2_PROJECT/experiments/robotwin/depthwam_policy" "$V2_ROBOTWIN/policy/depthwam_policy"
私有 curobo_left.yml、curobo_right.yml 中有源端绝对 URDF/碰撞球路径。只在新的 $V2_ROBOTWIN 副本中把前缀
/DATA/disk0/lyy/depthwam_robotwin_share/RoboTwin2.0_official/assets/
改为 $V2_ROBOTWIN/assets/。记录替换前后哈希,确认引用文件存在;球体数值、URDF、网格、关节顺序均不改。*_tmp.yml 的 ${ASSETS_PATH} 是原运行时代换变量,不直接当作 shell 变量展开。检查资产中的其他绝对路径:rg -n '/DATA/|/home/' "$V2_ROBOTWIN" -g '*.yml' -g '*.yaml' -g '*.py',区分运行路径与历史注释,禁止全目录盲目字符串替换。
10. 冻结案例迁移和评测命令
不要直接执行 scripts/run_robotwin_mixed_ee_eval.py 的原服务器 campaign。它目前仍有 V1 task/source、旧训练目录和旧 GPU 的硬编码。单任务入口可用,但必须设置 protocol 所需环境变量。
先复制 evaluation/preparation/ 到接收方新的 $V2_DEPLOY/eval_preparation_v2/,原包内证据保持原样。在新副本中:
- 修改
evaluation_protocol.json的private_harness_root、training_environment_seeds为接收方路径。 - 记录 V2 模型/编码/时间策略,将
mixed_inference.forward_rounds从 50 改成 51,明确depth_progress_spread_v1和body_piecewise_10_80_10_v1。旧camera_scales_m仅是 legacy 尺度,不能把它当作 V2 主相机完整编码规则。资源字段改为接收方实际允许卡号。 - 保持 stages、seed 范围、base seed、任务、unseen 指令、动作32步、相机 D435、官方成功标准及
selection_rule_id不变。 - 新协议文件重算 SHA256;仅在新副本各
evaluation_cases/{smoke,screen,final}/{source}/{task}/frozen_cases.json的protocol_sha256更新为新值,并增加迁移来源记录。断言所有cases数组逐项等于交付源文件,包括顺序、episode、seed、instruction。训练 seed 文件内容和哈希不变。 - 保留原
protocol_attempts.jsonl、origin_frozen_cases.json作为历史专家筛选证据。它们不是新模型结果,不能改写成 V2 的成功记录。
这属于路径/模型分布元数据迁移,不重新抽 seed、不重新选择示范。协议校验器要求 frozen cases 与协议 SHA 一致,不能通过删除校验绕过。
接收方应在 CPU 下对全部 18 个 task/source/stage 组合实例化 experiments.robotwin.six_task_protocol.ProtocolAudit,验证新协议、案例与训练 seed 无重叠,且每个阶段回合数为 1/10/100。保持冻结案例评测模式,不运行旧 expert-fix 版本。
上述迁移可使用以下 CPU 脚本;先确认 V2_GPUS 是目标机器允许的数字卡号,且机器人资产路径已经按上一节修正:
export V2_EVAL_PREP="$V2_DEPLOY/eval_preparation_v2"
CUDA_VISIBLE_DEVICES='' "$V2_TRAIN_PY" - <<'PY'
import copy, hashlib, json, os, shutil
from pathlib import Path
from omegaconf import OmegaConf
from scripts.check_mixed_depth_v2 import compose
from experiments.robotwin.six_task_protocol import ProtocolAudit
src=Path(os.environ['V2_UNPACK'])/'evaluation/preparation'
dst=Path(os.environ['V2_EVAL_PREP'])
gpu_ids=[int(v) for v in os.environ['V2_GPUS'].split(',')]
assert len(gpu_ids)==4 and len(set(gpu_ids))==4 and min(gpu_ids)>=0
shutil.copytree(src,dst) # 目标必须不存在
old=(src/'evaluation_protocol.json').read_bytes()
protocol=json.loads(old)
protocol['private_harness_root']=os.environ['V2_ROBOTWIN']
protocol['training_environment_seeds']=str(dst/'training_environment_seeds.json')
protocol['mixed_inference'].update(forward_rounds=51,schedule='depth_progress_spread_v1')
cfg=compose('robotwin_three_mixed_ee_v2_local')
protocol['body_depth_encoding']=OmegaConf.to_container(cfg.data.train.body_depth_encoding,resolve=True)
protocol['training_time_policy']=OmegaConf.to_container(cfg.model.flow.depth_time_policy,resolve=True)
protocol['resource_gate'].update(allowed_gpu_ids=gpu_ids,memory_gate_enabled=False,
max_used_percent=100,min_free_mib=0,max_workers=4,workers_per_gpu=1)
path=dst/'evaluation_protocol.json'
path.write_text(json.dumps(protocol,ensure_ascii=False,indent=2)+'\n')
sha=hashlib.sha256(path.read_bytes()).hexdigest()
case_records=[]
for f in sorted((dst/'evaluation_cases').glob('*/*/*/frozen_cases.json')):
source_file=src/f.relative_to(dst)
original=json.loads(source_file.read_text()); migrated=copy.deepcopy(original)
migrated['protocol_sha256']=sha
migrated['receiver_migration']=dict(source_sha256=hashlib.sha256(source_file.read_bytes()).hexdigest(),
original_protocol_sha256=original['protocol_sha256'],case_identity_unchanged=True)
assert migrated['cases']==original['cases']
f.write_text(json.dumps(migrated,ensure_ascii=False,indent=2)+'\n')
case_records.append(dict(file=str(f.relative_to(dst)),cases=len(migrated['cases']),
original_sha256=hashlib.sha256(source_file.read_bytes()).hexdigest(),
migrated_sha256=hashlib.sha256(f.read_bytes()).hexdigest()))
assert len(case_records)==18 and sum(x['cases'] for x in case_records)==666
os.environ['DEPTHWAM_SIX_PROTOCOL']=str(path)
os.environ['DEPTHWAM_SIX_FREEZE_ONLY']='0'
for stage,definition in protocol['stages'].items():
os.environ['DEPTHWAM_SIX_STAGE']=stage
os.environ['DEPTHWAM_SIX_CASES_ROOT']=str(dst/'evaluation_cases'/stage)
for scene in protocol['sources']:
for task in protocol['tasks']:
a=ProtocolAudit(task,'demo_'+scene,definition['base_seed'],
definition['episodes_per_task_source'],dst/'cpu_audit'/stage/scene/task)
for i,c in enumerate(a.frozen): assert a.seed_for_attempt(i,0)==c['seed']
for task in protocol['tasks']:
seeds={stage:{c['seed'] for scene in protocol['sources'] for c in
json.loads((dst/'evaluation_cases'/stage/scene/task/'frozen_cases.json').read_text())['cases']}
for stage in protocol['stages']}
assert not seeds['smoke']&seeds['screen'] and not seeds['smoke']&seeds['final'] and not seeds['screen']&seeds['final']
(dst/'receiver_migration.json').write_text(json.dumps(dict(
original_protocol_sha256=hashlib.sha256(old).hexdigest(),local_protocol_sha256=sha,
cases=case_records,cpu_protocol_check='PASS',simulator_run=False),indent=2)+'\n')
print('PROTOCOL_MIGRATION_CPU_PASS')
PY
单任务 smoke 示例(要使用**本次已训练的 V2 权重及同名 .interface.json**;第一次闭环 smoke 不计入初筛或最终分母):
cd "$V2_PROJECT"
export V2_EVAL_PREP="$V2_DEPLOY/eval_preparation_v2"
export DEPTHWAM_SIX_PROTOCOL="$V2_EVAL_PREP/evaluation_protocol.json"
export DEPTHWAM_SIX_STAGE=smoke
export DEPTHWAM_SIX_CASES_ROOT="$V2_EVAL_PREP/evaluation_cases/smoke"
export DEPTHWAM_SIX_FREEZE_ONLY=0
export V2_CKPT="$V2_RUN/formal/checkpoints/weights/step_008000.pt"
export V2_EVAL_GPU=AUTHORIZED_PHYSICAL_GPU
CUDA_VISIBLE_DEVICES="$V2_EVAL_GPU" "$V2_EVAL_PY" -m experiments.robotwin.eval_robotwin_single \
--config-name sim_robotwin_three_mixed_ee_v2_local \
task=robotwin_three_mixed_ee_v2_local_eval \
ckpt="$V2_CKPT" gpu_id=0 seed=61 \
EVALUATION.robotwin_root="$V2_ROBOTWIN" \
EVALUATION.task_name=click_bell EVALUATION.task_config=demo_clean \
EVALUATION.eval_num_episodes=1 \
EVALUATION.output_dir="$V2_DEPLOY/eval_v2/smoke/step_008000/clean" \
EVALUATION.inference_seed_mode=xwam_derived EVALUATION.instruction_type=unseen \
EVALUATION.replan_steps=32 EVALUATION.camera_type=D435 \
EVALUATION.visualize_future_video=false \
EVALUATION.visualization_depth_read_mode=isolated \
EVALUATION.visualize_future_video_preserve_fast_action=true
CUDA_VISIBLE_DEVICES 填物理卡,gpu_id=0 是受限可见列表内的下标,不代表使用目标服务器物理 GPU0。实际允许卡由接收方分配。
对其余任务/场景更换 task/config,各自使用独立输出目录。初筛/完整评测沿用同一命令,仅按表同时调整 stage、cases root、base seed、回合数和 checkpoint:
| 阶段 | DEPTHWAM_SIX_STAGE |
base seed |
每 task/source | 每 checkpoint 分母 |
|---|---|---|---|---|
| smoke | smoke | 61 | 1 | 6(单独记录) |
| 初筛 | screen | 62 | 10 | 60 |
| 完整评测 | final | 63 | 100 | 600 |
当前交付协议的筛选规则实际为 **top2_second_within_3**:初筛取前两名,再纳入距离第二名成功数不超过 3/60(5个百分点)的其他 checkpoint;同分按较早步数排序。它不是历史对话中的“只取最好、差2以内取两个”。以本次已冻结协议为准;如要改变,必须在查看 V2 分数之前另行冻结。
初筛完整四组共 240 回合;完整评测每入选 checkpoint 600。最终案例与本轮 smoke/screen 独立,但为公平对照复用过往 body-only/V1 的最终案例,所以不能声称它们是从未使用的新 benchmark。
实际 case seed、unseen 指令和 X-WAM derived 推理 seed 均要在输出日志核对。保留所有模型失败和环境错误;不换 seed、不把视频副本算 episode、不根据最终成绩再择优后只报告赢家。遇到资源失败或进程中断,保存现场后从明确边界处理,不能直接重跑覆盖同目录。
正式批量调度仍需接收方 AI 基于这些单任务命令在独立部署目录生成新 launcher,记录分卡、退出码和每组完成状态。任务书已将这一步列为明确工作项,包含 dry-run、断点边界和选模/分母要求;本 README 没有声称原服务器调度器已经适配完目标机器,也不授权修改评测协议来替代适配。
11. 交付验收与查询
训练查询:
tmux ls
tail -n 30 "$V2_RUN/formal/train.log"
tail -n 3 "$V2_RUN/formal/training_audit_rank0.jsonl"
cat "$V2_RUN/upload/status.json"
# 退出后再读取:
cat "$V2_RUN/formal/exit_status.json"
接收方回传:
SOURCE_SNAPSHOT.json、源码包 SHA256、基准与 overlay 哈希、所有_local配置、窗口路径重绑定记录、实际环境和 GPU 信息。- 下载/解压校验结果,CPU 检查结果,4卡 smoke/两种恢复结果,官方权重逐项迁移报告。
- 正式训练配置、分项 loss/梯度、显存峰值、每步耗时、SwanLab 地址、退出状态和 2k–20k 全部 10 组 checkpoint 清单。
- 新评测协议/案例迁移审计、每次启动命令及退出码、逐回合结果;分开报告 smoke、四组初筛、每个入选模型最终 600 回合。
12. 当前已验证和仍待验证
源端已完成:105.72 GiB/81 文件上传;逐文件远端 SHA256/大小确认;完成标记确认;传输工具 8 项 CPU 测试通过。V2 的历史小模型/真实数据验证见代码包内 docs/mixed_depth_v2.md,其中记录的是对应日期的验证范围。
本 README 的代码块在源端执行语法检查;完整源码恢复、20k 配置生成和无 Git 历史的 CPU 测试结果记录于相邻 README.validation.json。路径/协议迁移沿用此前轻量 CPU 演练结果。这不等于在目标服务器安装成功。
接收方仍需完成:ModelScope 下载审批、环境安装、公共组件下载、路径/协议绑定、真实 VAE/官方权重 GPU 冒烟、4卡通信、完整状态恢复、机器人闭环和正式实验。源端此前 V2 运行在首次 2k 保存前结束,迁移包没有可用的生产 V2 完整续训 checkpoint;本次按官方权重新初始化。
本 README、source_complete_v1/ 和 ai_handoff_v1/ 是迁移包上传完成后新增的交付文件,**不属于原封存迁移包的 81 个文件,也未加入原 transfer_manifest.json**。交接清单 ai_handoff_v1/HANDOFF_MANIFEST.json 另行记录完整源码、部署说明及 AI 文档的哈希;不修改已封存分包、清单或旧完成标记。
- Downloads last month
- 3,485