diff --git a/examples/mimo/scripts/run_hetero_nemotron_54l_hel_train.sh b/examples/mimo/scripts/run_hetero_nemotron_54l_hel_train.sh
index 3d206961c43..70828a8c10b 100755
--- a/examples/mimo/scripts/run_hetero_nemotron_54l_hel_train.sh
+++ b/examples/mimo/scripts/run_hetero_nemotron_54l_hel_train.sh
@@ -96,7 +96,7 @@ LR_WSD_DECAY_SAMPLES="${LR_WSD_DECAY_SAMPLES:-}"
 LR_WSD_DECAY_STYLE="${LR_WSD_DECAY_STYLE:-}"
 TRAIN_SAMPLES="${TRAIN_SAMPLES:-}"
 PACKING_BUFFER_SIZE="${PACKING_BUFFER_SIZE:-128}"
-NUM_WORKERS="${NUM_WORKERS:-1}"
+NUM_WORKERS="${NUM_WORKERS:-2}"
 SHUFFLE_BUFFER_SIZE="${SHUFFLE_BUFFER_SIZE:-100}"
 MAX_SAMPLES_PER_SEQUENCE="${MAX_SAMPLES_PER_SEQUENCE:-100}"
 LOG_INTERVAL="${LOG_INTERVAL:-10}"
diff --git a/examples/mimo/scripts/sbatch_hetero_prod_gbs768_100n.sh b/examples/mimo/scripts/sbatch_hetero_prod_gbs768_100n.sh
index ff6d5ad1c99..ee30b406ecd 100755
--- a/examples/mimo/scripts/sbatch_hetero_prod_gbs768_100n.sh
+++ b/examples/mimo/scripts/sbatch_hetero_prod_gbs768_100n.sh
@@ -8,7 +8,7 @@
 #     LR_WSD_DECAY_SAMPLES=18310547, LR_WSD_DECAY_STYLE=minus_sqrt
 #   * PACKING_BUFFER_SIZE=128
 #   * SAVE_INTERVAL=1000 (LOG_INTERVAL=1 for per-iter visibility)
-#   * NUM_WORKERS=1
+#   * NUM_WORKERS=2
 # Deviations: LLM_EP=8 (vs 16), hetero TP=2 (vs 4), force-LB=0, no MTP.
 
 #SBATCH -A nemotron_n4_pre
@@ -72,7 +72,7 @@ TRAINING_STAGE=stage2
 MODEL_PROVIDER=nemotron-moe-vlm-54l
 ENABLE_EXPERIMENTAL=1
 MOE_ROUTER_FORCE_LOAD_BALANCING=0
-NUM_WORKERS=1
+NUM_WORKERS=2
 PACKING_BUFFER_SIZE=128
 SHUFFLE_BUFFER_SIZE=100
 MAX_SAMPLES_PER_SEQUENCE=100
diff --git a/examples/mimo/scripts/sbatch_hetero_prod_gbs768_33n_ep8.sh b/examples/mimo/scripts/sbatch_hetero_prod_gbs768_33n_ep8.sh
index 1caf4221f79..6ec8336fa0b 100755
--- a/examples/mimo/scripts/sbatch_hetero_prod_gbs768_33n_ep8.sh
+++ b/examples/mimo/scripts/sbatch_hetero_prod_gbs768_33n_ep8.sh
@@ -8,7 +8,7 @@
 #     LR_WSD_DECAY_SAMPLES=18310547, LR_WSD_DECAY_STYLE=minus_sqrt
 #   * PACKING_BUFFER_SIZE=128
 #   * SAVE_INTERVAL=1000 (LOG_INTERVAL=1 for per-iter visibility)
-#   * NUM_WORKERS=1
+#   * NUM_WORKERS=2
 # Deviations from Sanjeev's baseline:
 #   * LLM_EP=8 (vs Sanjeev's EP=16) — kept from our scaling study
 #   * Hetero topology TP=2 (vs Sanjeev's TP=4)
@@ -77,7 +77,7 @@ TRAINING_STAGE=stage2
 MODEL_PROVIDER=nemotron-moe-vlm-54l
 ENABLE_EXPERIMENTAL=1
 MOE_ROUTER_FORCE_LOAD_BALANCING=0
-NUM_WORKERS=1
+NUM_WORKERS=2
 PACKING_BUFFER_SIZE=128
 SHUFFLE_BUFFER_SIZE=100
 MAX_SAMPLES_PER_SEQUENCE=100
diff --git a/examples/mimo/scripts/sbatch_hetero_prod_gbs768_68n_ep8.sh b/examples/mimo/scripts/sbatch_hetero_prod_gbs768_68n_ep8.sh
index c6676343901..3f178cfad6f 100755
--- a/examples/mimo/scripts/sbatch_hetero_prod_gbs768_68n_ep8.sh
+++ b/examples/mimo/scripts/sbatch_hetero_prod_gbs768_68n_ep8.sh
@@ -8,7 +8,7 @@
 #     LR_WSD_DECAY_SAMPLES=18310547, LR_WSD_DECAY_STYLE=minus_sqrt
 #   * PACKING_BUFFER_SIZE=128
 #   * SAVE_INTERVAL=1000 (LOG_INTERVAL=1 for per-iter visibility)
-#   * NUM_WORKERS=1
+#   * NUM_WORKERS=2
 # Deviations: LLM_EP=8 (vs 16), hetero TP=2 (vs 4), force-LB=0, no MTP.
 
 #SBATCH -A nemotron_n4_pre
@@ -72,7 +72,7 @@ TRAINING_STAGE=stage2
 MODEL_PROVIDER=nemotron-moe-vlm-54l
 ENABLE_EXPERIMENTAL=1
 MOE_ROUTER_FORCE_LOAD_BALANCING=0
-NUM_WORKERS=1
+NUM_WORKERS=2
 PACKING_BUFFER_SIZE=128
 SHUFFLE_BUFFER_SIZE=100
 MAX_SAMPLES_PER_SEQUENCE=100