Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code: DatasetGenerationError
Exception: CastError
Message: Couldn't cast
status: string
tests: int64
GPU_smoke_passed: bool
mean_l1_flow_loss: double
rank_samples: list<item: struct<global_sample_index: int64, loss: double>>
child 0, item: struct<global_sample_index: int64, loss: double>
child 0, global_sample_index: int64
child 1, loss: double
backend: struct<backend: string, version: int64, fp32_ieee: bool, fixed_reduce: bool, stable_route_order: boo (... 77 chars omitted)
child 0, backend: string
child 1, version: int64
child 2, fp32_ieee: bool
child 3, fixed_reduce: bool
child 4, stable_route_order: bool
child 5, num_modules: int64
child 6, failure_policy: string
child 7, implementation_sha256: string
samples: int64
parent_seen_validation: bool
seconds: double
scope: string
to
{'status': Value('string'), 'samples': Value('int64'), 'mean_l1_flow_loss': Value('float64'), 'rank_samples': List({'global_sample_index': Value('int64'), 'loss': Value('float64')}), 'seconds': Value('float64'), 'backend': {'backend': Value('string'), 'version': Value('int64'), 'fp32_ieee': Value('bool'), 'fixed_reduce': Value('bool'), 'stable_route_order': Value('bool'), 'num_modules': Value('int64'), 'failure_policy': Value('string'), 'implementation_sha256': Value('string')}, 'scope': Value('string'), 'parent_seen_validation': Value('bool')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
for key, table in generator:
^^^^^^^^^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
for item in generator(*args, **kwargs):
~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, 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 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
status: string
tests: int64
GPU_smoke_passed: bool
mean_l1_flow_loss: double
rank_samples: list<item: struct<global_sample_index: int64, loss: double>>
child 0, item: struct<global_sample_index: int64, loss: double>
child 0, global_sample_index: int64
child 1, loss: double
backend: struct<backend: string, version: int64, fp32_ieee: bool, fixed_reduce: bool, stable_route_order: boo (... 77 chars omitted)
child 0, backend: string
child 1, version: int64
child 2, fp32_ieee: bool
child 3, fixed_reduce: bool
child 4, stable_route_order: bool
child 5, num_modules: int64
child 6, failure_policy: string
child 7, implementation_sha256: string
samples: int64
parent_seen_validation: bool
seconds: double
scope: string
to
{'status': Value('string'), 'samples': Value('int64'), 'mean_l1_flow_loss': Value('float64'), 'rank_samples': List({'global_sample_index': Value('int64'), 'loss': Value('float64')}), 'seconds': Value('float64'), 'backend': {'backend': Value('string'), 'version': Value('int64'), 'fp32_ieee': Value('bool'), 'fixed_reduce': Value('bool'), 'stable_route_order': Value('bool'), 'num_modules': Value('int64'), 'failure_policy': Value('string'), 'implementation_sha256': Value('string')}, 'scope': Value('string'), 'parent_seen_validation': Value('bool')}
because column names don't match
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1879, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
status string | samples int64 | mean_l1_flow_loss float64 | rank_samples list | seconds float64 | backend dict | scope string | parent_seen_validation bool |
|---|---|---|---|---|---|---|---|
passed | 16 | 0.272728 | [
{
"global_sample_index": 0,
"loss": 0.23442745208740234
},
{
"global_sample_index": 4,
"loss": 0.286819726228714
},
{
"global_sample_index": 8,
"loss": 0.20375239849090576
},
{
"global_sample_index": 12,
"loss": 0.3037784993648529
}
] | 8.750602 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.274583 | [
{
"global_sample_index": 0,
"loss": 0.2356635332107544
},
{
"global_sample_index": 4,
"loss": 0.2917252779006958
},
{
"global_sample_index": 8,
"loss": 0.2034410685300827
},
{
"global_sample_index": 12,
"loss": 0.30443015694618225
}
] | 10.677722 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.272728 | [
{
"global_sample_index": 1,
"loss": 0.3678995668888092
},
{
"global_sample_index": 5,
"loss": 0.2966793179512024
},
{
"global_sample_index": 9,
"loss": 0.28742343187332153
},
{
"global_sample_index": 13,
"loss": 0.21449312567710876
}
] | 8.740261 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.274583 | [
{
"global_sample_index": 1,
"loss": 0.375865638256073
},
{
"global_sample_index": 5,
"loss": 0.29706254601478577
},
{
"global_sample_index": 9,
"loss": 0.28871771693229675
},
{
"global_sample_index": 13,
"loss": 0.21663805842399597
}
] | 11.113681 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.272728 | [
{
"global_sample_index": 2,
"loss": 0.1911766231060028
},
{
"global_sample_index": 6,
"loss": 0.3328573703765869
},
{
"global_sample_index": 10,
"loss": 0.3075946569442749
},
{
"global_sample_index": 14,
"loss": 0.24543029069900513
}
] | 8.642949 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.274583 | [
{
"global_sample_index": 2,
"loss": 0.19174520671367645
},
{
"global_sample_index": 6,
"loss": 0.3358871042728424
},
{
"global_sample_index": 10,
"loss": 0.3047950863838196
},
{
"global_sample_index": 14,
"loss": 0.2462688833475113
}
] | 11.090173 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.272728 | [
{
"global_sample_index": 3,
"loss": 0.3611660897731781
},
{
"global_sample_index": 7,
"loss": 0.206782266497612
},
{
"global_sample_index": 11,
"loss": 0.2087506353855133
},
{
"global_sample_index": 15,
"loss": 0.31462061405181885
}
] | 8.625415 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.274583 | [
{
"global_sample_index": 3,
"loss": 0.3684968650341034
},
{
"global_sample_index": 7,
"loss": 0.20669400691986084
},
{
"global_sample_index": 11,
"loss": 0.20951424539089203
},
{
"global_sample_index": 15,
"loss": 0.31639036536216736
}
] | 11.113739 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.272728 | [
{
"global_sample_index": 0,
"loss": 0.23442745208740234
},
{
"global_sample_index": 4,
"loss": 0.286819726228714
},
{
"global_sample_index": 8,
"loss": 0.20375239849090576
},
{
"global_sample_index": 12,
"loss": 0.3037784993648529
}
] | 8.31806 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.274583 | [
{
"global_sample_index": 0,
"loss": 0.2356635332107544
},
{
"global_sample_index": 4,
"loss": 0.2917252779006958
},
{
"global_sample_index": 8,
"loss": 0.2034410685300827
},
{
"global_sample_index": 12,
"loss": 0.30443015694618225
}
] | 8.160912 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.272728 | [
{
"global_sample_index": 1,
"loss": 0.3678995668888092
},
{
"global_sample_index": 5,
"loss": 0.2966793179512024
},
{
"global_sample_index": 9,
"loss": 0.28742343187332153
},
{
"global_sample_index": 13,
"loss": 0.21449312567710876
}
] | 8.46461 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.274583 | [
{
"global_sample_index": 1,
"loss": 0.375865638256073
},
{
"global_sample_index": 5,
"loss": 0.29706254601478577
},
{
"global_sample_index": 9,
"loss": 0.28871771693229675
},
{
"global_sample_index": 13,
"loss": 0.21663805842399597
}
] | 8.512253 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.272728 | [
{
"global_sample_index": 2,
"loss": 0.1911766231060028
},
{
"global_sample_index": 6,
"loss": 0.3328573703765869
},
{
"global_sample_index": 10,
"loss": 0.3075946569442749
},
{
"global_sample_index": 14,
"loss": 0.24543029069900513
}
] | 8.398312 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.274583 | [
{
"global_sample_index": 2,
"loss": 0.19174520671367645
},
{
"global_sample_index": 6,
"loss": 0.3358871042728424
},
{
"global_sample_index": 10,
"loss": 0.3047950863838196
},
{
"global_sample_index": 14,
"loss": 0.2462688833475113
}
] | 8.625894 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.272728 | [
{
"global_sample_index": 3,
"loss": 0.3611660897731781
},
{
"global_sample_index": 7,
"loss": 0.206782266497612
},
{
"global_sample_index": 11,
"loss": 0.2087506353855133
},
{
"global_sample_index": 15,
"loss": 0.31462061405181885
}
] | 8.476547 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
passed | 16 | 0.274583 | [
{
"global_sample_index": 3,
"loss": 0.3684968650341034
},
{
"global_sample_index": 7,
"loss": 0.20669400691986084
},
{
"global_sample_index": 11,
"loss": 0.20951424539089203
},
{
"global_sample_index": 15,
"loss": 0.31639036536216736
}
] | 8.297037 | {
"backend": "validation_deterministic_fp32_moe_v1",
"version": 1,
"fp32_ieee": true,
"fixed_reduce": true,
"stable_route_order": true,
"num_modules": 36,
"failure_policy": "SystemExit; no upstream BF16 fallback",
"implementation_sha256": "c906854d3afd9f94126b213405fc94e1c522bc8ea6f9fd19598ea29bdf5a266a... | fixed smoke subset, not full validation | true |
前置 E5+C3＋投影＋第二路门控实验
源端四 A100 的完整 smoke 25897517 已通过并独立审计：4 步主训练、两次原路径与恒等门控 loss/实际 noise-time/全骨干梯度逐位一致、门控更新、DCP 模型和优化器新进程精确重载、FP32 导出、同一临时父的 2+2 步尾部转接与首步 loss/顺序/梯度/更新/Adam exact，两个尾部最终 delta 也完全一致。所有容差为零。证据见 evidence/joint_independent_audit.json 与 evidence/source_smoke/。
这不是训练质量或闭环成功率结论。源端没有运行完整17180＋1000训练；目标机器尚未运行 GPU。包中不含训练后的4步权重，也没有完整80回合仿真环境、冻结确认集和部署评测入口；离线验证不替代闭环评测。
接收与准备
从 START_HERE.md 固定 commit/manifest/bootstrap SHA 获取本包。不要从可变 main 下载训练载荷。bootstrap.py 匿名获取固定公共资源、校验哈希并准备隔离 workspace，不启动 GPU 或训练。它仅复用旧三路交付中的公共数据/环境；旧三路任务说明和旧父权重不是本实验。前置主训练从 robbyant/lingbot-vla-v2-6b@11c703bf6a5c1f45b3b69168482da11fdbba53d7 的完整6片权重重新开始，不从25723372或本次smoke权重开始。
需要 Python3.12.13、CUDA12.8、Torch2.8.0、Transformers4.57.3、flash-attn2.8.3、Triton3.4.0；完整环境冻结见 environment/requirements.txt 和 source-freeze.txt。Dockerfile 是可选构建入口，未在新机器上实测构建。依赖安装失败时保留日志、修复环境，不自动更换训练精度或内核。
源端：4×A100 40GB、64CPU、440GB内存请求。主训练峰值 allocated约31.20GiB/卡、reserved约37.51GiB/卡，尾部约26.35GiB/卡。目标应预留同等级显存、主存及足够本地编译空间；四卡完整权重/优化器DCP和数据建议准备约1TB空盘。预检会要求容器只暴露4卡且每卡至少38GiB。不同硬件能否通过原exact门禁尚待实测，不自动降低micro或放宽容差。
原清单含固定绝对路径。为保持配置和数据顺序字节不变，必须按下述容器路径绑定，不改YAML或manifest里的路径。工作目录必须持久化；编译缓存默认容器本地 /tmp，不能放NFS/Lustre。需要时将本机高速盘绑定到 /tmp/deepstack-local。
export PACKAGE=/data/deepstack-joint-package
export WORKSPACE=/data/deepstack-joint
# bootstrap 已完成，再构建环境；构建不运行训练。
docker build -t deepstack-joint -f "$PACKAGE/environment/Dockerfile" "$PACKAGE"
# 以下命令才实际运行目标机器GPU smoke。
docker run --rm --gpus '"device=0,1,2,3"' --ipc=host --shm-size=64g \
-v "$PACKAGE:/package:ro" -v "$WORKSPACE:/workspace" \
-v "$WORKSPACE/runtime/source:/mnt/FAST/Lingbot-Challenge" \
deepstack-joint /opt/deepstack/bin/python /workspace/joint_tools/run.py smoke --workspace /workspace
必须得到 latest_joint_smoke.json 的 passed、根 acceptance，以及四rank原重复/恒等、门控梯度/更新、新进程DCP模型/优化器重载、FP32导出和两个尾部的全部严格收据。诊断exit0、CPU测试、单个rank或主训练4步完成均不算完整通过。失败不要盲目重复，保留日志并检查真实张量/运行状态；不选择恰好匹配的参照、不更换种子、不放宽容差。
正式实验：先另获执行授权，绝不自动启动
两臂拟比较各为17180步E5+C3主训练＋1000步相同尾部，最终模型有效更新18180步。只改变第二路门控是否在主训练期间学习；前置组从原预训练底座新训，全部原E5+C3权重加第二路门控可学；后置的旧25723372是否可公平复用仍需同一性核验，不能仅凭此smoke认定。
主阶段：50任务2500 clean轨迹、549787帧，原固定顺序/seed7/C3/历史配置；4卡micro8/GA2/global64、BF16/FSDP2/activation checkpoint/compile flex_cached；原Muon/原fallback组和LR时标17180不变，门控单独AdamW(.9,.95)、LR1e-4、同cosine时标、weight decay0。深度/未来深度/视频原辅助权重各.004，原路由损失保留。门控gamma初始1、仅active[1]、范围0–2。初始门控与后置参考逐位一致，文件固定在源码assets中。
尾部：骨干冻结，三路现有投影＋第二路门控训练；4卡micro1/GA16/global64、eager/joint、固定64000帧与逐样本noise/time、原图/C3双视图监督、投影LR1e-5/门控LR1e-4、AdamW(.9,.999)、wd0、identity1e-3、consistency0；正常观测/指令/状态，无隐藏供体、分割真值或未见指令训练。主训练原LR不是尾部LR，不混用。
prepare_formal.py 仅复制已smoke源码并把验收步数/回执边界推广到17180主步＋1000尾步，写出逐文件前后SHA；原smoke源码不改写，模型/训练器/YAML/数据/优化器数学不变。先从该正式主阶段最终父跑2步丢弃参考，再从同父重新跑1000步并严格对照首步；2步参考不进入最终模型。此长程编排仅经CPU检查，尚无完整长程GPU运行，不能说已训练通过。
目标smoke已通过且用户另行明确授权完整训练后才执行：
# 在同一容器环境准备独立formal_source，不运行GPU。
docker run --rm -v "$WORKSPACE:/workspace" \
deepstack-joint /opt/deepstack/bin/python /workspace/joint_tools/prepare_formal.py \
--workspace /workspace --confirm-formal-start
# 以下才开始正式17180+1000训练。不是本包发布时已启动的作业。
docker run --rm --gpus '"device=0,1,2,3"' --ipc=host --shm-size=64g \
-v "$WORKSPACE:/workspace" \
-v "$WORKSPACE/formal_source:/mnt/FAST/Lingbot-Challenge" \
deepstack-joint /opt/deepstack/bin/python /workspace/joint_tools/run.py train \
--workspace /workspace --confirm-formal-start
无自动重启、自动OOM改配置或覆盖旧输出。正式入口拒绝无目标smoke、变更硬件/环境或重复formal_launch。中断时保留完整DCP、RNG、数据游标及日志，接收agent先核验后设计严格续跑，不能假称从头重跑等价续跑。
交回与评估
交回目标smoke完整证据、环境/硬件/源码与数据manifest SHA；正式完成后交回step17180完整DCP/FP32导出/逐rank首步与最终收据、250帧尾部验证覆盖、1000步完整顺序和最后delta及SHA。包内历史4步父只用于工程验证，不能标成正式step17180父。开发验证数据父已见，不作独立泛化证据。
后续闭环必须另部署该最终FP32模型，先GPU启动与RPC逐位验收，再在共同冻结场景/指令/初始观测上配对评估原E5+C3、后置与前置。此包未提供完整80回合仿真评测runner，也不自动提交任何评测。不要仅根据离线loss选择赢家。
- Downloads last month
- 25