fix: lower steps in smolvlm nightly test (#1239)

Signed-off-by: Terry Kong <[email protected]>

Author: terrykong

…-clevr-1n2g-dtensor2tp1.v1.yaml.disabled …-clevr-1n2g-dtensor2tp1.v2.yaml.disabledexamples/configs/recipes/vlm/vlm_grpo-smolvlm2-2.2b-instruct-clevr-1n2g-dtensor2tp1.v1.yaml.disabled renamed to examples/configs/recipes/vlm/vlm_grpo-smolvlm2-2.2b-instruct-clevr-1n2g-dtensor2tp1.v2.yaml.disabled examples/configs/recipes/vlm/vlm_grpo-smolvlm2-2.2b-instruct-clevr-1n2g-dtensor2tp1.v1.yaml.disabled renamed to examples/configs/recipes/vlm/vlm_grpo-smolvlm2-2.2b-instruct-clevr-1n2g-dtensor2tp1.v2.yaml.disabled

@@ -24,7 +24,7 @@ tests/test_suites/vlm/vlm_grpo-qwen2.5-vl-3b-instruct-clevr-1n2g-dtensor2tp1.v1.
 tests/test_suites/vlm/vlm_grpo-qwen2.5-vl-3b-instruct-clevr-1n2g-megatrontp2.v1.sh
 
 # Removing this until this issue is resolved: https://github.com/huggingface/transformers/issues/41190
-# tests/test_suites/vlm/vlm_grpo-smolvlm2-2.2b-instruct-clevr-1n2g-dtensor2tp1.v1.sh
+# tests/test_suites/vlm/vlm_grpo-smolvlm2-2.2b-instruct-clevr-1n2g-dtensor2tp1.v2.sh
 
 # Deepscaler (short tests)
 tests/test_suites/llm/grpo-deepscaler-1.5b-16K.sh

tests/test_suites/nightly.txt

@@ -4,8 +4,8 @@ source $SCRIPT_DIR/common.env
 
 # ===== BEGIN CONFIG =====
 NUM_NODES=1
-STEPS_PER_RUN=200
-MAX_STEPS=200
+STEPS_PER_RUN=130
+MAX_STEPS=130
 NUM_RUNS=$(( (MAX_STEPS + STEPS_PER_RUN - 1) / STEPS_PER_RUN ))  # Round up
 NUM_MINUTES=180
 # ===== END CONFIG =====
@@ -34,7 +34,7 @@ uv run tests/json_dump_tb_logs.py $LOG_DIR --output_path $JSON_METRICS
 # Only run metrics if the target step is reached
 if [[ $(jq 'to_entries | .[] | select(.key == "train/loss") | .value | keys | map(tonumber) | max' $JSON_METRICS) -ge $MAX_STEPS ]]; then
     uv run tests/check_metrics.py $JSON_METRICS \
-        'data["train/loss"]["200"] < 0.1' \
-        'data["train/reward"]["200"] > 0.7'  # less performant than qwen
+        'data["train/loss"]["130"] < 0.1' \
+        'mean(data["train/reward"], -6, -1) > 0.6'  # less performant than qwen
 fi