change plotting

plot_and_analyze/plot.py

@@ -9,7 +9,9 @@ import matplotlib.pyplot as plt
 def process_log(folder):
     folder = Path(folder)
     # 获取文件夹内所有的 .out 文件
-    log_files = list(folder.glob("*.out"))
+    log_files = []
+    for suffix in ["out", "log"]:
+        log_files.extend(list(folder.glob(f"*.{suffix}")))
 
     key_value_pattern = re.compile(r"([\w/%(),[\]]+):\s*(-?\d+\.?\d*)")
     all_data = {}  # 用于存储每个 step 对应的最新数据

plot_and_analyze/timing.py

+import pandas as pd
+from pathlib import Path
+import argparse
+
+
+def calculate_timing(folder):
+    csv_file_path = Path(folder) / 'stats.csv'
+    df = pd.read_csv(csv_file_path)
+
+    # 筛选出以 'timing_s' 开头的列
+    timing_columns = [col for col in df.columns if col.startswith('timing_s')]
+
+    # 取前 10 行计算平均值
+    if timing_columns:
+        average_values = df[timing_columns].head(10).mean()
+        print("以 'timing_s' 开头的字段前 10 行的平均值：")
+        print(average_values)
+    else:
+        print("未找到以 'timing_s' 开头的字段。")
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Process log files and generate CSV.')
+    parser.add_argument('--folder', type=str, help='Path to the folder containing the log file.',
+                        default='ckpt/codev_distill_16k')
+    args = parser.parse_args()
+    calculate_timing(args.folder)
+    
\ No newline at end of file

recipe/dapo/run_dapo_codev_7b_7.4k_r1_gt.sh

@@ -71,7 +71,7 @@ python3 -m verl.trainer.main_ppo \
     algorithm.filter_groups.enable=${enable_filter_groups} \
     algorithm.filter_groups.max_num_gen_batches=999 \
     algorithm.filter_groups.metric=acc \
-    data.gen_batch_size=$((($train_prompt_bsz * 4 / 3 + $num_gpu - 1) / $num_gpu * $num_gpu)) \
+    data.gen_batch_size=$((($train_prompt_bsz * 3 / 2 + $num_gpu - 1) / $num_gpu * $num_gpu)) \
     actor_rollout_ref.model.path=$MODEL_PATH \
     +actor_rollout_ref.model.override_config.attention_dropout=0. \
     +actor_rollout_ref.model.override_config.embd_pdrop=0. \

recipe/grpo/run_grpo_codev_7b_3.4kdata_128x8.sh

+#!/bin/bash
+set -x
+
+# Warning: Export VLLM_ATTENTION_BACKEND on every machine before starting Ray cluster.
+# vLLM without XFORMERS will results in CUDA errors.
+# export VLLM_ATTENTION_BACKEND=XFORMERS
+
+# Parse command line arguments
+while [[ $# -gt 0 ]]; do
+    case $1 in
+        --model)
+            MODEL_PATH="$2"
+            shift 2
+            ;;
+        *)
+            break
+            ;;
+    esac
+done
+
+# Set default model path if not provided
+if [ -z "$MODEL_PATH" ]; then
+    MODEL_PATH="/nfs_global/S/lvhanqi/LLaMA-Factory/saves/Qwen2.5-Coder-7B-Instruct-codev-r1-87k/full/sft_6epoch"
+fi
+
+echo "GLOO_SOCKET_IFNAME is $GLOO_SOCKET_IFNAME!"
+
+MAX_TOKEN_PER_GPU=32768
+GPU_MEMORY_UTILIZATION=0.8
+
+python3 -m verl.trainer.main_ppo \
+    algorithm.adv_estimator=grpo \
+    data.train_files=$CURR_DIR/data/codev/v1/4.8k_r1_filtered/train.parquet \
+    data.val_files=$CURR_DIR/data/codev/v1_1/10k_qwq/test.parquet \
+    data.train_batch_size=128 \
+    data.val_batch_size=512 \
+    data.max_prompt_length=1152 \
+    data.max_response_length=16384 \
+    actor_rollout_ref.model.path=$MODEL_PATH \
+    actor_rollout_ref.actor.optim.lr=2e-6 \
+    actor_rollout_ref.actor.optim.weight_decay=0.0 \
+    +actor_rollout_ref.actor.optim.betas=[0.9,0.999] \
+    actor_rollout_ref.actor.grad_clip=0.5 \
+    actor_rollout_ref.actor.use_dynamic_bsz=True\
+    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=$MAX_TOKEN_PER_GPU \
+    actor_rollout_ref.model.use_remove_padding=True \
+    actor_rollout_ref.actor.ppo_mini_batch_size=64 \
+    actor_rollout_ref.actor.use_kl_loss=True \
+    actor_rollout_ref.actor.kl_loss_coef=0.000 \
+    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
+    actor_rollout_ref.model.enable_gradient_checkpointing=True \
+    actor_rollout_ref.actor.fsdp_config.param_offload=False \
+    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
+    actor_rollout_ref.rollout.log_prob_max_token_len_per_gpu=$(($MAX_TOKEN_PER_GPU*2)) \
+    actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+    actor_rollout_ref.rollout.name=vllm \
+    actor_rollout_ref.rollout.n=8 \
+    actor_rollout_ref.rollout.val_kwargs.n=2 \
+    actor_rollout_ref.rollout.temperature=1.0 \
+    actor_rollout_ref.rollout.val_kwargs.temperature=1.0 \
+    actor_rollout_ref.rollout.val_kwargs.do_sample=True \
+    actor_rollout_ref.rollout.gpu_memory_utilization=$GPU_MEMORY_UTILIZATION \
+    actor_rollout_ref.rollout.enforce_eager=False \
+    actor_rollout_ref.rollout.free_cache_engine=False \
+    actor_rollout_ref.ref.fsdp_config.param_offload=False \
+    algorithm.kl_ctrl.kl_coef=0.000 \
+    reward_model.reward_manager=prime \
+    trainer.critic_warmup=0 \
+    trainer.logger=['console','wandb'] \
+    trainer.project_name='codev' \
+    trainer.experiment_name='codev-distill-7b-16k-t1-kl0' \
+    trainer.n_gpus_per_node=8 \
+    trainer.nnodes=2 \
+    +trainer.val_before_train=True \
+    trainer.default_local_dir=$SAVE_DIR \
+    trainer.default_hdfs_dir=null \
+    trainer.save_freq=50 \
+    trainer.test_freq=50 \
+    trainer.total_epochs=100 "${@:1}"
\ No newline at end of file