ARBS / training /pretrain.py
CLIWorks's picture
Upload folder using huggingface_hub
d8bc908 verified
"""Unified multi-modal ARB pure-ternary pre-trainer.
Supports text, code, image, audio, and video modalities with weighted mixing,
checkpoint resume, and packed ternary state updates. Core pretraining freezes
all IEEE-float parameters; LoRA/AdamW paths live under ``training/finetuning``.
Usage:
# Phase 1a — Text pre-training smoke test (100M tokens on RTX 6000 Pro)
python training/pretrain.py --text-data training/data/tinyshakespeare.txt \\
--text-weight 1.0 --steps 50000 --batch 8 --ctx 1024
# Phase 1b — Full text + code pre-training
python training/pretrain.py --text-weight 0.95 --code-weight 0.05 \\
--steps 1000000 --batch 16 --ctx 2048
# Phase 2 — Add vision (freeze text, train vision adapters)
python training/pretrain.py --resume models/checkpoints/phase1b/best.pt \\
--image-weight 0.3 --text-weight 1.0
# Phase 3 — Add audio
python training/pretrain.py --resume models/checkpoints/phase2/best.pt \\
--audio-weight 0.2 --text-weight 1.0
# Phase 4 — Add video
python training/pretrain.py --resume models/checkpoints/phase3/best.pt \\
--video-weight 0.1 --text-weight 1.0
# Smoke test (1 step, CPU)
python training/pretrain.py --steps 1 --batch 1 --ctx 4 --cpu --no-save
"""
import argparse, os, random, sys, time
from dataclasses import dataclass
from pathlib import Path
from typing import Optional
import torch
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm
sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
from arbitor import ARBModel
from arbitor.config import CTX
from arbitor.kernel.ternary_audit import audit_model, format_audit, freeze_float_parameters, trainable_parameters
from arbitor.kernel.ternary_scale import TScaleType
from training.data import (
FineWebStream, FineWebConfig,
StarCoderStream, StarCoderConfig,
CC12MStream, CC12MConfig,
LibriSpeechStream, LibriSpeechConfig,
WebVidStream, WebVidConfig,
)
@dataclass
class TrainConfig:
steps: int = 5000
batch: int = 8
ctx: int = min(CTX, 1024)
accum: int = 1
tscale_type: str = "T32"
backend: str = "triton"
freeze_text: bool = False
freeze_vision: bool = False
freeze_audio: bool = False
freeze_video: bool = False
enable_vq: bool = True
enable_graph: bool = True
enable_moe: bool = True
enable_attention: bool = True
enable_output_router: bool = False
text_weight: float = 1.0
code_weight: float = 0.0
image_weight: float = 0.0
audio_weight: float = 0.0
video_weight: float = 0.0
text_data: Optional[str] = None
data_dir: str = "training/data"
out_dir: str = "models/checkpoints"
run: str = "pretrain"
resume: Optional[str] = None
no_save: bool = False
save_interval: int = 5000
eval_interval: int = 500
log_interval: int = 10
seed: int = 42
cpu: bool = False
max_moe_iters: int = 4
class LocalByteStream:
"""Small local byte stream for smoke tests and phase-1 text bootstrap."""
def __init__(self, path: str, ctx: int, batch_size: int):
self.path = Path(path)
self.ctx = ctx
self.batch_size = batch_size
def _load(self) -> torch.Tensor:
if not self.path.exists():
raise FileNotFoundError(f"Local text data not found: {self.path}")
if self.path.suffix == ".pt":
data = torch.load(self.path, weights_only=True).long().cpu()
else:
data = torch.tensor(list(self.path.read_bytes()), dtype=torch.long)
if data.numel() <= self.ctx + 1:
raise ValueError(f"Local text data has {data.numel()} tokens but ctx={self.ctx}")
return data
def batches(self):
data = self._load()
while True:
ix = torch.randint(0, data.numel() - self.ctx - 1, (self.batch_size,))
x = torch.stack([data[i : i + self.ctx] for i in ix])
yield x, x[:, 3:].contiguous()
def build_model(cfg: TrainConfig, device: torch.device):
model = ARBModel(
enable_image=cfg.image_weight > 0,
enable_audio=cfg.audio_weight > 0,
enable_vq=cfg.enable_vq,
enable_graph=cfg.enable_graph,
enable_memory_modules=False,
enable_moe=cfg.enable_moe,
max_moe_iters=cfg.max_moe_iters,
tscale_type=getattr(TScaleType, cfg.tscale_type.upper(), TScaleType.T32),
enable_attention=cfg.enable_attention and cfg.enable_graph and cfg.enable_vq,
enable_output_router=cfg.enable_output_router,
enable_video_output=cfg.video_weight > 0,
enable_talker_output=cfg.audio_weight > 0,
).to(device)
freeze_float_parameters(model)
print(format_audit(audit_model(model)))
return model
def create_streams(cfg: TrainConfig):
streams = {}
if cfg.text_weight > 0:
if cfg.text_data:
streams['text'] = LocalByteStream(cfg.text_data, ctx=cfg.ctx, batch_size=cfg.batch)
else:
streams['text'] = FineWebStream(FineWebConfig(ctx=cfg.ctx, batch_size=cfg.batch))
if cfg.code_weight > 0:
streams['code'] = StarCoderStream(StarCoderConfig(ctx=cfg.ctx, batch_size=cfg.batch))
if cfg.image_weight > 0:
streams['image'] = CC12MStream(CC12MConfig(batch_size=max(1, cfg.batch // 2)))
if cfg.audio_weight > 0:
streams['audio'] = LibriSpeechStream(LibriSpeechConfig(batch_size=max(1, cfg.batch // 2)))
if cfg.video_weight > 0:
streams['video'] = WebVidStream(WebVidConfig(batch_size=max(1, cfg.batch // 4)))
return streams
def sample_modality(cfg: TrainConfig) -> str:
weights = {
'text': cfg.text_weight,
'code': cfg.code_weight,
'image': cfg.image_weight,
'audio': cfg.audio_weight,
'video': cfg.video_weight,
}
active = {k: v for k, v in weights.items() if v > 0}
if not active:
return 'text'
total = sum(active.values())
r = random.random() * total
cumulative = 0.0
for k, v in active.items():
cumulative += v
if r <= cumulative:
return k
return list(active.keys())[-1]
def compute_loss(model, modality: str, batch, device):
if modality in ('text', 'code'):
x = batch[0].to(device, non_blocking=True)
targets = x[:, 3:].contiguous()
_, losses, _, _ = model(x, targets=targets)
return losses.total
if modality == 'image':
images, captions = batch
images = images.to(device, non_blocking=True)
targets = captions.to(device, non_blocking=True)
if targets.size(1) < 4:
raise ValueError("Image caption batch must contain at least 4 byte tokens")
_, losses, _, _ = model(x=targets, images=images, targets=targets[:, 3:])
return losses.total
if modality == 'audio':
waves, vq_targets = batch
waves = waves.to(device, non_blocking=True)
targets = vq_targets.to(device, non_blocking=True)
if targets.size(1) < 4:
raise ValueError("Audio token batch must contain at least 4 tokens")
_, losses, _, _ = model(x=targets, audio=waves, targets=targets[:, 3:])
return losses.total
if modality == 'video':
text_tokens, latent_targets = batch
text_tokens = text_tokens.to(device, non_blocking=True)
latents = latent_targets.to(device, non_blocking=True)
embedded = model.embedding(text_tokens)
seq_out = model.multimodal_sequencer({'text': embedded})
rel = seq_out['text']
pred = model.video_head(rel)
latents = match_latents(latents, pred)
loss = torch.nn.functional.mse_loss(pred, latents)
return loss
raise ValueError(f"Unknown modality: {modality}")
def match_latents(target: torch.Tensor, pred: torch.Tensor) -> torch.Tensor:
if target.shape[0] == 1 and pred.shape[0] > 1:
target = target.expand(pred.shape[0], -1, -1, -1, -1).contiguous()
if target.shape[1] != pred.shape[1]:
if target.shape[1] > pred.shape[1]:
target = target[:, :pred.shape[1]]
else:
pad = target.new_zeros(target.shape[0], pred.shape[1] - target.shape[1], *target.shape[2:])
target = torch.cat([target, pad], dim=1)
if target.shape[2:] != pred.shape[2:]:
target = torch.nn.functional.interpolate(
target, size=pred.shape[2:], mode="trilinear", align_corners=False
)
return target
def save_checkpoint(path: Path, model, step: int, loss: float, cfg: TrainConfig):
if cfg.no_save:
return
path.parent.mkdir(parents=True, exist_ok=True)
state = {
'step': step,
'loss': loss,
'model': model.state_dict(),
'config': vars(cfg),
}
torch.save(state, path)
def load_checkpoint(path: str, model, device):
ckpt_path = Path(path)
if ckpt_path.is_dir():
if (ckpt_path / "latest.pt").exists():
ckpt_path = ckpt_path / "latest.pt"
elif (ckpt_path / "best.pt").exists():
ckpt_path = ckpt_path / "best.pt"
elif (ckpt_path / "final.pt").exists():
ckpt_path = ckpt_path / "final.pt"
state = torch.load(ckpt_path, map_location=device, weights_only=True)
missing, unexpected = model.load_state_dict(state['model'], strict=False)
if missing or unexpected:
print(
"Checkpoint loaded with architecture drift: "
f"{len(missing)} missing keys, {len(unexpected)} unexpected keys"
)
return state.get('step', 0), state.get('loss', float('inf'))
def train(cfg: TrainConfig):
torch.manual_seed(cfg.seed)
random.seed(cfg.seed)
os.environ["ARB_TERNARY_BACKEND"] = cfg.backend
if cfg.backend == "tilelang" and os.environ.get("ARB_TILELANG_TRAINING", "0").lower() not in {"1", "true", "yes"}:
raise ValueError(
"TileLang BigInt training is unfinished and disabled by default. "
"Use --backend triton for production training."
)
device = torch.device("cuda" if torch.cuda.is_available() and not cfg.cpu else "cpu")
print(f"Device: {device}")
print(f"Ternary backend: {cfg.backend}")
model = build_model(cfg, device)
streams = create_streams(cfg)
if not streams:
raise ValueError("No active training streams. Set at least one modality weight above 0.")
print(f"Active modalities: {', '.join(streams.keys())}")
params = trainable_parameters(model)
if params:
raise RuntimeError(
"Pure ternary pretrain found trainable torch Parameters after freeze. "
"Use training/finetuning for LoRA adapters."
)
start_step = 0
if cfg.resume:
ckpt_path = Path(cfg.resume)
if ckpt_path.exists():
start_step, _ = load_checkpoint(str(ckpt_path), model, device)
print(f"Resumed from step {start_step}")
run_dir = Path(cfg.out_dir) / cfg.run
writer = SummaryWriter(str(run_dir))
model.train()
stream_iters = {k: s.batches() for k, s in streams.items()}
best_loss = float('inf')
last_loss = float('inf')
step = start_step
accum_loss = 0.0
accum_steps = 0
start_time = time.perf_counter()
pbar = tqdm(range(start_step, cfg.steps), desc="train", dynamic_ncols=True,
initial=start_step, total=cfg.steps)
for step in pbar:
modality = sample_modality(cfg)
stream = stream_iters.get(modality)
if stream is None:
continue
try:
batch = next(stream)
except StopIteration:
stream_iters[modality] = streams[modality].batches()
batch = next(stream_iters[modality])
model.zero_grad(set_to_none=True)
raw_loss = compute_loss(model, modality, batch, device)
last_loss = raw_loss.detach().item()
loss = raw_loss
if cfg.accum > 1:
loss = raw_loss / cfg.accum
if not torch.isfinite(loss).all():
raise FloatingPointError(f"Non-finite {modality} pretraining loss; aborting before ternary update")
model.prepare_ternary_backward(loss.detach(), update_scales=True)
loss.backward()
accum_loss += raw_loss.detach().item()
accum_steps += 1
if accum_steps >= cfg.accum:
model._ternary_update_memory(accum_threshold=3, update_scales=True,
loss_signal=raw_loss.detach())
model.zero_grad(set_to_none=True)
report_step = step + 1
if cfg.log_interval and (step + 1) % cfg.log_interval == 0:
avg = accum_loss / cfg.accum
writer.add_scalar("loss/train", avg, step)
pbar.set_postfix(loss=f"{avg:.4f}", mod=modality)
print(f"step {report_step:>6d} loss={avg:.4f} mod={modality}")
if cfg.eval_interval and (step + 1) % cfg.eval_interval == 0:
avg_loss = accum_loss / cfg.accum
if avg_loss < best_loss:
best_loss = avg_loss
save_checkpoint(run_dir / "best.pt", model, step, avg_loss, cfg)
print(f"step {report_step:>6d} loss={avg_loss:.4f} mod={modality}")
if cfg.save_interval and (step + 1) % cfg.save_interval == 0:
save_checkpoint(run_dir / "latest.pt", model, step, accum_loss / cfg.accum, cfg)
accum_loss = 0.0
accum_steps = 0
total_time = time.perf_counter() - start_time
print(f"Training complete. {cfg.steps - start_step} steps in {total_time / 3600:.1f}h")
save_checkpoint(run_dir / "final.pt", model, step, last_loss, cfg)
writer.close()
def parse_args():
p = argparse.ArgumentParser(description="Unified ARB multi-modal pre-trainer")
p.add_argument("--steps", type=int, default=5000)
p.add_argument("--batch", type=int, default=8)
p.add_argument("--ctx", type=int, default=min(CTX, 1024))
p.add_argument("--accum", type=int, default=1)
p.add_argument("--tscale-type", type=str, default="T32")
p.add_argument("--backend", choices=("triton", "torch", "auto", "tilelang"), default="triton",
help="Training backend. Triton is the production BigInt ternary path.")
p.add_argument("--no-save", action="store_true")
p.add_argument("--save-interval", type=int, default=5000)
p.add_argument("--eval-interval", type=int, default=500)
p.add_argument("--log-interval", type=int, default=10)
p.add_argument("--seed", type=int, default=42)
p.add_argument("--cpu", action="store_true")
p.add_argument("--max-moe-iters", type=int, default=4)
p.add_argument("--out-dir", type=str, default="models/checkpoints")
p.add_argument("--run", type=str, default="pretrain")
p.add_argument("--resume", type=str, default=None,
help="Path to checkpoint .pt or directory with latest.pt")
p.add_argument("--freeze-text", action="store_true", help=argparse.SUPPRESS)
p.add_argument("--freeze-vision", action="store_true", help=argparse.SUPPRESS)
p.add_argument("--freeze-audio", action="store_true", help=argparse.SUPPRESS)
p.add_argument("--freeze-video", action="store_true", help=argparse.SUPPRESS)
p.add_argument("--no-vq", dest="enable_vq", action="store_false")
p.add_argument("--no-graph", dest="enable_graph", action="store_false")
p.add_argument("--no-moe", dest="enable_moe", action="store_false")
p.add_argument("--no-attention", dest="enable_attention", action="store_false")
p.add_argument("--enable-output-router", action="store_true", default=False)
p.set_defaults(enable_vq=True, enable_graph=True, enable_moe=True, enable_attention=True)
p.add_argument("--text-weight", type=float, default=1.0)
p.add_argument("--code-weight", type=float, default=0.0)
p.add_argument("--image-weight", type=float, default=0.0)
p.add_argument("--audio-weight", type=float, default=0.0)
p.add_argument("--video-weight", type=float, default=0.0)
p.add_argument("--text-data", type=str, default=None,
help="Optional local .txt/.pt byte data for text pretraining smoke/bootstrap")
return p.parse_args()
if __name__ == "__main__":
cfg = TrainConfig(**vars(parse_args()))
train(cfg)