code-HFD / src /psi /trainers /distributed_flow_sampling.py
Tinghong-Ye22's picture
Add files using upload-large-folder tool
06f9561 verified
Raw History Blame Contribute Delete
15.1 kB
"""Deterministic distributed random sampling for flow-matching training.
The normal PyTorch CUDA RNG is process-local. Seeding every distributed rank
with the same value therefore gives equal random draws whenever ranks execute
the same sequence of random operations. This module avoids that coupling and
also makes the intended global-batch coverage explicit:
* Gaussian action noise uses a rank-specific stream.
* Sigma uses one jittered stratum per sample in the *global* batch, followed by
a deterministic global permutation.
* RTC delay is either the deployment delay or a globally balanced draw from an
explicit categorical distribution.
Sampling is keyed by ``(seed, draw_index, component[, rank])`` instead of by an
opaque long-lived generator state. Saving ``state_dict()`` after a batch and
loading it on resume reproduces the next batch exactly. Register an instance
with ``Accelerator.register_for_checkpointing`` so Accelerate saves that small
state next to the model and optimizer state.
"""
from __future__ import annotations
from dataclasses import dataclass
import math
from typing import Any, Literal, Mapping, Sequence
import torch
_STATE_VERSION = 1
_MASK64 = (1 << 64) - 1
_MAX_TORCH_SEED = (1 << 63) - 1
# Component identifiers are deliberately fixed constants. Changing one
# component's implementation must not shift the streams used by the others.
_SIGMA_JITTER_COMPONENT = 0x5349474D415F4A49
_SIGMA_PERM_COMPONENT = 0x5349474D415F504D
_DELAY_OFFSET_COMPONENT = 0x44454C41595F4F46
_DELAY_PERM_COMPONENT = 0x44454C41595F504D
_NOISE_COMPONENT = 0x4E4F4953455F524B
def _splitmix64(value: int) -> int:
"""Return a stable 64-bit mix without relying on Python's salted hash."""
value = (value + 0x9E3779B97F4A7C15) & _MASK64
value = ((value ^ (value >> 30)) * 0xBF58476D1CE4E5B9) & _MASK64
value = ((value ^ (value >> 27)) * 0x94D049BB133111EB) & _MASK64
return (value ^ (value >> 31)) & _MASK64
def _component_seed(
base_seed: int,
draw_index: int,
component: int,
*,
rank: int | None = None,
) -> int:
"""Derive a torch-compatible seed for one component and one draw."""
value = _splitmix64(int(base_seed) & _MASK64)
value ^= _splitmix64(int(draw_index) & _MASK64)
value ^= _splitmix64(int(component) & _MASK64)
if rank is not None:
value ^= _splitmix64(int(rank) & _MASK64)
return _splitmix64(value) & _MAX_TORCH_SEED
@dataclass(frozen=True)
class DelaySamplingConfig:
"""RTC delay policy.
``fixed`` trains every example with the delay used at deployment.
``balanced`` uses systematic categorical sampling over the global batch.
For a requested probability ``p``, the number of examples is always close
to ``p * global_batch_size`` instead of being allowed to clump as it would
under independent ``randint`` draws.
"""
mode: Literal["fixed", "balanced"]
deployment_delay: int | None = None
values: tuple[int, ...] = ()
probabilities: tuple[float, ...] = ()
def __post_init__(self) -> None:
object.__setattr__(self, "values", tuple(int(value) for value in self.values))
object.__setattr__(
self,
"probabilities",
tuple(float(probability) for probability in self.probabilities),
)
if self.mode == "fixed":
if self.deployment_delay is None or int(self.deployment_delay) < 0:
raise ValueError("fixed RTC delay requires a non-negative deployment_delay")
if self.values or self.probabilities:
raise ValueError("fixed RTC delay must not also define values/probabilities")
return
if self.mode != "balanced":
raise ValueError(f"unknown RTC delay sampling mode: {self.mode!r}")
if self.deployment_delay is not None:
raise ValueError("balanced RTC delay must not define deployment_delay")
if not self.values:
raise ValueError("balanced RTC delay requires at least one delay value")
if any(value < 0 for value in self.values):
raise ValueError("RTC delay values must be non-negative")
if len(set(self.values)) != len(self.values):
raise ValueError("RTC delay values must be unique")
if self.probabilities and len(self.probabilities) != len(self.values):
raise ValueError("RTC delay values and probabilities must have equal length")
probabilities = self.normalized_probabilities()
if any(not math.isfinite(probability) or probability < 0 for probability in probabilities):
raise ValueError("RTC delay probabilities must be finite and non-negative")
if sum(probabilities) <= 0:
raise ValueError("at least one RTC delay probability must be positive")
@classmethod
def fixed(cls, deployment_delay: int) -> "DelaySamplingConfig":
return cls(mode="fixed", deployment_delay=int(deployment_delay))
@classmethod
def balanced(
cls,
values: Sequence[int],
probabilities: Sequence[float] | None = None,
) -> "DelaySamplingConfig":
return cls(
mode="balanced",
values=tuple(int(value) for value in values),
probabilities=tuple(float(value) for value in probabilities or ()),
)
@classmethod
def uniform(cls, max_delay: int) -> "DelaySamplingConfig":
"""Balance all valid training delays ``0 .. max_delay - 1``."""
if int(max_delay) <= 0:
raise ValueError("max_delay must be positive")
return cls.balanced(tuple(range(int(max_delay))))
def normalized_probabilities(self) -> tuple[float, ...]:
if self.mode != "balanced":
return ()
raw = self.probabilities or tuple(1.0 for _ in self.values)
total = float(sum(raw))
if not math.isfinite(total) or total <= 0:
raise ValueError("RTC delay probability sum must be finite and positive")
return tuple(float(value) / total for value in raw)
def signature(self) -> dict[str, Any]:
return {
"mode": self.mode,
"deployment_delay": self.deployment_delay,
"values": self.values,
"probabilities": self.probabilities,
}
@dataclass(frozen=True)
class FlowRandomBatch:
"""Random quantities consumed by one local flow-matching microbatch."""
sigmas: torch.Tensor
noise: torch.Tensor
delay: torch.Tensor | None
draw_index: int
class DistributedFlowRandomSampler:
"""Generate reproducible, non-duplicated distributed flow randomness.
Every rank must use the same ``seed``, ``world_size``, sigma bounds, delay
policy, and local batch size at a given draw. ``rank`` is the global rank,
not the node-local CUDA index.
"""
def __init__(
self,
*,
seed: int,
rank: int,
world_size: int,
sigma_min: float = 0.05,
sigma_max: float = 1.0,
delay: DelaySamplingConfig | None = None,
draw_index: int = 0,
) -> None:
if int(seed) < 0:
raise ValueError("seed must be non-negative")
if int(world_size) <= 0:
raise ValueError("world_size must be positive")
if int(rank) < 0 or int(rank) >= int(world_size):
raise ValueError("rank must satisfy 0 <= rank < world_size")
if not math.isfinite(float(sigma_min)) or not math.isfinite(float(sigma_max)):
raise ValueError("sigma bounds must be finite")
if float(sigma_min) < 0 or float(sigma_max) <= float(sigma_min):
raise ValueError("sigma bounds must satisfy 0 <= sigma_min < sigma_max")
if int(draw_index) < 0:
raise ValueError("draw_index must be non-negative")
self.seed = int(seed)
self.rank = int(rank)
self.world_size = int(world_size)
self.sigma_min = float(sigma_min)
self.sigma_max = float(sigma_max)
self.delay_config = delay
self._draw_index = int(draw_index)
@property
def draw_index(self) -> int:
"""Number of already generated microbatches."""
return self._draw_index
def state_dict(self) -> dict[str, Any]:
"""Small checkpoint state accepted by Accelerate's checkpoint hooks."""
return {
"version": _STATE_VERSION,
"draw_index": self._draw_index,
"config": self._config_signature(),
}
def load_state_dict(self, state_dict: Mapping[str, Any]) -> None:
version = int(state_dict.get("version", -1))
if version != _STATE_VERSION:
raise ValueError(
f"unsupported distributed flow sampler state version {version}; "
f"expected {_STATE_VERSION}"
)
saved_config = state_dict.get("config")
current_config = self._config_signature()
if saved_config != current_config:
raise ValueError(
"distributed flow sampler configuration changed across resume: "
f"saved={saved_config!r}, current={current_config!r}"
)
draw_index = int(state_dict["draw_index"])
if draw_index < 0:
raise ValueError("checkpointed draw_index must be non-negative")
self._draw_index = draw_index
def sample_like(self, actions: torch.Tensor) -> FlowRandomBatch:
"""Sample sigma/noise/delay matching an action tensor."""
if not actions.is_floating_point():
raise ValueError("actions must have a floating-point dtype")
return self.sample(
actions.shape,
device=actions.device,
noise_dtype=actions.dtype,
)
def sample(
self,
action_shape: Sequence[int],
*,
device: torch.device | str,
noise_dtype: torch.dtype = torch.float32,
) -> FlowRandomBatch:
"""Sample one local microbatch and advance ``draw_index`` by one."""
shape = tuple(int(size) for size in action_shape)
if not shape or shape[0] <= 0 or any(size <= 0 for size in shape):
raise ValueError("action_shape must contain only positive dimensions")
if not noise_dtype.is_floating_point:
raise ValueError("noise_dtype must be floating point")
local_batch_size = shape[0]
target_device = torch.device(device)
draw_index = self._draw_index
sigmas = self._sample_sigmas(local_batch_size, draw_index).to(
device=target_device,
dtype=torch.float32,
)
delay = self._sample_delay(local_batch_size, draw_index)
if delay is not None:
delay = delay.to(device=target_device, dtype=torch.long)
noise = self._sample_noise(shape, target_device, noise_dtype, draw_index)
# Advance only after all component draws succeeded.
self._draw_index += 1
return FlowRandomBatch(
sigmas=sigmas,
noise=noise,
delay=delay,
draw_index=draw_index,
)
def _config_signature(self) -> dict[str, Any]:
return {
"seed": self.seed,
# Rank is intentionally omitted. Accelerate may restore one shared
# custom state file on all ranks; the live rank still selects its
# own global-batch slice and rank-specific noise stream.
"world_size": self.world_size,
"sigma_min": self.sigma_min,
"sigma_max": self.sigma_max,
"delay": self.delay_config.signature() if self.delay_config is not None else None,
}
@staticmethod
def _cpu_generator(seed: int) -> torch.Generator:
generator = torch.Generator(device="cpu")
generator.manual_seed(int(seed))
return generator
def _local_slice(self, local_batch_size: int) -> slice:
start = self.rank * local_batch_size
return slice(start, start + local_batch_size)
def _sample_sigmas(self, local_batch_size: int, draw_index: int) -> torch.Tensor:
global_batch_size = self.world_size * local_batch_size
jitter_generator = self._cpu_generator(
_component_seed(self.seed, draw_index, _SIGMA_JITTER_COMPONENT)
)
permutation_generator = self._cpu_generator(
_component_seed(self.seed, draw_index, _SIGMA_PERM_COMPONENT)
)
jitter = torch.rand(
global_batch_size,
generator=jitter_generator,
dtype=torch.float64,
)
unit_sigmas = (
torch.arange(global_batch_size, dtype=torch.float64) + jitter
) / float(global_batch_size)
sigmas = self.sigma_min + (self.sigma_max - self.sigma_min) * unit_sigmas
permutation = torch.randperm(global_batch_size, generator=permutation_generator)
return sigmas[permutation][self._local_slice(local_batch_size)]
def _sample_delay(self, local_batch_size: int, draw_index: int) -> torch.Tensor | None:
config = self.delay_config
if config is None:
return None
if config.mode == "fixed":
return torch.full(
(local_batch_size,),
int(config.deployment_delay),
dtype=torch.long,
)
global_batch_size = self.world_size * local_batch_size
offset_generator = self._cpu_generator(
_component_seed(self.seed, draw_index, _DELAY_OFFSET_COMPONENT)
)
permutation_generator = self._cpu_generator(
_component_seed(self.seed, draw_index, _DELAY_PERM_COMPONENT)
)
# A single randomized systematic offset makes every categorical count
# floor/ceil(probability * global_batch_size), eliminating IID clumps.
offset = torch.rand((), generator=offset_generator, dtype=torch.float64)
quantiles = (
torch.arange(global_batch_size, dtype=torch.float64) + offset
) / float(global_batch_size)
probabilities = torch.tensor(
config.normalized_probabilities(),
dtype=torch.float64,
)
cumulative = torch.cumsum(probabilities, dim=0)
cumulative[-1] = 1.0
category_ids = torch.searchsorted(cumulative, quantiles, right=False)
values = torch.tensor(config.values, dtype=torch.long)[category_ids]
permutation = torch.randperm(global_batch_size, generator=permutation_generator)
return values[permutation][self._local_slice(local_batch_size)]
def _sample_noise(
self,
shape: tuple[int, ...],
device: torch.device,
dtype: torch.dtype,
draw_index: int,
) -> torch.Tensor:
generator = torch.Generator(device=device)
generator.manual_seed(
_component_seed(
self.seed,
draw_index,
_NOISE_COMPONENT,
rank=self.rank,
)
)
return torch.randn(shape, device=device, dtype=dtype, generator=generator)