File size: 5,520 Bytes
fa2d87b | 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 | from __future__ import annotations
import json
import os
from pathlib import Path
from typing import Any
import torch
SPENT_PREDICTIONS = frozenset({"noise_pred", "audio_noise_pred"})
class DenoiseCheckpointStop(RuntimeError):
def __init__(self, *, completed_steps: int, total_steps: int):
self.completed_steps = completed_steps
self.total_steps = total_steps
super().__init__(f"stopped after durable checkpoint {completed_steps} of {total_steps}")
def _partial_path(path: Path) -> Path:
return path.with_name(f"{path.name}.partial-{os.getpid()}")
def _sync_and_replace(partial: Path, target: Path) -> None:
with partial.open("rb") as handle:
os.fsync(handle.fileno())
os.replace(partial, target)
directory_fd = os.open(target.parent, os.O_RDONLY)
try:
os.fsync(directory_fd)
finally:
os.close(directory_fd)
def atomic_torch_save(payload: Any, path: Path) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
partial = _partial_path(path)
try:
torch.save(payload, partial)
_sync_and_replace(partial, path)
finally:
partial.unlink(missing_ok=True)
def atomic_json_write(payload: Any, path: Path) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
partial = _partial_path(path)
try:
with partial.open("w", encoding="utf-8") as handle:
json.dump(payload, handle, indent=2, ensure_ascii=False)
handle.write("\n")
handle.flush()
os.fsync(handle.fileno())
os.replace(partial, path)
directory_fd = os.open(path.parent, os.O_RDONLY)
try:
os.fsync(directory_fd)
finally:
os.close(directory_fd)
finally:
partial.unlink(missing_ok=True)
def snapshot_to_cpu(value: Any) -> Any:
if isinstance(value, torch.Tensor):
return value.detach().to(device="cpu").contiguous().clone()
if isinstance(value, dict):
return {key: snapshot_to_cpu(item) for key, item in value.items()}
if isinstance(value, list):
return [snapshot_to_cpu(item) for item in value]
if isinstance(value, tuple):
return tuple(snapshot_to_cpu(item) for item in value)
return value
def write_step_checkpoint(
directory: Path,
*,
step_index: int,
total_steps: int,
block_state,
metadata: dict[str, Any],
) -> tuple[Path, Path]:
if not 0 <= step_index < total_steps:
raise ValueError(f"invalid checkpoint step {step_index} for total {total_steps}")
completed_step = step_index + 1
stem = f"step-{completed_step:03d}-of-{total_steps:03d}"
state = snapshot_to_cpu(block_state.as_dict())
for name in SPENT_PREDICTIONS:
state.pop(name, None)
payload = {
"schema_version": 1,
"completed_step": completed_step,
"resume_step_index": completed_step,
"total_steps": total_steps,
"metadata": dict(metadata),
"state": state,
}
checkpoint = directory / f"{stem}.pt"
manifest = directory / f"{stem}.json"
atomic_torch_save(payload, checkpoint)
atomic_json_write(
{
"status": "checkpointed",
"completed_step": completed_step,
"resume_step_index": completed_step,
"total_steps": total_steps,
"checkpoint": checkpoint.name,
**metadata,
},
manifest,
)
atomic_json_write(
{
"status": "checkpointed",
"completed_step": completed_step,
"resume_step_index": completed_step,
"total_steps": total_steps,
"checkpoint": checkpoint.name,
"manifest": manifest.name,
**metadata,
},
directory / "latest.json",
)
return checkpoint, manifest
def install_loop_checkpointing(
loop_wrapper_cls,
directory: Path,
*,
metadata: dict[str, Any],
stop_after_steps: int | None = None,
):
if stop_after_steps is not None and stop_after_steps < 1:
raise ValueError("stop_after_steps must be positive")
original_call = loop_wrapper_cls.__call__
@torch.no_grad()
def checkpointed_call(self, components, state):
block_state = self.get_block_state(state)
total_steps = len(block_state.timesteps)
with self.progress_bar(total=total_steps) as progress_bar:
for step_index, timestep in enumerate(block_state.timesteps):
components, block_state = self.loop_step(
components,
block_state,
i=step_index,
t=timestep,
)
write_step_checkpoint(
directory,
step_index=step_index,
total_steps=total_steps,
block_state=block_state,
metadata=metadata,
)
progress_bar.update()
completed_steps = step_index + 1
if stop_after_steps is not None and completed_steps >= stop_after_steps:
self.set_block_state(state, block_state)
raise DenoiseCheckpointStop(
completed_steps=completed_steps,
total_steps=total_steps,
)
self.set_block_state(state, block_state)
return components, state
loop_wrapper_cls.__call__ = checkpointed_call
return original_call
|