X-WING Brown: shared tables + per-order entropy gating, no cubric

PR openai#798's approach on our engine: per-order entropy centers
(7:3.0, 6:3.2, 5:3.5, 4:3.8, 3:4.2, 2:4.5) without cubric.
Testing if cubric was hurting when combined with per-order gating.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: Octavian

concepts/xwing_brown/run.sh

@@ -0,0 +1,49 @@
+#!/bin/bash
+set -euo pipefail
+# X-WING BROWN: shared tables + per-order entropy gating (no cubric)
+# PR#798 per-order entropy centers on our podracer engine + shared tables
+
+SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd -- "${SCRIPT_DIR}/../.." && pwd)"
+cd "${REPO_ROOT}"
+export PYTHONPATH="${REPO_ROOT}/flash-attention/hopper:${PYTHONPATH:-}"
+
+SEED="${SEED:-2045}"
+NPROC_PER_NODE="${NPROC_PER_NODE:-8}"
+
+echo "============================================"
+echo "  X-WING BROWN (shared tables + per-order entropy gating)"
+echo "  Seed: ${SEED}"
+echo "  No cubric — per-order gating only"
+echo "  Chunk tokens: ${NGRAM_CHUNK_TOKENS:-1048576}"
+echo "============================================"
+
+SEED="$SEED" \
+F1_CORR_RANK=0 \
+DISTILL_ENABLED=0 \
+MLP_ACT=leaky_relu_sq \
+MLP_LEAKY_SLOPE=0.5 \
+XSA_LAST_N=4 \
+BIGRAM_VOCAB_SIZE=1536 \
+TTT_EVAL_ENABLED=0 \
+ROPE_DIMS=24 \
+NGRAM_EVAL_ORDER=7 \
+NGRAM_EVAL_MIN_ORDER=2 \
+NGRAM_EVAL_ADAPTIVE=1 \
+NGRAM_EVAL_ALPHA=0.30 \
+NGRAM_EVAL_ALPHA_MIN=0.05 \
+NGRAM_EVAL_ALPHA_MAX=0.70 \
+NGRAM_EVAL_ENTROPY_CENTER=3.0 \
+NGRAM_EVAL_ENTROPY_SCALE=2.0 \
+NGRAM_EVAL_MIN_COUNT=2 \
+NGRAM_EVAL_BUCKETS=8388608 \
+NGRAM_EVAL_MAX_SECONDS=300 \
+CUBRIC_CADENCE=0 \
+COMPILE_FULLGRAPH=0 \
+torchrun --standalone --nproc_per_node="${NPROC_PER_NODE}" \
+    "${SCRIPT_DIR}/train_gpt.py" \
+    2>&1 | tee "logs/xwing_brown_s${SEED}_$(date +%Y%m%d_%H%M%S).log"
+
+echo "============================================"
+echo "  DONE"
+echo "============================================"