test: Perf recipe for v0.5

examples/configs/recipes/llm/grpo-moonlight-16ba3b-4n8g-megatron-fp8-e2e.yaml

@@ -16,8 +16,6 @@ policy:
   max_total_sequence_length: 8192
   dtensor_cfg:
     enabled: false
-  sequence_packing:
-    algorithm: modified_ffd
   make_sequence_length_divisible_by: ${policy.megatron_cfg.tensor_model_parallel_size}
   optimizer: null
   megatron_cfg:
@@ -45,7 +43,6 @@ policy:
       precision: fp8
       use_deep_gemm: true
       gpu_memory_utilization: 0.5
-      expert_parallel_size: 4
       quantization_ignored_layer_kws: [
         a_proj,
         b_proj

examples/configs/recipes/llm/performance/grpo-deepseek-v3-32n4g.yaml

@@ -0,0 +1,21 @@
+defaults: ./grpo-deepseek-v3-32n8g.yaml
+checkpointing:
+  checkpoint_dir: results/grpo-deepseek-v3-32n4g
+policy:
+  sequence_packing:
+    enabled: false
+  megatron_cfg:
+    pipeline_model_parallel_size: 8
+    expert_model_parallel_size: 16
+    num_layers_in_first_pipeline_stage: 7
+    num_layers_in_last_pipeline_stage: 6
+  generation:
+    vllm_cfg:
+      tensor_parallel_size: 16
+logger:
+  log_dir: logs/grpo-deepseek-v3-32n4g
+  wandb:
+    name: grpo-deepseek-v3-32n4g
+cluster:
+  gpus_per_node: 4
+  num_nodes: 32

examples/configs/recipes/llm/performance/grpo-deepseek-v3-64n4g-async-1off.yaml

@@ -0,0 +1,22 @@
+defaults: ./grpo-deepseek-v3-64n8g-async-1off.yaml
+checkpointing:
+  checkpoint_dir: results/grpo-deepseek-v3-64n4g-async-1off
+policy:
+  sequence_packing:
+    enabled: false
+  megatron_cfg:
+    pipeline_model_parallel_size: 8
+    expert_model_parallel_size: 16
+    num_layers_in_first_pipeline_stage: 7
+    num_layers_in_last_pipeline_stage: 6
+  generation:
+    vllm_cfg:
+      tensor_parallel_size: 16
+      gpu_memory_utilization: 0.8
+logger:
+  log_dir: logs/grpo-deepseek-v3-64n4g-async-32T32G-1off
+  wandb:
+    name: grpo-deepseek-v3-64n4g-async-32T32G-1off
+cluster:
+  gpus_per_node: 4
+  num_nodes: 64

examples/configs/recipes/llm/performance/grpo-deepseek-v3-64n8g-fp8-async-1off.yaml

@@ -0,0 +1,26 @@
+defaults: ./grpo-deepseek-v3-64n8g-async-1off.yaml
+checkpointing:
+  checkpoint_dir: results/grpo-deepseek-v3-64n8g-fp8-async-1off
+policy:
+  megatron_cfg:
+    fp8_cfg:
+      enabled: true
+      fp8: "e4m3"
+      fp8_recipe: "blockwise"
+      fp8_param: false
+    moe_router_dtype: fp32
+    env_vars:
+      NVTE_FP8_BLOCK_SCALING_FP32_SCALES: "1"
+  generation:
+    vllm_cfg:
+      tensor_parallel_size: 16
+      precision: "fp8"
+      use_deep_gemm: true
+      quantization_ignored_layer_kws: [
+        a_proj,
+        b_proj
+      ]
+logger:
+  log_dir: logs/grpo-deepseek-v3-64n8g-fp8-async-1off
+  wandb:
+    name: grpo-deepseek-v3-64n8g-fp8-async-1off

examples/configs/recipes/llm/performance/grpo-llama3.1-8b-instruct-2n8g-async-1off.yaml

@@ -9,6 +9,8 @@ loss_fn:
 checkpointing:
   checkpoint_dir: results/grpo-llama3.1-8b-instruct-2n8g-async-1off
 policy:
+  megatron_cfg:
+    pipeline_model_parallel_size: 1
   generation:
     colocated:
       enabled: false

examples/configs/recipes/llm/performance/grpo-llama3.1-8b-instruct-2n8g-fp8-async-1off.yaml

@@ -0,0 +1,20 @@
+defaults: ./grpo-llama3.1-8b-instruct-2n8g-async-1off.yaml
+checkpointing:
+  checkpoint_dir: results/grpo-llama3.1-8b-instruct-2n8g-fp8-async-1off
+policy:
+  megatron_cfg:
+    fp8_cfg:
+      enabled: true
+      fp8: "e4m3"
+      fp8_recipe: "blockwise"
+      fp8_param: false
+    env_vars:
+      NVTE_FP8_BLOCK_SCALING_FP32_SCALES: "1"
+  generation:
+    vllm_cfg:
+      precision: "fp8"
+      use_deep_gemm: true
+logger:
+  log_dir: logs/grpo-llama3.1-8b-instruct-2n8g-fp8-async-1off
+  wandb:
+    name: grpo-llama3.1-8b-instruct-2n8g-fp8-async-1off

examples/configs/recipes/llm/performance/grpo-qwen3-235b-16n4g.yaml

@@ -0,0 +1,20 @@
+defaults: ./grpo-qwen3-235b-16n8g.yaml
+checkpointing:
+  checkpoint_dir: results/grpo-qwen3-235b-16n4g
+policy:
+  sequence_packing:
+    enabled: false
+  megatron_cfg:
+    pipeline_model_parallel_size: 4
+    num_layers_in_first_pipeline_stage: 23
+    num_layers_in_last_pipeline_stage: 23
+  generation:
+    vllm_cfg:
+      tensor_parallel_size: 8
+logger:
+  log_dir: logs/grpo-qwen3-235b-16n4g
+  wandb:
+    name: grpo-qwen3-235b-16n4g
+cluster:
+  gpus_per_node: 4
+  num_nodes: 16

examples/configs/recipes/llm/performance/grpo-qwen3-235b-32n4g-async-1off.yaml

@@ -0,0 +1,20 @@
+defaults: ./grpo-qwen3-235b-32n8g-async-1off.yaml
+checkpointing:
+  checkpoint_dir: results/grpo-qwen3-235b-32n4g-async-1off
+policy:
+  sequence_packing:
+    enabled: false
+  megatron_cfg:
+    pipeline_model_parallel_size: 4
+    num_layers_in_first_pipeline_stage: 23
+    num_layers_in_last_pipeline_stage: 23
+  generation:
+    vllm_cfg:
+      tensor_parallel_size: 8
+logger:
+  log_dir: logs/grpo-qwen3-235b-32n4g-async-1off
+  wandb:
+    name: grpo-qwen3-235b-32n4g-async-1off
+cluster:
+  gpus_per_node: 4
+  num_nodes: 32

tests/test_suites/llm/performance/grpo-deepseek-v3-32n4g.sh

@@ -0,0 +1,45 @@
+#!/bin/bash
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd)
+source $SCRIPT_DIR/common.env
+# disable NVLS to avoid OOM issue
+export NCCL_NVLS_ENABLE=0
+# allow user to pass an existing HF checkpoint path based on instruction in https://github.com/NVIDIA-NeMo/RL/blob/main/docs/guides/deepseek.md
+export MODEL_NAME=${NRL_DEEPSEEK_V3_HF_CKPT:-"unsloth/DeepSeek-V3-0324-BF16"}
+
+# ===== BEGIN CONFIG =====
+NUM_NODES=32
+STEPS_PER_RUN=10
+MAX_STEPS=10
+NUM_RUNS=$(( (MAX_STEPS + STEPS_PER_RUN - 1) / STEPS_PER_RUN ))  # Round up
+NUM_MINUTES=240
+# ===== END CONFIG =====
+
+exit_if_max_steps_reached
+
+# Run the experiment
+cd $PROJECT_ROOT
+uv run examples/run_grpo_math.py \
+    --config $CONFIG_PATH \
+    grpo.max_num_steps=$MAX_STEPS \
+    policy.model_name=$MODEL_NAME \
+    policy.tokenizer.name=$MODEL_NAME \
+    logger.log_dir=$LOG_DIR \
+    logger.wandb_enabled=True \
+    logger.wandb.project=nemo-rl \
+    logger.wandb.name=$EXP_NAME \
+    logger.monitor_gpus=True \
+    logger.tensorboard_enabled=True \
+    checkpointing.enabled=True \
+    checkpointing.checkpoint_dir=$CKPT_DIR \
+    $@ \
+    2>&1 | tee $RUN_LOG
+
+# Convert tensorboard logs to json
+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"]) < 1.1' \
+        'data["train/token_mult_prob_error"]["10"] < 1.1'
+fi

tests/test_suites/llm/performance/grpo-deepseek-v3-64n4g-async-1off.sh

@@ -0,0 +1,45 @@
+#!/bin/bash
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd)
+source $SCRIPT_DIR/common.env
+# disable NVLS to avoid OOM issue
+export NCCL_NVLS_ENABLE=0
+# allow user to pass an existing HF checkpoint path based on instruction in https://github.com/NVIDIA-NeMo/RL/blob/main/docs/guides/deepseek.md
+export MODEL_NAME=${NRL_DEEPSEEK_V3_HF_CKPT:-"unsloth/DeepSeek-V3-0324-BF16"}
+
+# ===== BEGIN CONFIG =====
+NUM_NODES=64
+STEPS_PER_RUN=10
+MAX_STEPS=10
+NUM_RUNS=$(( (MAX_STEPS + STEPS_PER_RUN - 1) / STEPS_PER_RUN ))  # Round up
+NUM_MINUTES=240
+# ===== END CONFIG =====
+
+exit_if_max_steps_reached
+
+# Run the experiment
+cd $PROJECT_ROOT
+uv run examples/run_grpo_math.py \
+    --config $CONFIG_PATH \
+    grpo.max_num_steps=$MAX_STEPS \
+    policy.model_name=$MODEL_NAME \
+    policy.tokenizer.name=$MODEL_NAME \
+    logger.log_dir=$LOG_DIR \
+    logger.wandb_enabled=True \
+    logger.wandb.project=nemo-rl \
+    logger.wandb.name=$EXP_NAME \
+    logger.monitor_gpus=True \
+    logger.tensorboard_enabled=True \
+    checkpointing.enabled=True \
+    checkpointing.checkpoint_dir=$CKPT_DIR \
+    $@ \
+    2>&1 | tee $RUN_LOG
+
+# Convert tensorboard logs to json
+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"]) < 1.1' \
+        'data["train/token_mult_prob_error"]["10"] < 1.1'
+fi

tests/test_suites/llm/performance/grpo-deepseek-v3-64n8g-fp8-async-1off.sh

@@ -0,0 +1,45 @@
+#!/bin/bash
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd)
+source $SCRIPT_DIR/common.env
+# disable NVLS to avoid OOM issue
+export NCCL_NVLS_ENABLE=0
+# allow user to pass an existing HF checkpoint path based on instruction in https://github.com/NVIDIA-NeMo/RL/blob/main/docs/guides/deepseek.md
+export MODEL_NAME=${NRL_DEEPSEEK_V3_HF_CKPT:-"unsloth/DeepSeek-V3-0324-BF16"}
+
+# ===== BEGIN CONFIG =====
+NUM_NODES=64
+STEPS_PER_RUN=10
+MAX_STEPS=10
+NUM_RUNS=$(( (MAX_STEPS + STEPS_PER_RUN - 1) / STEPS_PER_RUN ))  # Round up
+NUM_MINUTES=240
+# ===== END CONFIG =====
+
+exit_if_max_steps_reached
+
+# Run the experiment
+cd $PROJECT_ROOT
+uv run examples/run_grpo_math.py \
+    --config $CONFIG_PATH \
+    grpo.max_num_steps=$MAX_STEPS \
+    policy.model_name=$MODEL_NAME \
+    policy.tokenizer.name=$MODEL_NAME \
+    logger.log_dir=$LOG_DIR \
+    logger.wandb_enabled=True \
+    logger.wandb.project=nemo-rl \
+    logger.wandb.name=$EXP_NAME \
+    logger.monitor_gpus=True \
+    logger.tensorboard_enabled=True \
+    checkpointing.enabled=True \
+    checkpointing.checkpoint_dir=$CKPT_DIR \
+    $@ \
+    2>&1 | tee $RUN_LOG
+
+# Convert tensorboard logs to json
+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"]) < 1.1' \
+        'data["train/token_mult_prob_error"]["10"] < 1.1'
+fi

tests/test_suites/llm/performance/grpo-llama3.1-8b-instruct-2n8g-fp8-async-1off.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd)
+source $SCRIPT_DIR/common.env
+
+# ===== BEGIN CONFIG =====
+NUM_NODES=2
+STEPS_PER_RUN=10
+MAX_STEPS=10
+NUM_RUNS=$(( (MAX_STEPS + STEPS_PER_RUN - 1) / STEPS_PER_RUN ))  # Round up
+NUM_MINUTES=100
+# ===== END CONFIG =====
+
+exit_if_max_steps_reached
+
+# Run the experiment
+cd $PROJECT_ROOT
+uv run examples/run_grpo_math.py \
+    --config $CONFIG_PATH \
+    grpo.max_num_steps=$MAX_STEPS \
+    logger.log_dir=$LOG_DIR \
+    logger.wandb_enabled=True \
+    logger.wandb.project=nemo-rl \
+    logger.wandb.name=$EXP_NAME \
+    logger.monitor_gpus=True \
+    logger.tensorboard_enabled=True \
+    checkpointing.enabled=True \
+    checkpointing.checkpoint_dir=$CKPT_DIR \
+    $@ \
+    2>&1 | tee $RUN_LOG
+
+# Convert tensorboard logs to json
+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"]) < 1.1' \
+        'data["train/token_mult_prob_error"]["10"] < 1.1'
+fi

tests/test_suites/llm/performance/grpo-qwen3-235b-16n4g.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd)
+source $SCRIPT_DIR/common.env
+# disable NVLS to avoid OOM issue
+export NCCL_NVLS_ENABLE=0
+
+# ===== BEGIN CONFIG =====
+NUM_NODES=16
+STEPS_PER_RUN=10
+MAX_STEPS=10
+NUM_RUNS=$(( (MAX_STEPS + STEPS_PER_RUN - 1) / STEPS_PER_RUN ))  # Round up
+NUM_MINUTES=100
+# ===== END CONFIG =====
+
+exit_if_max_steps_reached
+
+# Run the experiment
+cd $PROJECT_ROOT
+uv run examples/run_grpo_math.py \
+    --config $CONFIG_PATH \
+    grpo.max_num_steps=$MAX_STEPS \
+    logger.log_dir=$LOG_DIR \
+    logger.wandb_enabled=True \
+    logger.wandb.project=nemo-rl \
+    logger.wandb.name=$EXP_NAME \
+    logger.monitor_gpus=True \
+    checkpointing.enabled=True \
+    checkpointing.checkpoint_dir=$CKPT_DIR \
+    $@ \
+    2>&1 | tee $RUN_LOG
+
+# Convert tensorboard logs to json
+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"]) < 1.1' \
+        'data["train/token_mult_prob_error"]["10"] < 1.1'
+fi