# Optional: run train_sft_qwen3.py under `accelerate launch` instead of `torchrun`.
# The script builds its own FSDP wrapper, so accelerate is used ONLY as a launcher here
# (it sets RANK/LOCAL_RANK/WORLD_SIZE). These fsdp_config values document the exact
# equivalent settings for anyone who prefers accelerate's Accelerator.prepare().
#
#   accelerate launch --config_file accelerate_fsdp_qwen3.yaml train_sft_qwen3.py ...
#
# If you DO switch to Accelerator.prepare(), you must keep:
#   * mixed_precision bf16 with fp32 master weights and fp32 gradient reduction
#   * FULL_SHARD + transformer-layer wrapping of Qwen3DecoderLayer
#   * use_orig_params: true   (the weight-decay parameter groups depend on it)
#   * accelerator.clip_grad_norm_ AFTER all accumulation micro-steps
#   * your own loss normalisation; accelerate's default 1/gradient_accumulation_steps
#     is NOT the recipe's per-microbatch token-weighted mean.
compute_environment: LOCAL_MACHINE
distributed_type: FSDP
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 8
rdzv_backend: static
same_network: true
use_cpu: false
enable_cpu_affinity: true
fsdp_config:
  fsdp_version: 1
  fsdp_sharding_strategy: FULL_SHARD
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: Qwen3DecoderLayer
  fsdp_backward_prefetch: BACKWARD_PRE
  fsdp_forward_prefetch: false
  fsdp_use_orig_params: true
  fsdp_cpu_ram_efficient_loading: true
  fsdp_sync_module_states: true
  fsdp_offload_params: false
  fsdp_state_dict_type: SHARDED_STATE_DICT
  fsdp_activation_checkpointing: false   # the script calls gradient_checkpointing_enable itself
  fsdp_reshard_after_forward: true
  fsdp_limit_all_gathers: true