Instructions to use taohu/fastgen-offline with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use taohu/fastgen-offline with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("taohu/fastgen-offline", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
File size: 14,780 Bytes
0839907 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 | # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
"""
Test utilities for fastgen, including:
- RunIf: Conditional test skipping based on hardware/software requirements
- Distributed testing infrastructure for multi-GPU FSDP tests
"""
import os
import pickle
import socket
import tempfile
import threading
import time
from pathlib import Path
from typing import Any, Callable, Dict, List, Optional, Union
import sys
import pytest
from pytest import MarkDecorator
from packaging.version import Version
from pkg_resources import get_distribution
import torch
import torch.distributed as dist
import torch.multiprocessing as mp
# =============================================================================
# Distributed Testing Infrastructure
# =============================================================================
def _init_distributed(rank: int, world_size: int, port: int, nccl_timeout: int = 600):
"""Initialize distributed process group for testing.
Args:
rank: Process rank
world_size: Total number of processes
port: Port for distributed communication
nccl_timeout: Timeout in seconds for NCCL operations (default: 600)
"""
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = str(port)
os.environ["RANK"] = str(rank)
os.environ["WORLD_SIZE"] = str(world_size)
os.environ["LOCAL_RANK"] = str(rank)
from datetime import timedelta
dist.init_process_group(
backend="nccl",
init_method=f"tcp://localhost:{port}",
rank=rank,
world_size=world_size,
timeout=timedelta(seconds=nccl_timeout),
)
torch.cuda.set_device(rank)
def _cleanup_distributed():
"""Clean up distributed process group."""
if dist.is_initialized():
dist.destroy_process_group()
def _error_watchdog(error_event: mp.Event, check_interval: float = 0.5):
"""Watchdog thread that monitors for errors from other ranks.
When the error_event is set by any rank, this thread will terminate
the current process to prevent hanging on NCCL collectives.
Args:
error_event: Shared multiprocessing Event that signals an error occurred
check_interval: How often to check for errors (seconds)
"""
while True:
if error_event.is_set():
# Another rank failed, force exit to prevent hanging on collectives
os._exit(1)
time.sleep(check_interval)
def _run_distributed_test(
rank: int,
world_size: int,
port: int,
test_fn: Callable,
result_file: str,
error_file: str,
error_event: mp.Event,
setup_fn: Optional[Callable] = None,
nccl_timeout: int = 600,
**kwargs,
):
"""Worker function to run a distributed test.
Args:
rank: Process rank
world_size: Total number of processes
port: Port for distributed communication
test_fn: The test function to run with signature fn(rank, world_size, **kwargs)
result_file: Path to file where rank 0 writes results
error_file: Path to file where any rank writes error details
error_event: Shared event to signal errors across ranks
setup_fn: Optional setup function to run after distributed init (e.g., set_env_vars)
nccl_timeout: Timeout in seconds for NCCL operations (default: 600)
**kwargs: Additional arguments to pass to test_fn
"""
# Start watchdog thread to monitor for errors from other ranks
watchdog = threading.Thread(target=_error_watchdog, args=(error_event,), daemon=True)
watchdog.start()
try:
_init_distributed(rank, world_size, port, nccl_timeout=nccl_timeout)
# Run optional setup function (e.g., set_env_vars for library-specific setup)
if setup_fn is not None:
setup_fn()
# Run the test function
result = test_fn(rank=rank, world_size=world_size, **kwargs)
# Only rank 0 writes the result to file
if rank == 0:
with open(result_file, "wb") as f:
pickle.dump(("success", result), f)
except Exception as e:
import traceback
error_info = {
"rank": rank,
"error": str(e),
"traceback": traceback.format_exc(),
}
# Signal error to all other ranks FIRST (before any cleanup)
error_event.set()
# Write error details to file (use rank-specific suffix to avoid race conditions)
try:
error_file_rank = f"{error_file}.rank{rank}"
with open(error_file_rank, "wb") as f:
pickle.dump(error_info, f)
except Exception:
pass # Best effort - error signaling via event is the priority
# Also write to main result file if rank 0
if rank == 0:
with open(result_file, "wb") as f:
pickle.dump(("error", (str(e), traceback.format_exc())), f)
finally:
_cleanup_distributed()
def run_distributed_test(
test_fn: Callable,
world_size: int = 2,
timeout: int = 300,
setup_fn: Optional[Callable] = None,
nccl_timeout: Optional[int] = None,
**kwargs,
) -> Optional[Dict]:
"""Run a test function in a distributed setting using multiprocessing.spawn.
This utility handles:
- Finding a free port for NCCL communication
- Spawning worker processes with proper distributed initialization
- Collecting results from rank 0 via pickle file (avoiding Queue issues with CUDA)
- Cross-rank error signaling to prevent hanging when one rank fails
- Proper cleanup of process groups
Args:
test_fn: Test function with signature fn(rank, world_size, **kwargs)
Should return a dict with test results from rank 0.
world_size: Number of processes to spawn (default: 2)
timeout: Timeout in seconds for each process to complete (default: 300)
setup_fn: Optional setup function to run after distributed init on each rank.
Useful for library-specific setup like set_env_vars().
nccl_timeout: Timeout in seconds for NCCL operations. Defaults to match `timeout`.
**kwargs: Additional arguments to pass to test_fn
Returns:
Result dict from rank 0 test execution, or None if no result was written
Raises:
TimeoutError: If any process exceeds the timeout
AssertionError: If any process exits with non-zero code or test fails
Example:
```python
def _test_my_distributed_feature(rank: int, world_size: int, some_arg: int) -> dict:
# Setup distributed model/training
...
# Run test logic
...
# Return results (only rank 0's result is collected)
return {"passed": True, "value": computed_value}
@RunIf(min_gpus=2)
def test_my_distributed_feature():
result = run_distributed_test(
test_fn=_test_my_distributed_feature,
world_size=2,
some_arg=42,
)
assert result["passed"]
```
"""
# Find a free port
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
s.bind(("", 0))
port = s.getsockname()[1]
# Create temp files for results and errors
with tempfile.NamedTemporaryFile(delete=False, suffix=".pkl") as f:
result_file = f.name
with tempfile.NamedTemporaryFile(delete=False, suffix=".err") as f:
error_file = f.name
# Create shared error event for cross-rank signaling
ctx = mp.get_context("spawn")
error_event = ctx.Event()
# Default nccl_timeout to match test timeout
if nccl_timeout is None:
nccl_timeout = timeout
processes = []
try:
# Spawn processes
for rank in range(world_size):
p = ctx.Process(
target=_run_distributed_test,
args=(rank, world_size, port, test_fn, result_file, error_file, error_event),
kwargs={"setup_fn": setup_fn, "nccl_timeout": nccl_timeout, **kwargs},
)
p.start()
processes.append(p)
# Wait for processes with periodic error checking
start_time = time.time()
while True:
# Check if error event was set
if error_event.is_set():
# Give processes a moment to write error files and exit
time.sleep(0.5)
# Terminate any still-running processes
for p in processes:
if p.is_alive():
p.terminate()
p.join(timeout=5)
break
# Check if all processes have finished
all_done = all(not p.is_alive() for p in processes)
if all_done:
break
# Check for timeout
elapsed = time.time() - start_time
if elapsed > timeout:
for p in processes:
if p.is_alive():
p.terminate()
p.join(timeout=5)
raise TimeoutError(f"Distributed test timed out after {timeout}s")
# Brief sleep to avoid busy waiting
time.sleep(0.1)
# Join all processes to collect exit codes
for p in processes:
p.join(timeout=5)
# Collect error information from all ranks
error_messages = []
for rank in range(world_size):
error_file_rank = f"{error_file}.rank{rank}"
if os.path.exists(error_file_rank):
try:
with open(error_file_rank, "rb") as f:
error_info = pickle.load(f)
error_messages.append(
f"Rank {error_info['rank']} failed:\n{error_info['error']}\n{error_info['traceback']}"
)
except Exception:
error_messages.append(f"Rank {rank} failed (unable to read error details)")
# If we collected errors, raise them
if error_messages:
raise AssertionError(
f"Distributed test failed on {len(error_messages)} rank(s):\n\n" + "\n---\n".join(error_messages)
)
# Check if any process failed with non-zero exit code
for i, p in enumerate(processes):
if p.exitcode != 0:
raise AssertionError(f"Process {i} exited with code {p.exitcode}")
# Read result from file
if os.path.exists(result_file) and os.path.getsize(result_file) > 0:
with open(result_file, "rb") as f:
status, result = pickle.load(f)
if status == "error":
error_msg, traceback_str = result
raise AssertionError(f"Distributed test failed:\n{error_msg}\n{traceback_str}")
return result
return None
finally:
# Ensure all processes are terminated
for p in processes:
if p.is_alive():
p.terminate()
p.join(timeout=5)
if p.is_alive():
p.kill()
# Clean up temp files
for f in [result_file, error_file]:
if os.path.exists(f):
try:
os.unlink(f)
except Exception:
pass
# Clean up rank-specific error files
for rank in range(world_size):
error_file_rank = f"{error_file}.rank{rank}"
if os.path.exists(error_file_rank):
try:
os.unlink(error_file_rank)
except Exception:
pass
# =============================================================================
# Conditional Test Skipping
# =============================================================================
class RunIf:
"""RunIf wrapper for conditional skipping of tests.
Fully compatible with `@pytest.mark`.
Example:
```python
@RunIf(min_torch="1.8")
@pytest.mark.parametrize("arg1", [1.0, 2.0])
def test_wrapper(arg1):
assert arg1 > 0
```
"""
def __new__(
cls,
min_gpus: int = 0,
min_torch: Optional[str] = None,
max_torch: Optional[str] = None,
min_python: Optional[str] = None,
supported_arch: Optional[List[str]] = None,
requires_file: Optional[Union[str, List[str]]] = None,
**kwargs: Dict[Any, Any],
) -> MarkDecorator:
"""Creates a new `@RunIf` `MarkDecorator` decorator.
:param min_gpus: Min number of GPUs required to run test.
:param min_torch: Minimum pytorch version to run test.
:param max_torch: Maximum pytorch version to run test.
:param min_python: Minimum python version required to run test.
:param requires_file: File or list of files required to run test.
:param kwargs: Native `pytest.mark.skipif` keyword arguments.
"""
conditions = []
reasons = []
if min_gpus:
conditions.append(torch.cuda.device_count() < min_gpus)
reasons.append(f"GPUs>={min_gpus}")
if min_torch:
torch_version = get_distribution("torch").version
conditions.append(Version(torch_version) < Version(min_torch))
reasons.append(f"torch>={min_torch}")
if max_torch:
torch_version = get_distribution("torch").version
conditions.append(Version(torch_version) >= Version(max_torch))
reasons.append(f"torch<{max_torch}")
if min_python:
py_version = f"{sys.version_info.major}.{sys.version_info.minor}.{sys.version_info.micro}"
conditions.append(Version(py_version) < Version(min_python))
reasons.append(f"python>={min_python}")
if requires_file:
if isinstance(requires_file, str):
requires_file = [requires_file]
conditions.extend([not Path(file).exists() for file in requires_file])
reasons.append(f"requires file={','.join(requires_file)}")
reasons = [rs for cond, rs in zip(conditions, reasons) if cond]
return pytest.mark.skipif(
condition=any(conditions),
reason=f"Requires: [{' + '.join(reasons)}]",
**kwargs,
)
def check_grad_zero(model: torch.nn.Module):
"""check if the gradients of the model are zero"""
for param in model.parameters():
if param.requires_grad and param.grad is not None:
assert torch.allclose(param.grad, torch.zeros_like(param.grad))
|