From 488c50e1bdbf8f420b2ad5b4021a7f950d121835 Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Mon, 6 Jul 2026 07:15:30 -0500 Subject: --holoavg shipped (gate 0.913->0.936 @t2sel160, batch2 +0.051) + farm wave-2 tooling trend-aware stop + plateau averaging recovers the semi-convergence victims exactly as predicted. Estimator pack now: holofast+sdpa+t2sel80(+holoavg). Also: bp_lm stdinit/beta2/sched flags (anchor archaeology), dipfarm_freezer, staged tol_sweep.sh (gated on hr2 verdict), bp_sweep.sh. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn --- ep_run/runs/bp_sweep.sh | 8 ++++++++ ep_run/runs/tol_sweep.sh | 24 ++++++++++++++++++++++++ 2 files changed, 32 insertions(+) create mode 100755 ep_run/runs/bp_sweep.sh create mode 100644 ep_run/runs/tol_sweep.sh (limited to 'ep_run/runs') diff --git a/ep_run/runs/bp_sweep.sh b/ep_run/runs/bp_sweep.sh new file mode 100755 index 0000000..e98c79c --- /dev/null +++ b/ep_run/runs/bp_sweep.sh @@ -0,0 +1,8 @@ +#!/bin/bash +cd /home/yurenh2/ept/ep_run +for cfg in "qk3 --qknorm --lr 3e-3" "qk1 --qknorm --lr 1e-3" "nq3 --lr 3e-3" "nq1 --lr 1e-3"; do + set -- $cfg; name=$1; shift + CUDA_VISIBLE_DEVICES=0 python3 bp_lm.py "$@" --steps 32000 --log 1000 --ckpt runs/bp_$name.pt > runs/bp_$name.log 2>&1 + grep "DONE" runs/bp_$name.log +done +echo BP_SWEEP_ALL_DONE diff --git a/ep_run/runs/tol_sweep.sh b/ep_run/runs/tol_sweep.sh new file mode 100644 index 0000000..0536f59 --- /dev/null +++ b/ep_run/runs/tol_sweep.sh @@ -0,0 +1,24 @@ +#!/bin/bash +# TOLERANCE SWEEP (hardware gate, task #15/#16): warm-start s2000, 1500 steps per config. +# GATED on the hr2 verdict (EP+regs from a TRAINED state must be Pascal-safe). 2 GPUs x 5 configs. +# Grid: device noise fnoise x weight-quantization wq_bits (core cells; wmis later). +cd /home/yurenh2/ept/ep_run +PY=/home/yurenh2/miniconda3/envs/ep_pascal/bin/python3 +BASE="--mode ep --attn_mode thick --B 24 --C 512 --H 16 --T 256 --c 1.0 --jacreg 0.1 --jr_floor 0.1 --jr_max 0.1 --resreg 0.2 --holo 2 --hr 0.02 --t2sel 40 --track --pema 0.999 --t1max 300 --res_est 1e-4 --res_gate 0 --qknorm --init_ckpt runs/redx_traj/s2000.pt --warmup 50 --T1 150 --T2 20 --lr 6e-4 --wsd 0.25 --steps 1500 --log 100 --save_every 1500 --abort_res 0.3 --data data/tinystories_bpe" +run() { local g=$1 n=$2; shift 2; env CUDA_VISIBLE_DEVICES=$g $PY lt_ep_train.py $BASE "$@" --ckpt runs/tol_$n.pt --state runs/tol_$n.state > runs/tol_$n.log 2>&1; grep -E "DONE|abort" runs/tol_$n.log | tail -1; } +( + run 6 base + run 6 fn1e3 --fnoise 1e-3 + run 6 fn3e3 --fnoise 3e-3 + run 6 fn1e2 --fnoise 1e-2 + run 6 fn3e2 --fnoise 3e-2 +) & +( + run 7 wq8 --wq_bits 8 + run 7 wq6 --wq_bits 6 + run 7 wq4 --wq_bits 4 + run 7 fn3e3wq8 --fnoise 3e-3 --wq_bits 8 + run 7 wm1e2 --wmis 1e-2 +) & +wait +echo TOL_SWEEP_ALL_DONE -- cgit v1.2.3