ngu-eidf-kit / RELEASE.json
wckwan's picture
NGU EIDF kit v1: WebShop Qwen3-8B NGU seed 2000 (two 4-GPU pods)
6945b09 verified
Raw History Blame Contribute Delete
3.17 kB
{
"release": "ngu-eidf-kit v1",
"built": "2026-11-13T16:21Z",
"run": {
"arm": "wsq3_ngu_s2000",
"name": "Webshop_Qwen3-8B_ngu_nostd_s2000",
"seed": 2000,
"wandb": "cyruskwan/PRM/w7kzz0mg",
"model": "Qwen/Qwen3-8B",
"model_revision": "b968826d9c46dd6066d109eabc6255188de91218",
"data.max_prompt_length": 5120,
"max_prompt_length_note": "user choice 2026-11-13: as NGU seed 1000; the GRPO seed-2000 control itself ran at 4096",
"control": "WebShop Qwen3 GRPO no-std seed-2000 = seed-1000 control file with env.seed, run name and dump dir changed",
"ngu_args": "--k 8 --banks 64 --replica-gpu-memory-utilization 0.95 --free-expired-rows --malloc-trim --record-all-rollouts",
"learner_profile_80gb": "actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 (train_ngu.sh, GPUs < 100 GB)"
},
"code": {
"repo": "KwanWaiChung/search_prm (private)",
"base_commit": "de09e82eeefa3edbe877cd060cf5c5e94ff2b443",
"snapshot": "git archive <base_commit> verl-agent + train_scripts/ngu/{train_ngu.sh,wandb_ids.txt} + train_scripts/dapo/controls/wsq3_nostd_s2000.overrides from the huawei tree",
"files_not_in_base_commit": {
"train_scripts/ngu/train_ngu.sh": "0cb13ccbffb00a0fc748fbbc0cdf1f576f1f665ba619bfedd82aa7e5345da2b6",
"train_scripts/ngu/wandb_ids.txt": "13447c063a357ca7bb5dd028af7d47a0be24c825180934469837ca2474357e40",
"train_scripts/dapo/controls/wsq3_nostd_s2000.overrides": "c53c46d226d13f758468b2f4f172df68aac14d1aef4dafe753a0d32e83995892"
}
},
"environment": {
"name": "verl-agent-webshop",
"python": "3.10.22",
"torch": "2.8.0+cu128",
"vllm": "0.11.0",
"ray": "2.50.0",
"flash_attn": "2.8.3 (sm_80/sm_86/sm_90)",
"packed_from": "/root/micromamba/envs/verl-agent-webshop (huawei, Ubuntu 22.04, glibc 2.35)",
"audit_after_unpack": {
"same": 85960,
"prefix_only": 356,
"differ": 107,
"missing": 0,
"extra": 6,
"overlay_with_prefix": 91
},
"audit_after_overlay": {
"same": 85978,
"prefix_only": 445,
"differ": 0,
"missing": 0,
"extra": 0
},
"overlay": "original bytes of files conda-pack could not reproduce (pip-installed packaging + console scripts); bootstrap rewrites the packing prefix in its text files"
},
"webshop_data": "data/ + search_engine/{indexes*,resources*} exactly as used by NGU seed 1000 on huawei",
"artefacts": {
"kit.tar.gz": {
"sha256": "1451834460baad6e205ac93b9b1056ce349cbe7e53ea39e08ad4b518a47d918c",
"bytes": 28368
},
"search_prm-ngu-kit.tar.gz": {
"sha256": "46564bae6b32a6cdbf4f3ad74f2e5e1a3766ca0c5fc19142e53830c359cee450",
"bytes": 40700504
},
"text-data.tar.gz": {
"sha256": "86c434abde9bb92fcbcfee26a71729d93752bcbed9b0f2b21e5928905425186e",
"bytes": 3194
},
"verl-agent-webshop-overlay-x86_64.tar.gz": {
"sha256": "e7dd50f4e0dc3fe5f46ce1c7e36ba1316ba5cca7b5935cc492cae1b0528faabc",
"bytes": 7338345
},
"verl-agent-webshop-x86_64.tar.gz": {
"sha256": "8e31c35901503795a37b10cba4e1d505335be3c3af8e8bec5f2245f9adf9a1d0",
"bytes": 6750303386
},
"webshop-data-index.tar.gz": {
"sha256": "76c870cdb4e73bc34286891016a1b955cbdc05468589d92b5ae61acb048c4285",
"bytes": 12760969
}
}
}