Initial commit

verl/trainer/config/generation.yaml

@@ -0,0 +1,35 @@
+trainer:
+  nnodes: 1
+  n_gpus_per_node: 8
+
+data:
+  path: ~/data/rlhf/math/test.parquet
+  prompt_key: prompt
+  n_samples: 5
+  output_path: /opt/tiger/math_Qwen2-7B-Instruct.parquet
+  batch_size: 128
+
+model:
+  path: ~/models/Qwen2-7B-Instruct
+  external_lib: null
+rollout:
+  name: vllm
+  temperature: 1.0
+  top_k: 50 # 0 for hf rollout, -1 for vllm rollout
+  top_p: 0.7
+  prompt_length: 1536
+  response_length: 512
+  # for vllm rollout
+  dtype: bfloat16 # should align with FSDP
+  gpu_memory_utilization: 0.5
+  ignore_eos: False
+  micro_batch_size: 256
+  enforce_eager: True
+  free_cache_engine: True
+  load_format: dummy_dtensor
+  tensor_model_parallel_size: 1
+  max_num_batched_tokens: 8192
+  max_num_seqs: 1024
+  log_prob_micro_batch_size: 8
+  # for hf rollout
+  do_sample: True