Spaces:
Running on Zero
Running on Zero
File size: 11,094 Bytes
a1833c8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 | # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: OpenMDW-1.1
"""CosmosDataLoader — slim orchestrator that wires the four dataflow roles
(DataDistributor -> RawItemProcessor -> SampleBatcher -> BatchCollator) inside
each DataLoader worker. The canonical training dataloader.
"""
from __future__ import annotations
import torch
import torch.utils.data
import numpy as np
from cosmos_framework.utils import log
from cosmos_framework.data.generator.dataflow.base import (
BatchCollator,
DataDistributor,
RawItemProcessor,
SampleBatcher,
)
from cosmos_framework.data.generator.dataflow.batchers import SimpleBatcher
from cosmos_framework.data.generator.dataflow.collators import DefaultBatchCollator
class _DataflowIterableDataset(torch.utils.data.IterableDataset):
"""Wires distributor -> processor -> batcher -> collator inside a worker."""
def __init__(
self,
distributor: DataDistributor,
processor: RawItemProcessor,
batcher: SampleBatcher,
collator: BatchCollator,
dp_rank: int,
dp_world_size: int,
):
super().__init__()
self._distributor = distributor
self._processor = processor
self._batcher = batcher
self._collator = collator
self._dp_rank = dp_rank
self._dp_world_size = dp_world_size
def __iter__(self):
info = torch.utils.data.get_worker_info()
worker_id, num_workers = (info.id, info.num_workers) if info else (0, 1)
raw = self._distributor.stream(self._dp_rank, self._dp_world_size, worker_id, num_workers)
def _processed():
for item in raw:
if isinstance(item, dict):
meta = {k: item.pop(k) for k in list(item) if k.startswith("_dp_")}
else:
meta = {}
s = self._processor.process(item)
if meta and isinstance(s, dict):
s.update(meta)
yield s
for group in self._batcher.batches(_processed()):
has_meta = bool(group) and isinstance(group[0], dict) and "_dp_epoch" in group[0]
if has_meta:
epochs = [s["_dp_epoch"] for s in group]
positions = [s["_dp_stream_pos"] for s in group]
max_epoch = max(epochs)
max_pos = max(positions)
# Resume records (max_epoch, max_pos) and fast-forwards to max_pos+1 —
# bit-for-bit with the legacy collate_batch. That is gap-free only when
# this batch is a single sample (max_batch_size=1, all live recipes) or a
# single-epoch contiguous run (sequential packing). A reordering batcher
# (pool packing) at batch_size>1, or a batch spanning an epoch boundary,
# would leave buffered lower positions unrecorded and skip them on resume.
# Fail loudly rather than silently drop samples in that unsupported combo.
if len(group) > 1:
contiguous = min(epochs) == max_epoch and sorted(positions) == list(
range(min(positions), max_pos + 1)
)
if not contiguous:
raise ValueError(
"Map-style resume cannot safely stamp a multi-sample batch whose "
"_dp_stream_pos values are non-contiguous or span epochs (reordering "
"batcher + batch_size>1). Use max_batch_size=1 with pool packing, a "
"sequential (order-preserving) batcher, or an iterable (non-resumable) "
"source."
)
clean = [{k: v for k, v in s.items() if not k.startswith("_dp_")} for s in group]
batch = self._collator.collate(clean)
batch["sample_worker_id"] = torch.tensor([worker_id] * len(group))
batch["sample_epoch"] = torch.tensor([max_epoch] * len(group))
batch["sample_index"] = torch.tensor([max_pos] * len(group))
else:
batch = self._collator.collate(group)
yield batch
class CosmosDataLoader(torch.utils.data.DataLoader):
"""Public entry point: bring any dataset into training via four roles.
Either pass an explicit ``batcher`` (and optional ``collator``), or pass a
bare ``batch_size=N`` for stock fixed-size batching — the loader then builds
``SimpleBatcher(N)`` + ``DefaultBatchCollator()``. Passing both is an error.
DP coordinates: ``parallel_dims.dp_coord`` > ``torch.distributed`` > (0, 1).
"""
def __init__(
self,
distributor: DataDistributor,
processor: RawItemProcessor,
batcher: SampleBatcher | None = None,
collator: BatchCollator | None = None,
batch_size: int | None = None,
num_workers: int = 0,
prefetch_factor: int | None = None,
persistent_workers: bool = False,
pin_memory: bool = False,
parallel_dims=None,
):
if batch_size is not None and batcher is not None:
raise ValueError(
"Pass either batch_size= (sugar) or an explicit batcher=, not both."
)
if batch_size is None and batcher is None:
raise ValueError("Provide either a batcher= or a batch_size=.")
if batch_size is not None:
batcher = SimpleBatcher(batch_size=batch_size)
if collator is None:
collator = DefaultBatchCollator()
if parallel_dims is not None:
dp_rank, dp_world_size = parallel_dims.dp_coord
elif torch.distributed.is_initialized():
dp_rank = torch.distributed.get_rank()
dp_world_size = torch.distributed.get_world_size()
if dp_world_size > 1:
log.info(
"CosmosDataLoader: using global rank for DP sharding. "
"For FSDP+TP/PP pass parallel_dims= for the correct DP rank.",
rank0_only=True,
)
else:
dp_rank, dp_world_size = 0, 1
dataset = _DataflowIterableDataset(
distributor=distributor,
processor=processor,
batcher=batcher,
collator=collator,
dp_rank=dp_rank,
dp_world_size=dp_world_size,
)
from cosmos_framework.data.generator.dataflow.distributors import MapDistributor
if isinstance(distributor, MapDistributor) and num_workers > 0 and not persistent_workers:
log.info(
"CosmosDataLoader: MapDistributor requires persistent_workers=True for "
"correct stateful resume; overriding to True.",
rank0_only=True,
)
persistent_workers = True
if persistent_workers and num_workers == 0:
log.info(
"CosmosDataLoader: persistent_workers=True ignored because num_workers=0.",
rank0_only=True,
)
persistent_workers = False
loader_kwargs: dict = dict(
num_workers=num_workers,
persistent_workers=persistent_workers,
pin_memory=pin_memory,
)
if num_workers > 0 and prefetch_factor is not None:
loader_kwargs["prefetch_factor"] = prefetch_factor
super().__init__(dataset, batch_size=None, **loader_kwargs)
class JointCosmosDataLoader:
"""Wraps multiple ``CosmosDataLoader`` instances with ratio-based seeded selection.
One output batch = one inner loader, selected deterministically by ratio at each
step. Adds a ``"dataset_name"`` key to every yielded batch so downstream callbacks
can route state updates to the correct inner loader.
Parameters
----------
dataloaders:
``{name: {"dataloader": CosmosDataLoader, "ratio": int}}`` mapping.
Entries with ``ratio <= 0`` are silently skipped.
seed:
Base seed for per-step dataset selection. Step ``i`` uses
``np.random.RandomState(seed + i)`` — fully reproducible on resume via
``set_start_iteration``.
"""
def __init__(
self,
dataloaders: dict,
seed: int = 42,
) -> None:
entries = [
(name, cfg["dataloader"], cfg["ratio"])
for name, cfg in dataloaders.items()
if cfg.get("ratio", 0) > 0
]
if not entries:
raise ValueError("JointCosmosDataLoader: no dataloaders with ratio > 0")
self._names: list[str] = [e[0] for e in entries]
if "global_id" in self._names:
raise ValueError(
"JointCosmosDataLoader: dataset name 'global_id' is reserved "
"by the checkpoint state format; use a different name."
)
self._loaders: list[CosmosDataLoader] = [e[1] for e in entries]
ratios = np.array([e[2] for e in entries], dtype=float)
self._probs: np.ndarray = ratios / ratios.sum()
self._seed = seed
self._global_id = 0
# Iterators are created lazily on the first __iter__ call so that
# DataLoaderStateCallback.load_state_dict can install resume env vars
# before workers are spawned (for num_workers > 0, iter(DataLoader)
# forks workers immediately; env vars must be set in the parent first).
self._iterators: list | None = None
total = ratios.sum()
lines = [f"JointCosmosDataLoader: {len(self._names)} streams"]
for name, ratio in zip(self._names, ratios):
lines.append(f" {name}: ratio={ratio:.4g} ({ratio / total:.1%})")
log.info("\n".join(lines))
def set_start_iteration(self, iteration: int) -> None:
"""Restore deterministic selection sequence after checkpoint resume.
Called by ``JointCosmosDataLoaderStateCallback.load_state_dict`` and by the
trainer (if present) via ``hasattr`` guard.
"""
self._global_id = iteration
def __iter__(self):
# Lazy init: create iterators here (not in __init__) so that
# load_state_dict can set resume env vars before workers fork.
if self._iterators is None:
self._iterators = [iter(loader) for loader in self._loaders]
while True:
rng = np.random.RandomState(self._seed + self._global_id)
idx = int(rng.choice(len(self._loaders), p=self._probs))
try:
batch = next(self._iterators[idx])
except StopIteration:
# Inner CosmosDataLoaders are infinite; this guard handles
# the unlikely case of a finite IterableDataset inner source.
self._iterators[idx] = iter(self._loaders[idx])
batch = next(self._iterators[idx])
batch["dataset_name"] = self._names[idx]
self._global_id += 1
yield batch
|