Sync from v0.13

This commit is contained in:
2026-01-19 10:38:50 +08:00
parent b2ef04d792
commit 5aef6c175a
3714 changed files with 854317 additions and 89342 deletions

View File

@@ -1,4 +1,6 @@
import os
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import random
import pytest
@@ -6,10 +8,14 @@ import ray
import torch
import torch.distributed as dist
from vllm.distributed import tensor_model_parallel_all_reduce
from vllm.distributed.device_communicators import custom_all_reduce
from vllm.test_utils import (init_test_distributed_environment,
multi_process_tensor_parallel)
from vllm.distributed.communication_op import tensor_model_parallel_all_reduce # noqa
from vllm.distributed.parallel_state import get_tp_group, graph_capture
from ..utils import (
ensure_model_parallel_initialized,
init_test_distributed_environment,
multi_process_parallel,
)
random.seed(42)
test_sizes = [random.randint(1024, 2048 * 1024) for _ in range(8)]
@@ -18,67 +24,109 @@ for i, v in enumerate(test_sizes):
@ray.remote(num_gpus=1, max_calls=1)
def graph_allreduce(world_size, rank, distributed_init_port):
del os.environ["CUDA_VISIBLE_DEVICES"]
device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device)
init_test_distributed_environment(1, world_size, rank,
distributed_init_port)
def graph_allreduce(
monkeypatch: pytest.MonkeyPatch,
tp_size,
pp_size,
rank,
distributed_init_port,
):
with monkeypatch.context() as m:
m.delenv("CUDA_VISIBLE_DEVICES", raising=False)
device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device)
init_test_distributed_environment(tp_size, pp_size, rank, distributed_init_port)
ensure_model_parallel_initialized(tp_size, pp_size)
group = get_tp_group().device_group
custom_all_reduce.init_custom_all_reduce()
for sz in test_sizes:
for dtype in [torch.float32, torch.float16, torch.bfloat16]:
with custom_all_reduce.capture():
# use integers so result matches NCCL exactly
inp1 = torch.randint(1,
16, (sz, ),
dtype=dtype,
device=torch.cuda.current_device())
inp2 = torch.randint(1,
16, (sz, ),
dtype=dtype,
device=torch.cuda.current_device())
torch.cuda.synchronize()
graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(graph):
out1 = tensor_model_parallel_all_reduce(inp1)
# the input buffer is immediately modified to test
# synchronization
dist.all_reduce(inp1)
out2 = tensor_model_parallel_all_reduce(inp2)
dist.all_reduce(inp2)
graph.replay()
assert torch.allclose(out1, inp1)
assert torch.allclose(out2, inp2)
# A small all_reduce for warmup.
# this is needed because device communicators might be created lazily
# (e.g. NCCL). This will ensure that the communicator is initialized
# before any communication happens, so that this group can be used for
# graph capture immediately.
data = torch.zeros(1)
data = data.to(device=device)
torch.distributed.all_reduce(data, group=group)
torch.cuda.synchronize()
del data
# we use the first group to communicate once
# and the second group to communicate twice
# and so on
# this is used to demonstrate that each group can
# communicate independently
num_communication = rank // tp_size + 1
for sz in test_sizes:
for dtype in [torch.float32, torch.float16, torch.bfloat16]:
with graph_capture(device=device) as graph_capture_context:
# use integers so result matches NCCL exactly
inp1 = torch.randint(
1, 16, (sz,), dtype=dtype, device=torch.cuda.current_device()
)
inp2 = torch.randint(
1, 16, (sz,), dtype=dtype, device=torch.cuda.current_device()
)
torch.cuda.synchronize()
graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(graph, stream=graph_capture_context.stream):
for i in range(num_communication):
out1 = tensor_model_parallel_all_reduce(inp1)
# the input buffer is immediately modified to test
# synchronization
dist.all_reduce(inp1, group=group)
out2 = tensor_model_parallel_all_reduce(inp2)
dist.all_reduce(inp2, group=group)
graph.replay()
torch.testing.assert_close(out1, inp1)
torch.testing.assert_close(out2, inp2)
@ray.remote(num_gpus=1, max_calls=1)
def eager_allreduce(world_size, rank, distributed_init_port):
del os.environ["CUDA_VISIBLE_DEVICES"]
device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device)
init_test_distributed_environment(1, world_size, rank,
distributed_init_port)
def eager_allreduce(
monkeypatch: pytest.MonkeyPatch,
tp_size,
pp_size,
rank,
distributed_init_port,
):
with monkeypatch.context() as m:
m.delenv("CUDA_VISIBLE_DEVICES", raising=False)
device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device)
init_test_distributed_environment(tp_size, pp_size, rank, distributed_init_port)
sz = 1024
custom_all_reduce.init_custom_all_reduce()
fa = custom_all_reduce.get_handle()
inp = torch.ones(sz, dtype=torch.float32, device=device)
out = fa.all_reduce_unreg(inp)
assert torch.allclose(out, inp * world_size)
# we use the first group to communicate once
# and the second group to communicate twice
# and so on
# this is used to demonstrate that each group can
# communicate independently
num_communication = rank // tp_size + 1
sz = 1024
fa = get_tp_group().device_communicator.ca_comm
inp = torch.ones(sz, dtype=torch.float32, device=device)
out = inp
for _ in range(num_communication):
out = fa.all_reduce(out, registered=False)
torch.testing.assert_close(out, inp * (tp_size**num_communication))
inp = torch.ones(sz * 4, dtype=torch.bfloat16, device=device)
out = fa.all_reduce_unreg(inp)
assert torch.allclose(out, inp * world_size)
inp = torch.ones(sz * 4, dtype=torch.bfloat16, device=device)
out = inp
for _ in range(num_communication):
out = fa.all_reduce(out, registered=False)
torch.testing.assert_close(out, inp * (tp_size**num_communication))
@pytest.mark.skipif(torch.cuda.device_count() < 2,
reason="Need at least 2 GPUs to run the test.")
@pytest.mark.parametrize("tensor_parallel_size", [2])
@pytest.mark.parametrize("tp_size", [2])
@pytest.mark.parametrize("pipeline_parallel_size", [1, 2])
@pytest.mark.parametrize("test_target", [eager_allreduce, graph_allreduce])
def test_multi_process_tensor_parallel(tensor_parallel_size, test_target):
multi_process_tensor_parallel(tensor_parallel_size, test_target)
if __name__ == "__main__":
multi_process_tensor_parallel(2, graph_allreduce)
def test_custom_allreduce(
monkeypatch: pytest.MonkeyPatch,
tp_size,
pipeline_parallel_size,
test_target,
):
world_size = tp_size * pipeline_parallel_size
if world_size > torch.cuda.device_count():
pytest.skip("Not enough GPUs to run the test.")
multi_process_parallel(monkeypatch, tp_size, pipeline_parallel_size, test_target)