update nightly test

Signed-off-by: Yuki Huang <yukih@nvidia.com>

Author: yuki-97

examples/configs/recipes/llm/grpo-llama3.2-1b-instruct-1n8g-fsdp2tp2-temp0.8-topp0.9-topk50.yaml

@@ -0,0 +1,14 @@
+defaults: grpo-llama3.2-1b-instruct-1n8g-fsdp2tp1.v3.yaml
+checkpointing:
+  checkpoint_dir: results/grpo-llama3.2-1b-instruct-1n8g-fsdp2tp2-temp0.8-topp0.9-topk50
+policy:
+  dtensor_cfg:
+    tensor_parallel_size: 2
+  generation:
+    temperature: 0.8
+    top_p: 0.9
+    top_k: 50
+logger:
+  log_dir: logs/grpo-llama3.2-1b-instruct-1n8g-fsdp2tp2-temp0.8-topp0.9-topk50
+  wandb:
+    name: grpo-llama3.2-1b-instruct-1n8g-fsdp2tp2-temp0.8-topp0.9-topk50

examples/configs/recipes/llm/grpo-llama3.2-1b-instruct-1n8g-megatron-sampling.yaml

@@ -0,0 +1,14 @@
+defaults: grpo-llama3.2-1b-instruct-1n8g-megatron.yaml
+checkpointing:
+  checkpoint_dir: results/grpo-llama3.2-1b-instruct-1n8g-megatron-temp0.8-topp0.9-topk50
+policy:
+  megatron_cfg:
+    tensor_model_parallel_size: 2
+  generation:
+    temperature: 0.8
+    top_p: 0.9
+    top_k: 50
+logger:
+  log_dir: logs/grpo-llama3.2-1b-instruct-1n8g-megatron-temp0.8-topp0.9-topk50
+  wandb:
+    name: grpo-llama3.2-1b-instruct-1n8g-megatron-temp0.8-topp0.9-topk50

examples/configs/recipes/llm/grpo-llama3.2-1b-instruct-1n8g-megatron-temp0.6.yaml

@@ -7,7 +7,7 @@ NUM_NODES=1
 STEPS_PER_RUN=500
 MAX_STEPS=500
 NUM_RUNS=$(( (MAX_STEPS + STEPS_PER_RUN - 1) / STEPS_PER_RUN ))  # Round up
-NUM_MINUTES=180
+NUM_MINUTES=120
 # ===== END CONFIG =====
 
 exit_if_max_steps_reached
@@ -34,10 +34,9 @@ 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 \
-        'mean(data["train/token_mult_prob_error"], ignore_top_p=0.01) < 1.05' \
+        'median(data["train/token_mult_prob_error"]) < 1.1' \
         'data["train/token_mult_prob_error"]["500"] < 1.1' \
-        'data["train/reward"]["500"] > 0.1' \
-        'mean(data["timing/train/total_step_time"], -6, -1) < 12.5'
+        'mean(data["timing/train/total_step_time"], -6, -1) < 10'
 
     # Clean up checkpoint directory after successful run to save space.
     rm -rf "$CKPT_DIR"

examples/configs/recipes/llm/grpo-llama3.2-1b-instruct-1n8g-megatron-temp0.8-topp0.9-topk50.yaml

@@ -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 \
-        'mean(data["train/token_mult_prob_error"], ignore_top_p=0.01) < 1.05' \
+        'median(data["train/token_mult_prob_error"]) < 1.1' \
         'data["train/token_mult_prob_error"]["500"] < 1.1' \
         'data["train/reward"]["500"] > 0.1' \
         'mean(data["timing/train/total_step_time"], -6, -1) < 10.5'

….2-1b-instruct-1n8g-megatron-sampling.sh …-1n8g-fsdp2tp2-temp0.8-topp0.9-topk50.shtests/test_suites/llm/grpo-llama3.2-1b-instruct-1n8g-megatron-sampling.sh renamed to tests/test_suites/llm/grpo-llama3.2-1b-instruct-1n8g-fsdp2tp2-temp0.8-topp0.9-topk50.sh tests/test_suites/llm/grpo-llama3.2-1b-instruct-1n8g-megatron-sampling.sh renamed to tests/test_suites/llm/grpo-llama3.2-1b-instruct-1n8g-fsdp2tp2-temp0.8-topp0.9-topk50.sh

@@ -23,8 +23,6 @@ tests/test_suites/llm/grpo-moonlight-16b-automodel-1n8g-ep8.sh
 # Megatron
 tests/test_suites/llm/grpo-llama3.2-1b-instruct-1n8g-megatron.sh
 tests/test_suites/llm/grpo-llama3.2-1b-instruct-1n8g-megatron_generation.sh
-tests/test_suites/llm/grpo-llama3.2-1b-instruct-1n8g-megatron-sampling.sh
-tests/test_suites/llm/grpo-llama3.2-1b-instruct-1n8g-megatron-temp0.6.sh
 
 # Functional 32b run
 tests/test_suites/llm/grpo-qwen2.5-32b-32n8g-fsdp2tp8-actckpt.v3.sh
@@ -80,6 +78,10 @@ tests/test_suites/llm/grpo-nanov3-30BA3B-2n8g-megatron-lora.sh
 tests/test_suites/llm/grpo-qwen3-8B-base-1n8g-fsdp2-lora.sh
 tests/test_suites/llm/grpo-qwen3-8b-base-1n8g-megatron-lora.sh
 
+# Sampling (temperature, top-p, top-k)
+tests/test_suites/llm/grpo-llama3.2-1b-instruct-1n8g-fsdp2tp2-temp0.8-topp0.9-topk50.sh
+tests/test_suites/llm/grpo-llama3.2-1b-instruct-1n8g-megatron-temp0.8-topp0.9-topk50.sh
+
 #######
 # SFT #
 #######