feat: optional multi-GPU (DDP) finetuning

[evasnow1992]

Summary

Adds opt-in data-parallel (DistributedDataParallel) finetuning so finetunes can be accelerated across multiple GPUs, mirroring the existing pretraining DDP path. Single-GPU finetuning is unchanged and remains the default — DDP only
engages through a dedicated launcher.

What's new

Core (feat(finetune): add optional multi-GPU DDP training)

• finetune_ddp.py: a one-process-per-GPU launcher (torch.multiprocessing spawn, WORLD_SIZE env, NCCL topology auto-config, per-rank determinism), routed through the normal cross_validate -> run_training finetune path.
• kermt/util/ddp_utils.py: shared ddp_setup + configure_nccl_for_topology so the pretrain and finetune launchers use one DDP bootstrap.
• task/train.py (run_training/train): thread rank/world_size; wrap the model in DDP (find_unused_parameters=True, since the finetune model carries parameters the FFN forward does not exercise); shard the train loader with a DistributedSampler (+ set_epoch); run validation on the unwrapped module over the full val set on every rank so best-model selection stays consistent; gate checkpoint saving, logging and test-set evaluation to rank 0; all-reduce the standalone MTLLoss.log_sigma gradient. Also uses a distinct train_loader instead of reassigning train_data, fixing an ensemble_size>1 DataLoader double-wrap.
• cross_validate: thread rank/world_size; gate W&B + reporting to rank 0.
• build_lr_scheduler: world-size-aware steps_per_epoch.

Skill (feat(kermt-finetune skill): expose optional multi-GPU DDP finetune)

• agent/scripts/run_finetune_local.py: new --num-gpus N option. Default 1 (single-process main.py finetune, unchanged); N>1 launches finetune_ddp.py with WORLD_SIZE=N. Under DDP it omits the single-device --gpu flag and records num_gpus/distributed in run.json.
• agent/skills/kermt-finetune/SKILL.md: documents --num-gpus in the hardware, arguments, invocation-template, and error-hint sections.

Design

• Opt-in, zero-impact default. DDP activates only when a process group is initialized (i.e. via finetune_ddp.py). python main.py finetune ... and --num-gpus 1 behave exactly as before.
• Per-GPU batch size. --batch_size is per process; effective global batch is batch_size * WORLD_SIZE. The LR schedule is world-size-aware, so the same --batch_size can be reused as GPUs are added.
• Rank 0 owns side effects. Logging, checkpointing, and test-set evaluation run on rank 0; saved checkpoints are ordinary (non-DDP) state dicts, portable to single-GPU inference/finetune.

Usage

# data-parallel finetune across all visible GPUs (or set WORLD_SIZE=N)
python finetune_ddp.py --data_path train.csv --separate_val_path val.csv \
    --separate_test_path test.csv --checkpoint_path <ckpt> --save_dir <dir> \
    --dataset_type regression --metric mae ...   # usual finetune flags, no subcommand

# via the finetune skill runner
python agent/scripts/run_finetune_local.py --ckpt <ckpt> \
    --prepare-manifest <prepare_data.json> --dataset-type regression \
    --out <run-dir> --num-gpus 4

Testing

• Single-process finetune (main.py finetune) runs unchanged end-to-end (2-epoch regression finetune from a pretrained checkpoint): trains, selects best by validation, saves checkpoints, evaluates on test.
• DDP finetune path runs end-to-end (finetune_ddp.py, WORLD_SIZE=1): full DDP wrap, DistributedSampler + set_epoch, unwrapped-module eval, rank-0 checkpoint/test — completes with comparable validation/test metrics to the single-process run.
• Skill runner: --num-gpus 1 dispatches to main.py finetune; --num-gpus N (N>1) dispatches to finetune_ddp.py with WORLD_SIZE=N (verified by dry-run command inspection and an actual launch). Existing finetune-runner unit tests (19) pass.
• Recommended before production reliance: a multi-GPU (2+) run to validate cross-rank gradient all-reduce and wall-clock scaling, which single-GPU CI cannot exercise.

[sveccham]

I wonder if we should have only 1 entrypoint for finetuning. Currently, we have two main.py finetune ... and finetune_ddp.py .... While they share most of the underlying code, it might be better to have 1 entrypoint.

[sveccham]

Thanks for the DDP implementation. Overall, useful useful feature to have. I have left comments in-line.

[sveccham]

There is a similar argument --gpus elsewhere. I think that one is mostly ignored. If so, we can remove that arg.

[sveccham]

Why do we need find_unused_parameters=True. I think there might be a a situation where we are not using some views from prertaining. It would be good to delete those. I am scared that this param would suppress some genuine bugs when params are not being trained correctly.