megatron backend for sglang rollout

examples/configs/recipes/llm/grpo-qwen2.5-math-1.5b-instruct-1n2g-fsdp2tp1-sglang.yaml

@@ -0,0 +1,66 @@
+defaults: ../../grpo_math_1B.yaml
+
+grpo:
+  max_num_steps: 450
+  val_batch_size: 128
+
+checkpointing:
+  checkpoint_dir: results/grpo-qwen2.5-math-1.5b-instruct-1n2g-fsdp2tp1-sglang
+
+policy:
+  model_name: Qwen/Qwen2.5-Math-1.5B-Instruct
+  tokenizer:
+    name: Qwen/Qwen2.5-Math-1.5B-Instruct
+  dynamic_batching:
+    enabled: true
+  sequence_packing:
+    enabled: false
+  make_sequence_length_divisible_by: 1
+  generation:
+    backend: "sglang"
+    max_new_tokens: 512
+    sglang_cfg:
+      model_path: ${policy.model_name}
+      dtype: ${policy.precision}
+      context_length: 512
+      allow_auto_truncate: true
+      dp_size: 1
+      pp_size: 1
+      ep_size: 1
+      random_seed: 42
+      max_running_requests: null
+      mem_fraction_static: 0.6
+      skip_server_warmup: true
+      # Piecewise CUDA graph currently crashes with "illegal memory access"
+      # (likely torch 2.10 + sglang incompatibility). Keep disabled until upstream fix.
+      disable_piecewise_cuda_graph: true
+      disable_cuda_graph: false
+      # Fault tolerance (RolloutHealthMonitor). Off by default; when enabled,
+      # a daemon thread health-checks each engine and kills hung actors.
+      use_fault_tolerance: false
+      rollout_health_check_interval: 60
+      rollout_health_check_timeout: 60
+      rollout_health_check_first_wait: 60
+    sglang_server:
+      needs_offload: true
+      cpu_weight_backup: true
+      sglang_server_concurrency: 1024
+      pause_generation_mode: retract
+      num_gpus: 2
+      num_gpus_per_engine: 2
+    sglang_router: {}
+
+data:
+  max_input_seq_length: 512
+
+logger:
+  log_dir: logs/grpo-qwen2.5-math-1.5b-instruct-1n2g-fsdp2tp1-sglang
+  wandb_enabled: true
+  tensorboard_enabled: true
+  wandb:
+    project: nemo-rl
+    name: grpo-qwen2.5-math-1.5b-instruct-1n2g-fsdp2tp1-sglang
+
+cluster:
+  gpus_per_node: 2
+  num_nodes: 1

examples/configs/recipes/llm/grpo-qwen2.5-math-1.5b-instruct-1n2g-fsdp2tp1-vllm.yaml

@@ -0,0 +1,38 @@
+defaults: ../../grpo_math_1B.yaml
+
+grpo:
+  max_num_steps: 450
+  val_batch_size: 128
+
+checkpointing:
+  checkpoint_dir: results/grpo-qwen2.5-math-1.5b-instruct-1n2g-fsdp2tp1-vllm
+
+policy:
+  model_name: Qwen/Qwen2.5-Math-1.5B-Instruct
+  tokenizer:
+    name: Qwen/Qwen2.5-Math-1.5B-Instruct
+  dynamic_batching:
+    enabled: true
+  sequence_packing:
+    enabled: false
+  make_sequence_length_divisible_by: 1
+  generation:
+    max_new_tokens: 512
+    vllm_cfg:
+      max_model_len: 512
+      tensor_parallel_size: 2
+
+data:
+  max_input_seq_length: 512
+
+logger:
+  log_dir: logs/grpo-qwen2.5-math-1.5b-instruct-1n2g-fsdp2tp1-vllm
+  wandb_enabled: true
+  tensorboard_enabled: true
+  wandb:
+    project: nemo-rl
+    name: grpo-qwen2.5-math-1.5b-instruct-1n2g-fsdp2tp1-vllm
+
+cluster:
+  gpus_per_node: 2
+  num_nodes: 1