finetrainers-ui/run_trainer.py at main · neph1/finetrainers-ui · GitHub

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
import os
import signal
import subprocess

import psutil

from config import Config

class RunTrainer:

    def __init__(self):
        self.running = False
        self.process = None
        pass

    def run(self, config: Config, finetrainers_path: str, log_file: str):

        assert finetrainers_path, "Path to finetrainers is required"
        assert config.get('pretrained_model_name_or_path'), "pretrained_model_name_or_path required"

        parallel_backend = config.get('parallel_backend')
        parallel_cmd = ["--parallel_backend", parallel_backend,
                            "--pp_degree", config.get('pp_degree'),
                            "--dp_degree", config.get('dp_degree'),
                            "--dp_shards", config.get('dp_shards'),
                            "--cp_degree", config.get('cp_degree'),
                            "--tp_degree", config.get('tp_degree')]

        model_cmd = ["--model_name", config.get('model_name'),
                     "--pretrained_model_name_or_path", config.get('pretrained_model_name_or_path'),
                     "--text_encoder_dtype", config.get('text_encoder_dtype'),
                     "--text_encoder_2_dtype", config.get('text_encoder_2_dtype'),
                     "--text_encoder_3_dtype", config.get('text_encoder_3_dtype'),
                     "--transformer_dtype", config.get('transformer_dtype'),
                     "--vae_dtype", config.get('vae_dtype')]
        if config.get('text_encoder_id'):
            model_cmd += ["--text_encoder_id", config.get('text_encoder_id')]
        if config.get('text_encoder_2_id'):
            model_cmd += ["--text_encoder_2_id", config.get('text_encoder_2_id')]
        if config.get('text_encoder_3_id'):
            model_cmd += ["--text_encoder_3_id", config.get('text_encoder_3_id')]
        if config.get('transformer_id'):
            model_cmd += ["--transformer_id", config.get('transformer_id')]
        if config.get('vae_id'):
            model_cmd += ["--vae_id", config.get('vae_id')]
        if config.get('tokenizer_id'):
            model_cmd += ["--tokenizer_id", config.get('tokenizer_id')]
        if config.get('tokenizer_2_id'):
            model_cmd += ["--tokenizer_2_id", config.get('tokenizer_2_id')]
        if config.get('tokenizer_3_id'):
            model_cmd += ["--tokenizer_3_id", config.get('tokenizer_3_id')]

        if config.get('layerwise_upcasting_modules') != 'none':
            model_cmd +=["--layerwise_upcasting_modules", config.get('layerwise_upcasting_modules'),
            "--layerwise_upcasting_storage_dtype", config.get('layerwise_upcasting_storage_dtype'),
            "--layerwise_upcasting_skip_modules_pattern", config.get('layerwise_upcasting_skip_modules_pattern')]

        dataset_cmd = ["--dataset_config", config.get('dataset_config'),
                    "--caption_dropout_p", config.get('caption_dropout_p'),
                    "--caption_dropout_technique", config.get('caption_dropout_technique'),
                    "--enable_precomputation" if config.get('enable_precomputation') else '',
                    "--precomputation_items", config.get('precomputation_items'),
                    "--precomputation_dir" if config.get('precomputation_dir') else '',
                    "--precomputation_once" if config.get('precomputation_once') else '']

        dataloader_cmd = ["--dataloader_num_workers", config.get('dataloader_num_workers')]

        # TODO: seems to have changed, need full options
        #diffusion_cmd = [config.get('diffusion_options')]

        training_cmd = ["--training_type", config.get('training_type'),
                "--seed", config.get('seed'),
                "--batch_size", config.get('batch_size'),
                "--train_steps", config.get('train_steps')]
        training_cmd += config.get('target_modules').split(' ')
        training_cmd += ["--gradient_accumulation_steps", config.get('gradient_accumulation_steps'),
                '--gradient_checkpointing' if config.get('gradient_checkpointing') else '',
                "--checkpointing_steps", config.get('checkpointing_steps'),
                "--checkpointing_limit", config.get('checkpointing_limit'),
                '--enable_slicing' if config.get('enable_slicing') else '',
                '--enable_tiling' if config.get('enable_tiling') else '']
        if config.get('enable_model_cpu_offload'):
            training_cmd += ["--enable_model_cpu_offload"]

        if config.get('resume_from_checkpoint'):
            training_cmd += ["--resume_from_checkpoint", config.get('resume_from_checkpoint')]

        optimizer_cmd = ["--optimizer", config.get('optimizer'),
                 "--lr", config.get('lr'),
                 "--lr_scheduler", config.get('lr_scheduler'),
                 "--lr_warmup_steps", config.get('lr_warmup_steps'),
                 "--lr_num_cycles", config.get('lr_num_cycles'),
                 "--beta1", config.get('beta1'),
                 "--beta2", config.get('beta2'),
                 "--weight_decay", config.get('weight_decay'),
                 "--epsilon", config.get('epsilon'),
                 "--max_grad_norm", config.get('max_grad_norm'),
                 '--use_8bit_bnb' if config.get('use_8bit_bnb') else '']

        validation_cmd = ["--validation_dataset_file" if config.get('validation_dataset_file') else '',
                  "--num_validation_videos", config.get('num_validation_videos'),
                  "--validation_steps", config.get('validation_steps')]

        control_cmd = ["--rank", config.get('rank'),
                "--lora_alpha", config.get('lora_alpha'),
                "--control_type", config.get('control_type'),
                "--frame_conditioning_index", config.get('frame_conditioning_index'),
                "--frame_conditioning_type", config.get('frame_conditioning_type')]

        miscellaneous_cmd = ["--tracker_name", config.get('tracker_name'),
                     "--output_dir", config.get('output_dir'),
                     "--nccl_timeout", config.get('nccl_timeout'),
                     "--report_to", config.get('report_to')]

        pre_command = ''
        num_gpus = config.get('num_gpus')
        address = config.get('master_address')
        port = config.get('master_port')
        if parallel_backend == 'accelerate':
            os.environ['WORLD_SIZE'] = f'{num_gpus}'
            os.environ['RANK'] = config.get('gpu_ids')
            os.environ['MASTER_ADDR'] = address
            os.environ['MASTER_PORT'] = f'{port}'
            pre_command = ["accelerate", "launch", "--config_file", f"{finetrainers_path}/accelerate_configs/{config.get('accelerate_config')}", "--gpu_ids", config.get('gpu_ids')]
        elif parallel_backend == 'ptd':
            pre_command = ["torchrun", "--standalone", "--nnodes", num_gpus, "--nproc_per_node", config.get('nproc_per_node'), "--rdzv_backend", "c10d", "--rdzv_endpoint", f"{address}:{port}"]
        cmd = pre_command + [f"{finetrainers_path}/train.py"] + parallel_cmd + model_cmd + dataset_cmd + dataloader_cmd + training_cmd + optimizer_cmd + validation_cmd + miscellaneous_cmd + control_cmd
        fixed_cmd = []
        for i in range(len(cmd)):
            if cmd[i] != '':
                fixed_cmd.append(f"{cmd[i]}")
        print(' '.join(fixed_cmd))
        self.running = True
        with open(log_file, "w") as output_file:
            self.process = subprocess.Popen(fixed_cmd, shell=False, stdout=output_file, stderr=output_file, text=True, preexec_fn=os.setsid)
            self.process.communicate()
            return self.process

        return "Unknown result"

    def stop(self):
        try:
            self.running = False
            if self.process:
                os.killpg(os.getpgid(self.process.pid), signal.SIGTERM)
                self.terminate_process_tree(self.process.pid)
        except Exception as e:
            return f"Error stopping training: {e}"
        finally:
            self.process.wait()
        return "Training forcibly stopped"

    def terminate_process_tree(pid):
        try:
            parent = psutil.Process(pid)
            children = parent.children(recursive=True)  # Get child processes
            for child in children:
                child.terminate()
            parent.terminate()
        except psutil.NoSuchProcess:
            pass