Repository navigation
Expand file tree
/
Copy pathtrain_chronos.py
More file actions
165 lines (148 loc) · 7.01 KB
/
Copy pathtrain_chronos.py
File metadata and controls
165 lines (148 loc) · 7.01 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
163
164
165
"""
train_chronos.py — Phase 1 training entry point for Project Chronos.
Usage:
python train_chronos.py \
--data_path /path/to/pretrain.jsonl \
--hidden_size 512 \
--num_hidden_layers 8 \
--num_experts 4 \
--lambda_temporal 1e-3 \
--epochs 2 \
--device cuda:0
"""
import sys
import argparse
from torch.utils.data import DataLoader
from chronos.backend import resolve_training_device
from chronos.model.kernel import Logger, setup_seed
from chronos.data.flexible_dataset import (
FlexibleDataset as PretrainDataset,
MultimodalStreamingDataset,
dataset_modality_summary,
multimodal_collate,
)
from chronos.trainer.chronos_trainer import ChronosTrainer
from chronos.trainer.stage_utils import add_topology_args, build_pretrain_config, load_stage_tokenizer
from chronos.model.checkpoint import save_state_dict_with_config
from chronos.trainer.device_utils import configure_cpu_threads, dataloader_kwargs, runtime_summary
def parse_args():
p = argparse.ArgumentParser(description="Project Chronos Pretraining")
p.add_argument("--data_path", type=str, required=True)
p.add_argument("--save_dir", type=str, default="./out")
p.add_argument("--epochs", type=int, default=2)
p.add_argument("--batch_size", type=int, default=16)
p.add_argument("--learning_rate", type=float, default=5e-4)
p.add_argument("--device", type=str, default="auto")
p.add_argument("--dtype", type=str, default="auto")
p.add_argument("--accumulation_steps", type=int, default=8)
p.add_argument("--grad_clip", type=float, default=1.0)
p.add_argument("--log_interval", type=int, default=100)
p.add_argument("--save_interval", type=int, default=1000)
p.add_argument("--max_seq_len", type=int, default=512)
p.add_argument("--num_workers", default="auto")
p.add_argument("--cpu_threads", default="auto")
p.add_argument("--cpu_budget_percent", default=100, type=float)
add_topology_args(p, defaults=True)
# Loss coefficients
p.add_argument("--lambda_balance", type=float, default=5e-4)
p.add_argument("--lambda_temporal", type=float, default=3e-3)
p.add_argument("--lambda_lookahead", type=float, default=0.1)
p.add_argument("--lambda_lookahead_topk", type=float, default=0.15)
p.add_argument("--lambda_lookahead_union", type=float, default=0.05)
p.add_argument("--lambda_router_locality", type=float, default=0.02)
p.add_argument("--lambda_cluster_locality", type=float, default=0.3,
help="cluster-locality constraint weight (default 0.3, measured; 0 = off)")
p.add_argument("--cluster_locality_clusters", type=int, default=0,
help="expert clusters for the constraint (0 = ceil(E / 2k))")
p.add_argument("--cluster_locality_window", type=int, default=32,
help="tokens per locality window (0 = whole sequence)")
p.add_argument("--fallback_mask_prob", type=float, default=0.08,
help="Randomly mask routed experts during training to teach the shared offload fallback.")
# VRAM budget
p.add_argument("--vram_budget_gb", type=float, default=4.0)
# Pipeline compat: accept but ignore args that other Stage scripts take.
# The 6-stage Web UI invokes every stage with the same arg signature.
p.add_argument("--steps", type=int, default=None,
help="If set, cap total steps (overrides epochs·len(loader)).")
p.add_argument("--from_weight", type=str, default="",
help="Ignored for Stage 1 (pretrain). Present for pipeline parity.")
return p.parse_args()
def main():
args = parse_args()
selected_backend, resolved_device = resolve_training_device(args.device)
args.device = resolved_device
threads = configure_cpu_threads(args.cpu_threads, budget_percent=args.cpu_budget_percent)
setup_seed(42)
config = build_pretrain_config(args)
modality_summary = dataset_modality_summary(args.data_path)
if config.auto_detect_modalities:
config.modalities = modality_summary["detected_modalities"]
tokenizer = load_stage_tokenizer(args, config)
if config.modalities:
dataset = MultimodalStreamingDataset(
args.data_path, tokenizer, max_length=args.max_seq_len,
image_size=config.image_size,
audio_sample_rate=config.audio_sample_rate,
max_audio_seconds=config.max_audio_seconds,
media_error_policy=config.media_error_policy,
)
else:
dataset = PretrainDataset(args.data_path, tokenizer, max_length=args.max_seq_len)
loader = DataLoader(
dataset,
batch_size=args.batch_size,
collate_fn=multimodal_collate if config.modalities else None,
**dataloader_kwargs(device=args.device, num_workers=args.num_workers, shuffle=True),
)
if selected_backend == "mlx":
from chronos.mlx.training import run_mlx_stage
rt = runtime_summary(args.device, args.dtype)
Logger(
f"Training backend: {selected_backend} device={args.device} "
f"dtype={rt.dtype} autocast={rt.autocast}"
)
Logger(f"CPU threads: {threads} DataLoader workers: {loader.num_workers}")
run_mlx_stage(
stage="pretrain",
config=config,
checkpoint_path=None,
save_dir=args.save_dir,
loader=loader,
args=args,
)
return
trainer = ChronosTrainer(config, args)
total_params = sum(p.numel() for p in trainer.model.parameters()) / 1e6
Logger(f"Chronos model: {total_params:.2f}M params | "
f"experts={config.num_experts} shared={config.num_shared_experts} "
f"lookahead={config.lookahead_steps} "
f"λ1={config.lambda_balance} λ2={config.lambda_temporal}")
Logger(f"Dataset modalities: {modality_summary}")
rt = runtime_summary(args.device, args.dtype)
Logger(
f"Training backend: {selected_backend} device={args.device} "
f"dtype={rt.dtype} autocast={rt.autocast}"
)
Logger(f"CPU threads: {threads} DataLoader workers: {loader.num_workers}")
iters = len(loader) if args.steps is None else min(int(args.steps), len(loader))
global_step = 0
for epoch in range(args.epochs):
trainer.train_epoch(epoch, loader, iters, start_step=global_step, max_steps=args.steps)
global_step += iters
if args.steps is not None and global_step >= int(args.steps):
Logger(f"Reached --steps={args.steps}, stopping.")
break
Logger("Training complete.")
# Always save a final checkpoint at the conventional path so that
# downstream stages (sft/dpo/orpo/grpo/distill) can pick it up via
# --from_weight chronos.
import os
os.makedirs(args.save_dir, exist_ok=True)
ckp = os.path.join(args.save_dir, f"chronos_{config.hidden_size}_moe.pth")
save_state_dict_with_config(
trainer.model, ckp, config, stage="chronos",
extra={"dataset_modality_summary": modality_summary},
)
Logger(f"Saved → {ckp}")
if __name__ == "__main__":
main()