fix(job_train): force CUDA_VISIBLE_DEVICES=0 to dodge TRL multi-GPU entropy bug

scripts/job_train.sh

@@ -5,6 +5,13 @@ set -e
 cd /w
 export PYTHONPATH="/w:${PYTHONPATH}"
 
+# Force single-GPU training. The 0.5B (and 1.5B) 4-bit + LoRA model fits on
+# one A10G easily, and TRL's SFTTrainer entropy logging path on multi-GPU
+# (a10g-largex2) crashes with `RuntimeError: tensor a (2) must match tensor
+# b (4) at non-singleton dimension 0` because the per-rank logits get mixed
+# with the full-batch attention mask. Hide the second GPU completely.
+export CUDA_VISIBLE_DEVICES=0
+
 # Install huggingface_hub first so we can download data files (parity with job_smoke.sh)
 pip install -q 'huggingface_hub>=0.28'