Upload scripts/t32_train.sh with huggingface_hub

scripts/t32_train.sh

@@ -0,0 +1,100 @@
+#!/usr/bin/env bash
+# =============================================================================
+# 32k training at the CORRECTED scale -- the cell that has never been run.
+#
+# WHY SCALE=32 AND NOT 8:
+#   The invariant SCALE*ORIG==CP_T would force SCALE=8 at T=32,768, putting the
+#   far end at virtual 4,096 -- the badly-conditioned edge, and exactly the
+#   phase-1 mistake that made 1,200 steps look worthless.
+#   YARNOps.cpp treats seq_len != scale*orig_maxpos as a WARNING, not an error
+#   ("(R2) ... warn, don't hard-fail"), so CP_T=32,768 with SCALE=32 builds a
+#   cache SIZED 32,768 but SCALED for 131,072: positions land at virtual
+#   0..1,023. The angle at a position is (pos/scale)*omega, independent of cache
+#   LENGTH, so this MATCHES the inference config we will evaluate at:
+#       train  CP_T=32,768 x32   -> virtual 0..1,023
+#       infer  65,512 tok   x64  -> virtual 0..1,023
+#   The R2 warning is EXPECTED here and is not a misconfiguration.
+#
+# PARENT: p1final, not base680. At x64 p1final beats base680 by +8/+11/+7pp,
+#   so it is the better starting point despite having been trained at x4.
+#
+# MEMORY: the 3-step probe peaked at 77,397 of 81,559 MB (95%). Tight. Saves
+#   every 200 steps so an OOM late in the run does not lose everything.
+# =============================================================================
+set -uo pipefail
+cd "$(dirname "$0")"
+ROOT="$(cd ../.. && pwd)"
+export LD_LIBRARY_PATH="$ROOT/Tensor-Implementations/lib:$ROOT/Profiler/lib:${LD_LIBRARY_PATH:-}"
+export OMPI_ALLOW_RUN_AS_ROOT=1 OMPI_ALLOW_RUN_AS_ROOT_CONFIRM=1
+export PATH=/usr/local/cuda/bin:$PATH
+PY=/root/Adhi/BluTrain/envv/bin/python
+EXE=./build/bluscriptCP_exec
+LOG=t32_logs; mkdir -p "$LOG"; M="$LOG/master.log"
+T=32768; NP=2; CPSZ=2; SCALE=32; ORIG=4096
+STEPS=${STEPS:-1200}; WARM=100; TAIL=2; FREQ=200; KEEP=3
+STABLE=$(( STEPS - WARM - TAIL ))
+GB=524288
+DATA=Data_Loader/Data/t32mix
+CK=checkpoints_bluscriptcp/rank_0
+PARENT="$CK/blumodelcp_run12_step_1200.ckpt"
+say(){ echo "[$(date -u +%H:%M:%S)] $*" | tee -a "$M" >&2; }
+die(){ say "FATAL: $*"; exit 1; }
+
+[ -x "$EXE" ] || die "$EXE missing"
+S=build/objects/.cp_rope_flags
+grep -q 'CP_FUSED_ROPE=1' "$S" && grep -q 'CP_ATTN_SPLIT=1' "$S" || die "binary lacks split kernels"
+[ -f "$PARENT" ] || die "parent p1final missing"
+[ "$(stat -c%s "$PARENT")" = 5141244514 ] || die "parent wrong size"
+[ -d "$DATA" ] || die "$DATA missing"
+U=$($PY - "$DATA" "$T" "$GB" <<'PY'
+import glob,os,sys
+d,BT,GBt=sys.argv[1],int(sys.argv[2]),int(sys.argv[3]); GA=GBt//BT; t=0
+for p in sorted(glob.glob(f"{d}/train_shard_*.bin")):
+    n=os.path.getsize(p)//2; pos=0; b=0
+    while pos+BT+1<=n:
+        b+=1; pos+=BT
+        if pos+(BT+1)>n: break
+    t+=b
+print(t//GA)
+PY
+)
+[ "$STEPS" -le "$U" ] || die "STEPS=$STEPS > usable=$U (loader would WRAP silently)"
+[ $(( WARM + STABLE )) -lt $(( STEPS - 1 )) ] || die "illegal WSD"
+
+ARCH="CP_N_EMBD=1536 CP_N_LAYER=24 CP_N_HEAD=12 CP_N_KVHEAD=4 CP_FFN=4096 CP_WEIGHT_TYING=1"
+BASE="CP_SIZE=$CPSZ CP_ATTN_MODE=ring CP_ATTN_FUSION=split CP_T=$T \
+YARN_SCALE=$SCALE YARN_ORIG_MAXPOS=$ORIG CP_ROPE_THETA=500000 \
+CP_ZERO=1 CP_ZERO_V2=1 CP_BF16=1 CP_OPTIMIZER=muon \
+CP_WARMUP=$WARM CP_REWARMUP=$WARM CP_REWARMUP_PEAK=0.3 \
+CP_STABLE=$STABLE CP_CKPT=1 CP_CKPT_FREQ=$FREQ CP_CKPT_KEEP=$KEEP"
+
+say "=== 680M 32k training at the CORRECTED scale ==="
+say "parent : p1final (run12 step 1200)"
+say "data   : $DATA  b4=50 b3=30 b1=20, usable $U steps, running $STEPS"
+say "YaRN   : CP_T=$T SCALE=$SCALE ORIG=$ORIG -> virtual 0..$(( (T-1)/SCALE ))"
+say "         (R2 seq_len != scale*orig warning is EXPECTED and intended)"
+say "mem    : probe peaked 77,397/81,559 MB -- 95%, tight"
+
+if [ ! -f "$LOG/probe.ok" ]; then
+  say "PROBE: 3 steps at the EXACT training config (grad_accum $((GB/T)))"
+  env $ARCH $BASE CP_STABLE=0 CP_CKPT=0 CP_INIT_FROM="$PWD/$PARENT" \
+      CP_DATA_ROOT="$PWD/$DATA" CP_B=1 CP_GLOBAL_BATCH=$GB CP_MAX_STEPS=3 \
+      CUDA_VISIBLE_DEVICES=0,1 mpirun --allow-run-as-root -np $NP "$EXE" > "$LOG/probe.log" 2>&1
+  RC=$?
+  grep -qaiE 'out of memory|bad_alloc' "$LOG/probe.log" && die "PROBE OOM at the real batch -- 32k will not fit"
+  N=$(grep -acE '^step +[0-9]+ \|' "$LOG/probe.log"); P=$(grep -aoE 'mem: [0-9]+MB' "$LOG/probe.log"|tail -1)
+  [ "$RC" -eq 0 ] && [ "$N" -ge 1 ] && [ -n "$P" ] || { say "PROBE FAILED rc=$RC steps=$N"; tail -12 "$LOG/probe.log" >&2; exit 1; }
+  say "PROBE OK: $N steps, peak $P, dt $(grep -aoE 'dt: [0-9.]+ms' "$LOG/probe.log"|tail -1)"
+  touch "$LOG/probe.ok"
+fi
+
+say "TRAINING $STEPS steps"
+env $ARCH $BASE CP_INIT_FROM="$PWD/$PARENT" CP_CKPT_NEW_RUN=1 CP_DATA_ROOT="$PWD/$DATA" \
+    CP_B=1 CP_GLOBAL_BATCH=$GB CP_MAX_STEPS=$STEPS \
+    CUDA_VISIBLE_DEVICES=0,1 mpirun --allow-run-as-root -np $NP "$EXE" >> "$LOG/train.log" 2>&1
+grep -qaiE 'out of memory|bad_alloc' "$LOG/train.log" && die "OOM mid-run -- checkpoints up to the last save survive"
+RN=$(grep -aoE 'CP_Training_log[0-9]+' "$LOG/train.log" | head -1 | grep -oE '[0-9]+')
+[ -n "$RN" ] || die "no run number"; echo "$RN" > "$LOG/runnum"
+say "done, run $RN, $(grep -aE '^step ' "$LOG/train.log"|tail -1|grep -oE 'loss: [0-9.]+')"
+say "peak mem $(grep -aoE 'mem: [0-9]+MB' "$LOG/train.log"|tail -1)"
+say "=== 32K TRAINING COMPLETE ==="