Smoke SFT: grad-accum=1 batch + batch-scaled LRs (verified)

hf-jobs/launch.sh

@@ -59,9 +59,15 @@ case "$STAGE" in
             --device-batch-size=16 --total-batch-size=16384 \
             --num-iterations=50 --eval-every=25 --eval-tokens=524288 \
             --core-metric-every=-1 --sample-every=-1 --run=dummy
+        # SFT: keep grad-accum=1 (total = device_batch*seq_len) so --num-iterations counts
+        # optimizer steps 1:1 (chat_sft's progress tracking is per micro-batch), and scale
+        # LRs to the batch like base_train does (sqrt(8192/524288) ~= 0.125 x defaults).
+        # Note: train/loss may log NaN due to fully-masked micro-batches (harmless, zero
+        # gradient) — judge SFT by val/bpb.
         python -m scripts.chat_sft \
-            --num-iterations=30 --total-batch-size=32768 \
-            --eval-every=15 --eval-tokens=524288 \
+            --num-iterations=20 --total-batch-size=8192 \
+            --embedding-lr=0.0375 --unembedding-lr=0.0005 --matrix-lr=0.0025 \
+            --eval-every=10 --eval-tokens=524288 \
             --chatcore-every=-1 --run=dummy
         python -m scripts.chat_cli -p "Why is the sky blue?"
         ;;