Allow checkpoint resume across shorter context configs

Pass eval and curriculum controls to RunPod pods

Expose eval and curriculum controls in RunPod trainer

Ensure at least one masked token per sample

Add eval, sharding, and curriculum controls

Pass pretrain shuffle controls to RunPod pods

Expose pretrain shuffle controls in RunPod trainer

Add buffered shuffle to pretrain streams

Pass TensorBoard settings to RunPod pods

Launch TensorBoard in RunPod trainer

Add TensorBoard training metrics

Install git-lfs when missing and skip LFS smudge

Propagate tokenizer setting to RunPod pods

Use SmolLM tokenizer by default and validate vocab

Pass tokenizer explicitly in RunPod training script

Add idle-start debug mode for RunPod

Enable global networking for RunPod pods

Remove HF token literal from RunPod bootstrap

Require explicit RunPod GPU type

Fix RunPod GPU default

Sync RunPod launcher and 2B trainer

Sync 2B training job files

Multi-GPU via Local SGD: split batch across GPUs, average params each step

Remove DataParallel (incompatible with unified mode in-place updates)

Unified PC training: fix energy normalization, add task signal to all blocks

ric

PC-SHO-DLM: full architecture with MSA integration

Upload src/msa.py with huggingface_hub

Upload train_2b.py with huggingface_hub

Upload src/model.py with huggingface_hub

Upload checkpoints/pretrain_step5000.pt with huggingface_hub

Upload checkpoints/pretrain_step4000.pt with huggingface_hub

Upload checkpoints/pretrain_step3000.pt with huggingface_hub

Upload checkpoints/pretrain_step2000.pt with huggingface_hub

Upload checkpoints/pretrain_step1000.pt with huggingface_hub

initial commit