Repository navigation
Expand file tree
/
Copy patheval_single_node.sh
More file actions
executable file
·162 lines (149 loc) · 7.47 KB
/
Copy patheval_single_node.sh
File metadata and controls
executable file
·162 lines (149 loc) · 7.47 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
#!/bin/bash
# Single-node evaluation (1 node, 8 GPUs)
set -ex
# ==================== USER CONFIGURATION (Edit these) ====================
export MODEL_PATH="/path/to/models/Qwen3-VL-8B-Instruct"
# export MODEL_PATH="/path/to/models/Qwen3-VL-32B-Instruct"
export TEXT_SEARCH_ADDRESS="127.0.0.1:8000"
export LOCAL_DATABASE_ADDRESS="127.0.0.1:8001"
export SUMMARIZER_BASE_URL="${SUMMARIZER_BASE_URL:-http://127.0.0.1:8123/v1}"
export SUMMARIZER_API_KEY="${SUMMARIZER_API_KEY:-EMPTY}"
# If you use Azure GPT judge, set this. If you use local judge (8181), you can leave it unset.
# export AZURE_OPENAI_API_KEY="<YOUR_AZURE_OPENAI_KEY>"
export LLM_JUDGE_URL="127.0.0.1:8181"
# For SGLang/vLLM OpenAI-compatible servers, API key is typically ignored; any placeholder works.
export LLM_JUDGE_API_KEY="token-abc123"
# ==========================================================================
num_nodes=1
# ==================== Environment Variables ====================
export CODE_DIR=$(dirname $0)/verl
export PYTHONPATH=$CODE_DIR:$PYTHONPATH
export PYTHONUNBUFFERED=1
export PYTHONHASHSEED=0
export CUBLAS_WORKSPACE_CONFIG=:4096:8
export MAX_PIXELS=8294400
export MIN_PIXELS=65536
# ==================== Ray Setup ====================
ray stop --force 2>/dev/null || true
sleep 3
pkill -9 -f "ray::" 2>/dev/null || true
pkill -9 -f "raylet" 2>/dev/null || true
sleep 2
ray start --head --num-gpus=8 --dashboard-host=0.0.0.0 --dashboard-port=8265 --disable-usage-stats
sleep 5
ray status || { echo "ERROR: Ray failed to start"; exit 1; }
# ==================== Evaluation ====================
TOOL_CONFIG_TRAIN="$(dirname $0)/config/tool_config/tools_train.yaml"
TOOL_CONFIG_VAL="$(dirname $0)/config/tool_config/tools_val.yaml"
train_files=/path/to/workspace/ECCV/KBSearch-v4/train_infoseek_val.json
val_files=/path/to/workspace/ECCV/KBSearch-v4/test_infoseek_val_v1.json
model_name=$(basename "$MODEL_PATH")
TIMESTAMP=$(date +%Y%m%d_%H%M%S)
exp_name=ProMSA_eval_${model_name}_${TIMESTAMP}
mkdir -p $(dirname $0)/rollout_data_PPU/$exp_name/train
mkdir -p $(dirname $0)/rollout_data_PPU/$exp_name/validation
mkdir -p $(dirname $0)/logs
LOG_TIMESTAMP=$(date +%Y%m%d_%H%M%S)
python3 -m verl.trainer.main_ppo \
algorithm.adv_estimator=grpo \
algorithm.use_kl_in_reward=False \
data.custom_cls.path="pkg://verl.utils.dataset.rl_dataset_json_v2" \
data.custom_cls.name="RLHFJSONDatasetV2" \
data.train_files=$train_files \
data.val_files=$val_files \
data.train_batch_size=128 \
data.val_batch_size=512 \
data.dataloader_num_workers=8 \
data.max_prompt_length=$((1024 * 16)) \
data.max_response_length=$((1024 * 16)) \
data.filter_overlong_prompts=False \
data.truncation='error' \
data.return_raw_chat=True \
data.return_multi_modal_inputs=False \
data.image_key=image \
data.image_patch_size=16 \
+data.image_search_max_results=4 \
data.tool_config_path="${TOOL_CONFIG_TRAIN}" \
+data.val_tool_config_path="${TOOL_CONFIG_VAL}" \
actor_rollout_ref.model.path=${MODEL_PATH} \
actor_rollout_ref.actor.optim.lr=5e-6 \
actor_rollout_ref.model.use_remove_padding=True \
actor_rollout_ref.actor.use_dynamic_bsz=True \
actor_rollout_ref.actor.ppo_mini_batch_size=128 \
actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2 \
actor_rollout_ref.actor.ppo_max_token_len_per_gpu=$((1024 * 16)) \
actor_rollout_ref.actor.use_kl_loss=False \
actor_rollout_ref.actor.kl_loss_coef=0.0 \
actor_rollout_ref.actor.kl_loss_type=low_var_kl \
actor_rollout_ref.actor.entropy_coeff=0 \
actor_rollout_ref.actor.clip_ratio_low=0.2 \
actor_rollout_ref.actor.clip_ratio_high=0.28 \
actor_rollout_ref.actor.ulysses_sequence_parallel_size=2 \
actor_rollout_ref.model.enable_gradient_checkpointing=True \
actor_rollout_ref.model.trust_remote_code=True \
actor_rollout_ref.actor.fsdp_config.param_offload=False \
actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
actor_rollout_ref.actor.fsdp_config.model_dtype=bfloat16 \
actor_rollout_ref.actor.freeze_vision_tower=True \
actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \
actor_rollout_ref.rollout.log_prob_max_token_len_per_gpu=65536 \
actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
actor_rollout_ref.rollout.name=sglang \
actor_rollout_ref.rollout.mode=async \
actor_rollout_ref.rollout.gpu_memory_utilization=0.5 \
actor_rollout_ref.rollout.enable_chunked_prefill=True \
actor_rollout_ref.rollout.enforce_eager=False \
actor_rollout_ref.rollout.max_num_batched_tokens=$((1024 * 32)) \
actor_rollout_ref.rollout.max_model_len=$((1024 * 32)) \
actor_rollout_ref.rollout.max_num_seqs=32 \
actor_rollout_ref.rollout.response_length=$((1024 * 16)) \
actor_rollout_ref.rollout.n=8 \
+actor_rollout_ref.rollout.seed=3407 \
+actor_rollout_ref.rollout.repetition_penalty=1.0 \
+actor_rollout_ref.rollout.presence_penalty=1.5 \
actor_rollout_ref.rollout.calculate_log_probs=True \
algorithm.rollout_correction.bypass_mode=True \
actor_rollout_ref.rollout.free_cache_engine=True \
+actor_rollout_ref.rollout.engine_kwargs.sglang.enable_deterministic_inference=True \
actor_rollout_ref.rollout.val_kwargs.n=1 \
actor_rollout_ref.rollout.val_kwargs.temperature=1.0 \
actor_rollout_ref.rollout.val_kwargs.top_p=1.0 \
actor_rollout_ref.rollout.val_kwargs.top_k=20 \
actor_rollout_ref.rollout.val_kwargs.do_sample=True \
+actor_rollout_ref.rollout.val_kwargs.repetition_penalty=1.0 \
+actor_rollout_ref.rollout.val_kwargs.presence_penalty=1.5 \
+actor_rollout_ref.rollout.val_kwargs.seed=3407 \
actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \
actor_rollout_ref.ref.fsdp_config.param_offload=False \
actor_rollout_ref.ref.fsdp_config.model_dtype=bfloat16 \
actor_rollout_ref.rollout.multi_turn.enable=True \
actor_rollout_ref.rollout.multi_turn.max_assistant_turns=8 \
actor_rollout_ref.rollout.multi_turn.max_tool_response_length=8192 \
+actor_rollout_ref.rollout.multi_turn.directly_provide_image_search=False \
actor_rollout_ref.rollout.multi_turn.tool_config_path="${TOOL_CONFIG_TRAIN}" \
+actor_rollout_ref.rollout.multi_turn.val_tool_config_path="${TOOL_CONFIG_VAL}" \
actor_rollout_ref.rollout.agent.default_agent_loop=tool_agent \
actor_rollout_ref.rollout.agent.num_workers=16 \
reward_model.reward_manager=tool \
reward_model.launch_reward_fn_async=True \
+reward_model.val_reward_kwargs.log_num_round=True \
+reward_model.val_reward_kwargs.format_score=0.5 \
+reward_model.val_reward_kwargs.penalize_tool_limit_reached=True \
+reward_model.val_reward_kwargs.llm_judge_model='Qwen-vl-32b' \
+reward_model.val_reward_kwargs.llm_judge_urls=["${LLM_JUDGE_URL}"] \
+reward_model.val_reward_kwargs.llm_judge_api_key="${LLM_JUDGE_API_KEY}" \
+reward_model.val_reward_kwargs.llm_judge_concurrency_limit=2 \
+reward_model.val_reward_kwargs.llm_judge_timeout=300 \
trainer.critic_warmup=0 \
trainer.val_before_train=True \
trainer.resume_mode='disable' \
trainer.logger=['console'] \
trainer.project_name='ProMSA_rl' \
trainer.experiment_name=$exp_name \
trainer.n_gpus_per_node=8 \
trainer.nnodes=$num_nodes \
trainer.save_freq=10 \
trainer.test_freq=0 \
trainer.rollout_data_dir=$(dirname $0)/rollout_data_PPU/$exp_name/train \
trainer.validation_data_dir=$(dirname $0)/rollout_data_PPU/$exp_name/validation \
trainer.total_epochs=0 2>&1 | tee $(dirname $0)/logs/${exp_name}.log