Upload scripts/p1_eval32k.sh with huggingface_hub

scripts/p1_eval32k.sh

@@ -0,0 +1,76 @@
+#!/usr/bin/env bash
+# =============================================================================
+# The question: does the 680M's advantage survive at 32k, and does phase-1
+# training add anything there?
+#
+# Evaluates BOTH models -- base680 (untrained, zero-shot YaRN) and p1final
+# (1200 steps) -- at 8,192 / 16,360 / 32,744 under a T=32,768 window.
+#
+# base680 is the control. Phase 1 is the run that taught us why: without the
+# untrained baseline, 77 and 42 look like results instead of noise.
+#
+# YaRN: SCALE=8 x ORIG_MAXPOS=4096 = 32,768. ORIG stays at the model's ORIGINAL
+# pretrain length (4096) and only SCALE rises -- build_rope_cache is rebuilt from
+# base theta every run so SCALE never composes, and find_dim() does not take
+# scale, so holding ORIG fixed keeps the interpolated dims at [14,32] exactly as
+# at 16k. Re-basing to 2 x 16384 would satisfy SCALE*ORIG==CP_T but freeze
+# interpolation and push the far end to virtual position 16,384.
+#
+# CAVEAT ON COMPARING TO PHASE 1: phase 1 measured 84 @8192 / 40 @16360 under a
+# T=16,384 cache (SCALE=4). These runs use a T=32,768 cache (SCALE=8), so the
+# 8k/16k numbers here are NOT directly comparable to those -- the cache changed,
+# not just the prompt length. What IS valid is base680 vs p1final within this
+# run, which is the actual question.
+# =============================================================================
+set -uo pipefail
+cd "$(dirname "$0")"
+export PATH=/root/Adhi/BluTrain/envv/bin:$PATH
+source /root/Adhi/BluTrain/envv/bin/activate 2>/dev/null
+L=e32k_logs; mkdir -p "$L"; M="$L/master.log"
+say(){ echo "[$(date -u +%H:%M:%S)] $*" | tee -a "$M" >&2; }
+die(){ say "FATAL: $*"; exit 1; }
+
+EXE=./build/bluscriptCP_exec
+HAY=Data_Loader/Data/b2_8192_16384/val_shard_00000.bin
+CK=checkpoints_bluscriptcp/rank_0
+T=32768; ORIG=4096; SCALE=8
+[ $((SCALE*ORIG)) -eq $T ] || die "YaRN invariant: $SCALE*$ORIG != $T"
+[ -x "$EXE" ] || die "$EXE not built"
+S=build/objects/.cp_rope_flags
+grep -q 'CP_FUSED_ROPE=1' "$S" && grep -q 'CP_ATTN_SPLIT=1' "$S" || die "binary lacks split kernels ($(cat $S))"
+[ -f "$HAY" ] || die "haystack missing"
+[ "$(stat -c%s "$HAY")" = 199754694 ] || die "haystack is NOT the file phase 1 used (size $(stat -c%s "$HAY"))"
+[ -f "$CK/blumodelcp_run12_step_1200.ckpt" ] || die "p1final missing"
+[ -f "$CK/varybatch.ckpt" ] || ln -sf "$PWD/checkpoints_680m/rank_0/varybatch.ckpt" "$CK/varybatch.ckpt"
+[ -s "$CK/varybatch.ckpt" ] || die "varybatch missing"
+ls Data_Loader/Data/evalsrc/train_shard_*.bin >/dev/null 2>&1 || die "no train split for the loader"
+python -c "import tiktoken" 2>/dev/null || die "tiktoken missing"
+
+A1B_680="CP_N_EMBD=1536 CP_N_LAYER=24 CP_N_HEAD=12 CP_N_KVHEAD=4 CP_FFN=4096 CP_WEIGHT_TYING=1"
+say "=== 680M @ 32k window: base680 (control) vs p1final ==="
+say "YaRN SCALE=$SCALE ORIG=$ORIG -> $((SCALE*ORIG)) == CP_T=$T"
+say "haystack $(stat -c%s "$HAY") B -- same file phase 1 used"
+
+for PL in 8192 16360 32744; do
+  O="e32k_len${PL}"
+  if [ -s "$O/summary.csv" ] && [ "$(wc -l < "$O/summary.csv")" -gt 1 ]; then
+    say "len $PL already done: $(grep '^niah_single,' "$O/summary.csv")"; continue; fi
+  say "eval prompt_len=$PL (2 GPUs, 2 jobs)"
+  A1B="$A1B_680" HAY="$HAY" GEN_DATA_ROOT="Data_Loader/Data/evalsrc" \
+  CKPT_EXTRA="p1final=blumodelcp_run12_step_1200.ckpt base680=varybatch.ckpt" \
+  CKPTS="p1final base680" TASKS=niah_single N=100 \
+  T=$T YARN=$SCALE YARN_ORIG=$ORIG PROMPT_LEN=$PL GPUS=0,1 NP=1 OUT="$O" \
+    ./run_longeval.sh >> "$L/eval_${PL}.log" 2>&1
+  if [ -s "$O/summary.csv" ] && [ "$(wc -l < "$O/summary.csv")" -gt 1 ]; then
+    say "  len $PL -> $(head -1 $O/summary.csv) | $(grep '^niah_single,' $O/summary.csv)"
+  else
+    say "  len $PL FAILED"; grep -ahA2 'what()' "$O"/gen/*.log 2>/dev/null | head -4 | tee -a "$M" >&2
+  fi
+done
+
+say "=== RESULTS (niah_single, N=100, T=$T, YaRN ${SCALE}x${ORIG}) ==="
+for PL in 8192 16360 32744; do
+  f="e32k_len${PL}/summary.csv"
+  [ -f "$f" ] && say "  len $PL  $(head -1 $f)  ->  $(grep '^niah_single,' $f)"
+done
+say "DONE"