File size: 5,535 Bytes
bc3dbfd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
163
164
# Search Environment Experiments

This doc covers the experiment scripts for the Search (HotpotQA + Dense Retrieval) environment.

## Overview

All experiments use:
- **Task**: SearchQA (HotpotQA multi-hop QA with Wikipedia dense retrieval)
- **Model**: `Qwen/Qwen2.5-3B-Instruct`
- **Algorithm**: PPO (`algorithm.adv_estimator=gae`)
- **Config**: `_9_search`

The sweep compares three rollout filtering strategies while keeping all other hyperparameters fixed.

| Experiment | Filter Strategy | Filter Value | Effective Batch | Description |
|-----------|----------------|-------------|----------------|-------------|
| No Filter | `top_p` | `1.0` | 128 | Baseline: all rollout groups kept |
| TopK 0.25 | `top_k` | `0.25` | 32 | Keep top 25% groups by reward variance |
| TopP 0.9 | `top_p` | `0.9` | ~115 | Keep groups covering 90% cumulative reward variance |

---

## Prerequisites

### 1. Prepare data

```bash
# HotpotQA train/val parquet
python scripts/prepare_search_data.py

# Wikipedia corpus + FAISS index (~74GB)
python scripts/download_search_index.py
```

### 2. Start retrieval server

The retrieval server provides dense retrieval over ~21M Wikipedia passages using E5-base-v2 + FAISS.

```bash
python scripts/retrieval/server.py \
    --data_dir ./search_data/prebuilt_indices \
    --port 8000 --host 127.0.0.1 \
    --device cuda:0 --gpu_memory_limit_mb 6144
```

**Important**: We recommend running the retrieval server on a **dedicated GPU** not used by training, or on CPU. Sharing a GPU with vLLM rollout and training causes CUDA OOM errors due to memory contention between processes.

---

## Experiment Scripts

All experiments use `scripts/runs/run_search_benchmark.sh`.

### Experiment 1: PPO + No Filter (baseline)

No filtering — all rollout groups are used for training.

```bash
bash scripts/runs/run_search_benchmark.sh \
    --algos PPO \
    --filter-strategy top_p --filter-value 1.0 \
    --gpus 0,1,2,3,4,5,6,7 --gpus-per-exp 8 \
    --micro-batch 4 --mini-batch 64 \
    --gpu-memory-utilization 0.65 \
    --save-freq 20 --steps 200 \
    --retrieval-port 8000
```

### Experiment 2: PPO + TopK=0.25

Keep only the top 25% of rollout groups ranked by reward variance.

```bash
bash scripts/runs/run_search_benchmark.sh \
    --algos PPO \
    --filter-strategy top_k --filter-value 0.25 \
    --gpus 0,1,2,3,4,5,6,7 --gpus-per-exp 8 \
    --micro-batch 4 --mini-batch 32 \
    --gpu-memory-utilization 0.65 \
    --save-freq 20 --steps 200 \
    --retrieval-port 8000
```

Note: `mini-batch` is reduced to 32 because effective batch after filtering is `16 groups * 8 group_size * 0.25 = 32`. The `ppo_mini_batch_size` must not exceed this value.

### Experiment 3: PPO + TopP=0.9

Keep rollout groups covering the top 90% cumulative reward variance (softmax-weighted).

```bash
bash scripts/runs/run_search_benchmark.sh \
    --algos PPO \
    --filter-strategy top_p --filter-value 0.9 \
    --gpus 0,1,2,3,4,5,6,7 --gpus-per-exp 8 \
    --micro-batch 4 --mini-batch 64 \
    --gpu-memory-utilization 0.65 \
    --save-freq 20 --steps 200 \
    --retrieval-port 8000
```

---

## W&B Runs

Project: [`cuhksz-gc/ragen_search_benchmark`](https://wandb.ai/cuhksz-gc/ragen_search_benchmark)

| Experiment | Run ID | Link |
|-----------|--------|------|
| PPO + No Filter | `2sbt8952` | [wandb](https://wandb.ai/cuhksz-gc/ragen_search_benchmark/runs/2sbt8952) |
| PPO + TopK=0.25 | `2h5c7kbb` | [wandb](https://wandb.ai/cuhksz-gc/ragen_search_benchmark/runs/2h5c7kbb) |
| PPO + TopP=0.9 | `tbgx0lpt` | [wandb](https://wandb.ai/cuhksz-gc/ragen_search_benchmark/runs/tbgx0lpt) |

---

## Shared Config

```yaml
# config/_9_search.yaml overrides
micro_batch_size_per_gpu: 4
ppo_mini_batch_size: 32-64  # depends on filter setting

agent_proxy:
  max_turn: 5
  max_actions_per_turn: 1

actor_rollout_ref:
  rollout:
    max_model_len: 5000       # TopK=0.25 experiment used 4000
    max_num_batched_tokens: 5000  # TopK=0.25 experiment used 4000
    gpu_memory_utilization: 0.65
    temperature: 1
  actor:
    use_kl_loss: False
    kl_loss_coef: 0.001
    entropy_coeff: 0.001
    loss_agg_mode: token-mean
    filter_loss_scaling: none

es_manager:
  train:
    env_groups: 16
    group_size: 8    # 16 * 8 = 128 rollouts per step
  val:
    env_groups: 256

collapse_detection:
  compute_freq: 999  # effectively disabled

trainer:
  total_training_steps: 200
  save_freq: 20
  val_before_train: True
  logger: ['console', 'wandb']
```

---

## Common Notes

- **Retrieval server GPU deployment**: Place the E5 retrieval server on a **dedicated GPU** not used by training. Co-locating with training on the same GPU causes CUDA OOM due to memory contention between vLLM, training, and the E5 server process. Do not use CPU mode — during rollout, hundreds of environments issue concurrent retrieval requests (256 env groups can produce 1000+ requests), and CPU cannot keep up.
- **mini-batch size adjustment**: When using aggressive filtering (e.g., `top_k=0.25`), reduce `ppo_mini_batch_size` so it does not exceed `env_groups * group_size * filter_value`. Otherwise training fails with an assertion error.
- **max_model_len**: Default is 5000 (in `_9_search.yaml`). The TopK=0.25 experiment used 4000 to save KV cache memory; the No Filter and TopP=0.9 experiments use the default 5000.
- **Checkpoint size**: Each checkpoint is ~35GB (model + optimizer, 8 FSDP shards). With `save_freq=20` and 200 steps, expect 10 checkpoints (~350GB). Monitor disk usage and delete old checkpoints as needed.