kermt-continue-pretrain

name	kermt-continue-pretrain
description	Continue pretraining from an existing KERMT checkpoint. The skill validates the user's checkpoint and pretrain CSV, prepares the data into shard/vocab/features form, then launches pretrain_ddp.py inside the kermt container (detached for long runs). Auto-dispatches `--pretrain_mode` based on the checkpoint type (grover_base vocab-only, cmim, or hybrid).
license	Apache-2.0
compatibility	Requires docker, nvidia-container-toolkit, and a CUDA-capable NVIDIA GPU. Designed for Claude Code, Codex, and Nemotron.
metadata	{"owner":"[e-mail]","classification":"workflow-skill","risk_tier":"skill"}

kermt-continue-pretrain

Continue pretraining from a user-supplied KERMT checkpoint (grover_base / cmim / hybrid). The skill is the workflow orchestrator: it validates inputs, prepares the corpus, launches the runner, and returns a run directory.

Hardware requirements

• GPUs: 1–N CUDA-capable NVIDIA GPUs. The runner auto-detects via torch.cuda.device_count(); --gpus 0,2 overrides. On a single GPU the runner falls back to --batch_size 32 --save_interval 500; on multi-GPU it uses the defaults_pretrain.json values (currently batch_size 256). Note: --gpus N uses torch.cuda indexing, which can differ from nvidia-smi's display order on multi-GPU hosts (PCI bus vs. CUDA enumeration). To target a specific physical GPU, set CUDA_VISIBLE_DEVICES before invoking, or run python -c "import torch; print([torch.cuda.get_device_name(i) for i in range(torch.cuda.device_count())])" to confirm which device you're picking.

• VRAM: the default --batch-size 256 is sized for A100-class hardware (80 GB VRAM). On smaller GPUs, downscale to avoid OOM:

  GPU class	VRAM	Suggested --batch-size
  L4, T4, V100 16 GB	16–24 GB	32–64
  A100 40 GB, L40, A40	40–48 GB	128
  A100 80 GB, H100, H200	80 GB	256 (default)

  These are rough starting points — pass --batch-size N to override.

• Disk: tens of GB depending on corpus size + epochs (each checkpoint is several hundred MB).

• Driver / CUDA: any host supporting CUDA 12.6 (the kermt image base). kermt-setup validates this up-front.

Inputs

Required:

• --csv <path> — the pretrain CSV (single column smiles). If you have separate train/val CSVs, pass --val-csv <path> too.

Checkpoint (optional — defaults to the released model if omitted):

• --ckpt <path> — the input pretrain checkpoint to continue from. Must be a grover_base (with vocab heads), cmim, or hybrid ckpt; the validator rejects everything else with a redirect to the correct workflow. If omitted, the skill offers to download the released pretrained hybrid model nvidia/NV-KERMT-70M-v2 and continue-pretrain from it — see "Resolve & validate the checkpoint" (workflow step 3). The released bundle ships its three vocab files alongside the ckpt, so the authoritative-vocab pass-through (step 5) works automatically.
• --pretrained-release — explicit opt-in to use the released model without the interactive prompt (for non-interactive / agent runs). Mutually exclusive with --ckpt.
• --model-dir <dir> — where to save the downloaded bundle (default $KERMT_REPO/models/NV-KERMT-70M-v2/). An already-complete bundle there is reused, not re-downloaded.

Optional:

• --val-csv <path> — separate validation CSV. Without it, the prep step auto-splits the input by --val-frac 0.1 (random shuffle with --seed).
• --epochs N / --batch-size N / --init-lr F / --max-lr F / --final-lr F / --warmup-epochs F / --weight-decay F / --dropout F / --save-interval N / --seed N — training-hyperparameter overrides. Anything not given is filled from agent/config/defaults_pretrain.json.
• --vocab-loss-weight F (hybrid only) / --latent-dim N / --contrastive-temperature F (cmim and hybrid only) — loss / decoder overrides.
• --wandb-project NAME / --wandb-run-name NAME — optional Weights & Biases logging. When --wandb-project is set, rank 0 logs train/val losses; the run name is honored only alongside a project. Off by default. (Independent of the ckpt's wandb_run_id continuity handling under --resume.)
• --resume — see "Modes" section below.
• --gpus 0,2 — restrict to a GPU subset. Default uses all visible GPUs.
• --from-prepare <dir> — skip the prepare step and reuse an existing prepare_data.json in <dir>. Useful when iterating on hyperparameters.

Modes

The runner has two modes for ingesting the input ckpt, dispatched on whether --resume is set. Pick based on intent:

Default (fresh-schedule continue-pretrain)

Use when: you have a finished pretrain ckpt and want to continue training it — on a new corpus, with a different objective, or just for more epochs than its original plan. The previous training's step counter and schedule shape are no longer relevant; you want a new learning-rate schedule for the new run.

What gets loaded from the ckpt:

• ✓ Model weights (encoder + vocab heads + contrast head + decoder, whatever is there)
• ✓ Optimizer state (Adam's running m1/m2 moments — warm-starts the new schedule so the first few hundred steps aren't dominated by noisy gradient-estimate startup)
• ✗ Scheduler step counter (reset to 0)
• ✗ Epoch counter (reset to 0)
• ✗ Batch counter (reset to 0)
• ✗ wandb run id (new wandb run, not a continuation)

Schedule shape (init/max/final LR, warmup epochs, total epochs): from your CLI args or defaults_pretrain.json. A fresh NoamLR is constructed from these values and starts at step 0.

--resume (true resume)

Use when: a previous run was interrupted (crash, OOM, Ctrl-C) and you want to pick up exactly where it left off — same dataset, same schedule, same training trajectory.

What gets loaded from the ckpt: everything in the save_model_for_restart format. Model weights + optimizer state + scheduler_step + epoch + batch_idx + wandb_run_id are all restored. The new run continues from the saved step in the saved schedule (which is recovered from the ckpt's saved_args). Mid-epoch resume works too — pretrain_ddp.py's sampler skip-count picks up at the saved batch index within the saved epoch.

Schedule shape: inherited from the ckpt's saved_args. CLI overrides of any schedule flag (--epochs / --warmup-epochs / --init-lr / --max-lr / --final-lr) are rejected with a hard error — pure resume means pure resume; if you want to change the schedule, drop --resume and start a fresh-schedule run.

Requirements: the ckpt must have been saved via save_model_for_restart (i.e., carry optimizer / scheduler_step / epoch / batch_idx keys). If any of these is missing, the runner errors with a clear message and suggests dropping --resume.

The default mode is the right choice ~90% of the time. Reach for --resume only when you genuinely need to continue a single interrupted training run.

Workflow

Let $KERMT_REPO be the path to your kermt repo checkout, and assume kermt-setup has already built kermt:latest. All paths below are on the host; the helper bind-mounts them at known container paths.

1. Pre-flight: ensure container + system probe.

   $KERMT_REPO/agent/scripts/kermt_container.sh check_system | python -c "
   import json, sys; d = json.load(sys.stdin)
   if not d['ok']:
       print('System check failed:', d['gaps']); sys.exit(1)
   print(f'OK: {len(d[\"gpus\"])} GPU(s); {d[\"disk\"][\"free_gb\"]} GB free; CUDA via container toolkit')
   "
   

   Surface any gaps to the user. Refuse to proceed if ok: false.

2. Compute run directory.

   RUN_DIR=$KERMT_REPO/runs/continue-pretrain_$(date -u +%Y-%m-%dT%H-%M-%SZ)
   

3. Resolve & validate the checkpoint.

   Resolve — only if --ckpt was omitted. Default to the released pretrained hybrid model nvidia/NV-KERMT-70M-v2:

   • Consent gate. Unless --pretrained-release was passed, ask the user: "No checkpoint given — download the released model nvidia/NV-KERMT-70M-v2 (NVIDIA Open Model License, https://huggingface.co/nvidia/NV-KERMT-70M-v2) and continue-pretrain from it? [y/N]". Never download without an explicit yes (or --pretrained-release). If both --ckpt and --pretrained-release are given, abort — they conflict.
   • Save location. Default $KERMT_REPO/models/NV-KERMT-70M-v2/; honor --model-dir <dir> if given. An already-complete bundle is reused.
   • Download (foreground; ~282 MB on first fetch):

     $KERMT_REPO/agent/scripts/kermt_container.sh run --model-dir <save-dir> -- \
         "python agent/scripts/fetch_released_model.py --out /model"
     

     Parse the JSON; abort on ok: false (surface errors). On success set <user-ckpt> = <save-dir>/kermt_contrastive_v2.0.pt. The bundle's three vocab files land in <save-dir> too, so step 5's --vocab-dir auto-detection (which looks in the ckpt's parent directory) finds them with no extra work.

   Validate the resolved (or user-provided) ckpt:

   $KERMT_REPO/agent/scripts/kermt_container.sh run --ckpt <user-ckpt> -- \
       "python agent/scripts/check_checkpoint.py --mode continue_pretrain --ckpt /ckpt"
   

   Parse the JSON. Abort on , showing the error verbatim. The error message redirects the user to for encoder-only ckpts, or to for finetuned ckpts.

Hard rules

• Never download the released model without consent. When --ckpt is omitted, download nvidia/NV-KERMT-70M-v2 only after an explicit user "yes" or an explicit --pretrained-release flag. --ckpt and --pretrained-release are mutually exclusive.
• Never modify the user's input ckpt. The runner symlinks it into the save_dir; the symlink is what pretrain_ddp.py auto-resumes from. The source file stays untouched.
• Never silently override arch. If the user passes a --hidden-size etc. that doesn't match the ckpt-derived value, the runner aborts loudly. Arch params come from the ckpt, period.
• Never block on the long-running pretrain itself. The runner is invoked via run_detached; the skill returns immediately after step 8. Use kermt-monitor for progress.
• Echo applied defaults back to the user. The args_applied field of run.json records every flag's value + source (user / default-config / auto-1gpu / auto-multi-gpu). Skill should surface a summary of any flag not user-specified so the user knows what was assumed.

Common errors

• model_type='finetuned' rejected → the ckpt is a downstream finetune, not a pretrain. The error redirects to the relevant workflow.
• grover_base ckpt has no vocab head → encoder-only ckpt (e.g. the original-grover grover_base.pt). The error redirects to kermt-add-cmim-pretrain.
• prepare_data manifest is missing required outputs → user passed --from-prepare to a directory where prepare was run with --skip-vocab or --skip-split. Re-run prepare without those flags.
• --gpus all not available → install nvidia-container-toolkit; check kermt_container.sh check_system.

Replayability

The run.json cmd_replay field is a single-line command that re-runs the pretrain with the same inputs, hyperparameters, and arch. To replay:

# Inside the kermt container:
$(jq -r .cmd_replay $RUN_DIR/run.json)

If ok_to_replay: false in the manifest (because the kermt repo working tree was dirty at launch time), the replay may not be bit-exact — pin the exact commit via the repo.commit field and git checkout it first.