test: fix bug in deselection and make fast tests even faster (#2038)

Signed-off-by: Terry Kong <terryk@nvidia.com>

Author: terrykong

tests/functional/L1_Functional_Tests_GPU.sh

@@ -36,7 +36,7 @@ run_test() {
 
 # This test is intentionally not run with uv run --no-sync to verify that the frozen environment is working correctly.
 run_test      bash ./tests/functional/grpo_frozen_env.sh
-run_test fast bash ./tests/functional/test_frozen_env.sh
+run_test      bash ./tests/functional/test_frozen_env.sh
 
 run_test fast uv run --no-sync bash ./tests/functional/distillation.sh
 run_test      uv run --no-sync bash ./tests/functional/distillation_megatron.sh
@@ -58,7 +58,7 @@ run_test      uv run --no-sync bash ./tests/functional/grpo_non_colocated.sh
 run_test      uv run --no-sync bash ./tests/functional/grpo_rm_env.sh
 run_test      uv run --no-sync bash ./tests/functional/grpo_sglang.sh
 run_test      uv run --no-sync bash ./tests/functional/prorlv2.sh
-run_test fast uv run --no-sync bash ./tests/functional/rm.sh
+run_test      uv run --no-sync bash ./tests/functional/rm.sh
 run_test fast uv run --no-sync bash ./tests/functional/sft.sh
 run_test      uv run --no-sync bash ./tests/functional/sft_automodel_lora.sh
 run_test      uv run --no-sync bash ./tests/functional/sft_avlm.sh

tests/functional/distillation.sh

@@ -26,16 +26,16 @@ uv run coverage run -a --data-file=$PROJECT_ROOT/tests/.coverage --source=$PROJE
     policy.model_name=Qwen/Qwen3-0.6B \
     teacher.model_name=Qwen/Qwen3-0.6B \
     cluster.gpus_per_node=2 \
-    policy.train_global_batch_size=16 \
+    policy.train_global_batch_size=8 \
     policy.dtensor_cfg.tensor_parallel_size=1 \
     policy.dtensor_cfg.context_parallel_size=2 \
-    policy.max_total_sequence_length=2048 \
+    policy.max_total_sequence_length=256 \
     teacher.dtensor_cfg.tensor_parallel_size=2 \
     teacher.dtensor_cfg.context_parallel_size=1 \
     distillation.max_num_steps=3 \
     distillation.num_prompts_per_step=16 \
-    distillation.max_val_samples=16 \
-    distillation.val_batch_size=8 \
+    distillation.max_val_samples=2 \
+    distillation.val_batch_size=4 \
     distillation.val_period=3 \
     data.train.dataset_name=OpenMathInstruct-2 \
     ++data.train.split_validation_size=0.05 \

tests/functional/grpo_megatron_async.sh

@@ -27,7 +27,7 @@ uv run coverage run -a --data-file=$PROJECT_ROOT/tests/.coverage --source=$PROJE
     policy.train_global_batch_size=4 \
     policy.train_micro_batch_size=1 \
     cluster.gpus_per_node=2 \
-    grpo.max_num_steps=20 \
+    grpo.max_num_steps=5 \
     grpo.async_grpo.enabled=true \
     grpo.async_grpo.max_trajectory_age_steps=1 \
     policy.generation.vllm_cfg.async_engine=true \