Upload scripts/abl_phase1.sh with huggingface_hub

scripts/abl_phase1.sh

@@ -30,6 +30,13 @@ EXE=./build/bluscriptCP_exec
 
 ARM=${ARM:?set ARM=const or ARM=cosine}
 GPU=${GPU:?set GPU=0 or 1}
+# OPT=muon (default, reproduces the 2026-09-18 arms) or OPT=adamw.
+# adamw is the CONTROL that separates the two variables yesterday's run confounded:
+#   p1final    1.8e-4 WSD  + AdamW
+#   const/muon 8e-5 const  + Muon      <- differs from p1final in BOTH
+#   const/adamw 8e-5 const + AdamW     <- differs from each in exactly ONE
+OPT=${OPT:-muon}
+case "$OPT" in muon|adamw) ;; *) echo "OPT must be muon or adamw"; exit 1 ;; esac
 T=16384; SCALE=4; ORIG=4096
 STEPS=${STEPS:-1200}
 WARM=$(( STEPS / 100 ))                 # 1% of this phase's steps
@@ -43,7 +50,7 @@ GB=524288
 DATA=Data_Loader/Data/ablmix16
 CK=checkpoints_bluscriptcp/rank_0
 PARENT="$CK/varybatch.ckpt"
-LOG=p1_${ARM}_logs; mkdir -p "$LOG"; M="$LOG/master.log"
+LOG=p1_${ARM}_${OPT}_logs; mkdir -p "$LOG"; M="$LOG/master.log"
 say(){ echo "[$(date -u +%H:%M:%S)] [$ARM] $*" | tee -a "$M" >&2; }
 die(){ say "FATAL: $*"; exit 1; }
 
@@ -56,18 +63,19 @@ grep -q 'CP_FUSED_ROPE=1' "$S" && grep -q 'CP_ATTN_SPLIT=1' "$S" || die "binary
 [ "$STABLE" -eq 0 ] || [ $(( WARM + STABLE )) -lt $(( STEPS - 1 )) ] || die "WSD guard would fire"
 $PY -c "import tiktoken" 2>/dev/null || die "tiktoken missing"
 
+OPTENV=""; [ "$OPT" = muon ] && OPTENV="CP_OPTIMIZER=muon"
 ARCH="CP_N_EMBD=1536 CP_N_LAYER=24 CP_N_HEAD=12 CP_N_KVHEAD=4 CP_FFN=4096 CP_WEIGHT_TYING=1"
 BASE="CP_SIZE=1 CP_ATTN_MODE=ring CP_ATTN_FUSION=split CP_T=$T \
 YARN_SCALE=$SCALE YARN_ORIG_MAXPOS=$ORIG CP_ROPE_THETA=500000 \
 CP_ZERO=1 CP_ZERO_V2=1 CP_BF16=1 \
-CP_OPTIMIZER=muon CP_MAX_LR=$MAXLR CP_MIN_LR=$MINLR \
+${OPTENV} CP_MAX_LR=$MAXLR CP_MIN_LR=$MINLR \
 CP_WARMUP=$WARM CP_REWARMUP=$WARM CP_REWARMUP_PEAK=1.0 \
 CP_STABLE=$STABLE CP_CKPT=1 CP_CKPT_FREQ=200 CP_CKPT_KEEP=3 \
-CP_CKPT_DIR=checkpoints_${ARM}"
+CP_CKPT_DIR=checkpoints_${ARM}_${OPT}"
 
-say "=== phase 1: 4k -> 16k, arm=$ARM on GPU $GPU ==="
+say "=== phase 1: 4k -> 16k, arm=$ARM opt=$OPT on GPU $GPU ==="
 say "LR      : max $MAXLR  min $MINLR  warmup $WARM (1% of $STEPS)  stable $STABLE"
-say "optim   : Muon (>=2-D matrices) + aux AdamW, both at the SAME lr"
+say "optim   : $([ "$OPT" = muon ] && echo "Muon (>=2-D matrices) + aux AdamW, same lr" || echo "AdamW only (ZeRO-1 v2)")"
 say "parent  : varybatch   data: $DATA   steps: $STEPS"
 
 if [ ! -f "$LOG/probe.ok" ]; then
@@ -82,7 +90,12 @@ if [ ! -f "$LOG/probe.ok" ]; then
   [ "$RC" -eq 0 ] && [ "$N" -ge 1 ] || { say "PROBE FAILED rc=$RC steps=$N"; tail -14 "$LOG/probe.log" >&2; exit 1; }
   # ASSERT the trainer actually took the new env vars -- all three were no-ops
   # until today, and a silent fallback would look like a healthy run.
-  grep -qai 'muon' "$LOG/probe.log" || die "no Muon in probe log -- CP_OPTIMIZER did not take"
+  # Assert the CONSTRUCTED optimizer, in both directions. The [ZeRO-1] line
+  # used to print configured state, so a Muon run and an AdamW run were
+  # indistinguishable in the log -- which is how CP_OPTIMIZER stayed a no-op
+  # across every run before 2026-09-18.
+  WANT=$([ "$OPT" = muon ] && echo "optimizer=MUON+adamw-aux" || echo "optimizer=adamw")
+  grep -qa "$WANT" "$LOG/probe.log" || die "probe built the WRONG optimizer -- wanted '$WANT', log says '$(grep -ao "optimizer=[A-Za-z+-]*" "$LOG/probe.log" | head -1)'"
   LR0=$(grep -aoE '^step +0 \|.*lr [0-9.e+-]+' "$LOG/probe.log" | grep -oE 'lr [0-9.e+-]+' | awk '{print $2}')
   say "PROBE OK: $N steps, peak $P, step-0 lr=$LR0 (expect ~8.0e-05 at warmup 1/$WARM)"
   touch "$LOG/probe.ok"