# Local launch mirrors the repo's node launcher, not standard torchrun
# semantics. eval.py spawns one worker per visible GPU by itself.
# Here RANK/WORLD_SIZE mean node_rank/node_count, so WORLD_SIZE=1 is a
# single-node local run; total GPU workers come from CUDA_VISIBLE_DEVICES.
export CUDA_VISIBLE_DEVICES=0,1,2,3
export MASTER_ADDR=127.0.0.1
export MASTER_PORT=29500
export RANK=0
export WORLD_SIZE=1

# Match this to the target model used by the draft checkpoint.
target_name_or_path=Qwen/Qwen3-4B

# Training writes checkpoints under ~/checkpoints/<project_name>/<exp_name>/step_*.
# Use step_latest for the most recent checkpoint, or replace it with step_<N>.
draft_name_or_path=${HOME}/checkpoints/deepspec/dspark_block7_qwen3_4b/step_latest
python eval.py \
    --target_name_or_path ${target_name_or_path} \
    --draft_name_or_path ${draft_name_or_path}