#!/bin/bash
# Feather prod9 autonomous launcher — no local cache, mid_val B=1, skip final eval on 6GB
set -euo pipefail
cd /home/mikeb/work/feather
HF=$(grep -ohP 'hf_[A-Za-z0-9_-]+' ~/.bashrc 2>/dev/null | head -1 || true)
pkill -9 -f "python.*train\.py" 2>/dev/null || true
sleep 1
rm -f /home/mikeb/.cache/autoresearch/packed_tokens_v1_T1024_V65536_train.bin*

export LD_LIBRARY_PATH=/usr/lib/wsl/lib:/usr/local/cuda/lib64
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export HF_TOKEN="$HF"
export HUGGINGFACE_HUB_TOKEN="$HF"
export WANDB_DISABLED=true
export HYDRA_USE_NEMOTRON=1
export HYDRA_USE_FULL_BLEND=1
export HYDRA_SAMPLED_SOFTMAX=1024
export HYDRA_SOFTCAP_CLAMP=1
export HYDRA_SEQ_LEN=1024
export HYDRA_HEADDIM=32
export HYDRA_D_STATE=64
export HYDRA_TIME_BUDGET=300
export HYDRA_ENGRAM_TOPK=64
export HYDRA_GDN_LAYERS=
export HYDRA_MTP_K=1
export HYDRA_USE_MDLM=0
export HYDRA_MUON_COMPILE=0
export HYDRA_MUON_NS_STEPS=2
# Generalization-recovery recipe: resume from best checkpoint, cool LR,
# increase regularization. Current latest overfits train BPB while val worsens.
export HYDRA_RESUME_CKPT=/home/mikeb/.cache/autoresearch/best_bpb.pt
export HYDRA_MATRIX_LR=0.004
export HYDRA_EMBED_LR=0.08
export HYDRA_UNEMBED_LR=0.0005
export HYDRA_DT_BIAS_LR=0.02
export HYDRA_SCALAR_LR=0.004
export HYDRA_WEIGHT_DECAY=0.03
export HYDRA_DROPOUT=0.30
export HYDRA_LABEL_SMOOTHING=0.05
export HYDRA_Z_LOSS_WEIGHT=0.0005
export HYDRA_WARMUP_RATIO=0.02
export HYDRA_LR_MIN_MULT=0.25
export HYDRA_WARMSTART=1
export HYDRA_STREAM_SHUFFLE_BUFFER=4096
export HYDRA_LOCAL_SHARDS_ONLY=0
export HYDRA_BACKGROUND_PREFETCH=0
export HYDRA_STREAM_PREFETCH=16
export HYDRA_TOKEN_PREFETCH=4
export HYDRA_TOKEN_CACHE_GB=4
export HYDRA_CKPT_INTERVAL=2000
export HYDRA_MID_VAL_INTERVAL=250
export HYDRA_MID_VAL_BATCH=1
export HYDRA_MID_VAL_TOKENS=51200
export HYDRA_EVAL_BATCH=1
export HYDRA_CKPT_ROTATIONS=3
export HYDRA_SKIP_FACTUAL_EVAL=1
export HYDRA_FORCE_OS_EXIT=1
export HYDRA_N_LAYER=6
export HYDRA_D_MODEL=192
export HYDRA_EXPAND=3
export HYDRA_BATCH_SIZE=16
export HYDRA_TOTAL_BATCH=32768
export HYDRA_HESTIA_INTERVAL=999999
export HYDRA_HTM_SUBSAMPLE=16
export UV_PYTHON=/usr/bin/python3

setsid -f taskset -c 0-15 ./.venv/bin/python -u train.py </dev/null >>run_3060_prod9.log 2>&1 &
TPID=$!
echo "Launched PID=$TPID"
sleep 2
pgrep -n -f 'python.*train\.py' && echo "Training running" || echo "WARNING: no process"