File size: 8,242 Bytes
fa6856c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
import json
import math
import os
import sys
from itertools import islice

import numpy as np
import torch
import tritonclient.grpc as client_util
from datasets import load_dataset
from huggingface_hub import snapshot_download
from torch import nn
from transformers import AutoModelForCausalLM, AutoTokenizer
from tritonclient.utils import np_to_triton_dtype

import trlx
from trlx.data.default_configs import (
    ModelConfig,
    OptimizerConfig,
    PPOConfig,
    SchedulerConfig,
    TokenizerConfig,
    TrainConfig,
    TRLConfig,
)

default_config = TRLConfig(
    train=TrainConfig(
        seq_length=1024,
        epochs=10000,
        total_steps=10000,
        batch_size=4,
        checkpoint_interval=10000,
        eval_interval=500,
        pipeline="PromptPipeline",
        trainer="AcceleratePPOTrainer",
        checkpoint_dir="checkpoints/ppo_hh",
    ),
    model=ModelConfig(model_path="EleutherAI/gpt-j-6B", num_layers_unfrozen=2),
    tokenizer=TokenizerConfig(tokenizer_path="EleutherAI/gpt-j-6B", truncation_side="left"),
    optimizer=OptimizerConfig(name="adamw", kwargs=dict(lr=8e-6, betas=(0.9, 0.95), eps=1.0e-8, weight_decay=1.0e-6)),
    scheduler=SchedulerConfig(name="cosine_annealing", kwargs=dict(T_max=10000, eta_min=8e-6)),
    method=PPOConfig(
        name="PPOConfig",
        num_rollouts=64,
        chunk_size=16,
        ppo_epochs=4,
        init_kl_coef=0.05,
        target=6,
        horizon=10000,
        gamma=1,
        lam=0.95,
        cliprange=0.2,
        cliprange_value=0.2,
        vf_coef=1,
        scale_reward="running",
        ref_mean=None,
        ref_std=None,
        cliprange_reward=10,
        gen_kwargs=dict(
            max_new_tokens=128,
            top_k=0,
            top_p=1.0,
            do_sample=True,
        ),
    ),
)


config_name = os.environ.get("CONFIG_NAME")
if config_name == "125M":
    default_config.train.batch_size = 32
    default_config.train.total_steps = 1500
    default_config.train.checkpoint_dir = "checkpoints/ppo_hh_125M"
    default_config.model.model_path = "Dahoas/pythia-125M-static-sft"
    default_config.tokenizer.tokenizer_path = "EleutherAI/gpt-neox-20b"
    default_config.method.num_rollouts = 128
elif config_name == "1B":
    default_config.train.batch_size = 8
    default_config.train.total_steps = 2500
    default_config.optimizer.kwargs["lr"] = 6e-6
    default_config.scheduler.kwargs["eta_min"] = 6e-6
    default_config.train.checkpoint_dir = "checkpoints/ppo_hh_1B"
    default_config.model.model_path = "Dahoas/pythia-1B-static-sft"
    default_config.tokenizer.tokenizer_path = "EleutherAI/gpt-neox-20b"
    default_config.method.chunk_size = 16
elif config_name == "6B":
    default_config.train.batch_size = 4
    default_config.train.seq_length = 512
    default_config.train.total_steps = 6000
    default_config.train.checkpoint_dir = "checkpoints/ppo_hh_6B"
    default_config.model.model_path = "Dahoas/pythia-6B-static-sft"
    default_config.tokenizer.tokenizer_path = "EleutherAI/gpt-neox-20b"
    default_config.method.chunk_size = 16
elif config_name == "20B":
    default_config.train.seq_length = 512
    default_config.train.batch_size = 1
    default_config.train.total_steps = 8000
    default_config.optimizer.kwargs["lr"] = 1e-6
    default_config.scheduler.kwargs["eta_min"] = 1e-6
    default_config.train.checkpoint_dir = "checkpoints/ppo_hh_20B"
    default_config.model.model_path = "EleutherAI/gpt-neox-20b"
    default_config.tokenizer.tokenizer_path = "EleutherAI/gpt-neox-20b"
    default_config.method.num_rollouts = 16
    default_config.method.chunk_size = 4
    default_config.method.ppo_epochs = 2


def prepare_tensor(name: str, input):
    t = client_util.InferInput(name, input.shape, np_to_triton_dtype(input.dtype))
    t.set_data_from_numpy(input)
    return t


def create_reward_fn():  # noqa:  C901
    reward_tokenizer = AutoTokenizer.from_pretrained("gpt2")
    reward_tokenizer.pad_token = reward_tokenizer.eos_token
    reward_tokenizer.truncation_side = "left"
    triton_host = os.environ.get("TRITON_HOST")

    if triton_host:
        triton_url, triton_model = triton_host.split("/")
        client = client_util.InferenceServerClient(url=triton_url, verbose=False)

        def reward_fn(samples, prompts, outputs):
            samples = [s + reward_tokenizer.eos_token for s in samples]
            input = reward_tokenizer(samples, padding=True, max_length=1024)

            mbs = 24
            out = []
            for i in range(math.ceil(len(samples) / mbs)):
                batch_ixs = slice(i * mbs, (i + 1) * mbs)
                input_ids = np.array(input.input_ids[batch_ixs], dtype=np.int32)

                result = client.infer(triton_model, [prepare_tensor("input_ids", input_ids)])
                rewards = result.as_numpy("rewards")
                out.extend(rewards)

            return out

    elif os.environ.get("RANK", "0") == "0":

        class RewardModel(nn.Module):
            def __init__(self, checkpoint_path, eos_token_id):
                super().__init__()
                model = AutoModelForCausalLM.from_pretrained(checkpoint_path)
                self.transformer = model.transformer
                self.v_head = nn.Linear(model.config.n_embd, 1, bias=False)
                self.eos_token_id = eos_token_id

            def forward(self, input_ids):
                states = self.transformer(input_ids)[0]
                rewards = self.v_head(states).squeeze(-1)
                ends = torch.argmax((input_ids == self.eos_token_id).float(), dim=1).view(-1, 1)
                returns = torch.gather(rewards, 1, ends).squeeze(-1)
                return returns

        reward_model = RewardModel("EleutherAI/gpt-j-6B", reward_tokenizer.eos_token_id)
        directory = snapshot_download("Dahoas/gptj-rm-static", revision="676bfd4d")
        for fpath in os.listdir(directory):
            if fpath.endswith(".pt") or fpath.endswith(".bin"):
                checkpoint = os.path.join(directory, fpath)
                break

        reward_model.load_state_dict(torch.load(checkpoint))
        reward_model.eval()
        reward_model.requires_grad_(False)
        reward_device = torch.cuda.device_count() - 1
        reward_model = reward_model.half().to(reward_device)
        reward_batch_size = 48
        delta_reward = True

        def get_reward(samples):
            input = reward_tokenizer(
                samples,
                padding=True,
                truncation=True,
                max_length=reward_tokenizer.max_len_single_sentence,
                return_tensors="pt",
            ).to(reward_device)

            mbs = reward_batch_size
            out = []
            for i in range(math.ceil(len(samples) / mbs)):
                batch_ixs = slice(i * mbs, (i + 1) * mbs)
                input_ids = input.input_ids[batch_ixs]
                rewards = reward_model(input_ids)
                out.extend(rewards)
            return torch.hstack(out)

        def reward_fn(samples, prompts, original_output, **kwargs):
            samples = [s + reward_tokenizer.eos_token for s in samples]
            rewards = get_reward(samples)

            if not delta_reward:
                return rewards

            original_samples = [p + o + reward_tokenizer.eos_token for p, o in zip(prompts, original_output)]
            original_rewards = get_reward(original_samples)
            return rewards - original_rewards

    else:
        reward_fn = True

    return reward_fn


def main(hparams={}):
    config = TRLConfig.update(default_config, hparams)

    dataset = load_dataset("Dahoas/rm-static")
    prompts = [{"prompt": x["prompt"], "original_output": x["chosen"]} for x in dataset["train"]]
    eval_prompts = [{"prompt": x["prompt"], "original_output": x["chosen"]} for x in islice(dataset["test"], 280)]
    reward_fn = create_reward_fn()

    trlx.train(
        prompts=prompts,
        eval_prompts=eval_prompts,
        reward_fn=reward_fn,
        config=config,
        stop_sequences=["Human:", "human:", "Assistant:", "assistant:"],
    )


if __name__ == "__main__":
    hparams = {} if len(sys.argv) == 1 else json.loads(sys.argv[1])
    main(hparams)