[Core][Distributed] remove graph mode function (#4818)
This commit is contained in:
@@ -50,7 +50,7 @@ def graph_allreduce(tp_size, pp_size, rank, distributed_init_port):
|
||||
|
||||
for sz in test_sizes:
|
||||
for dtype in [torch.float32, torch.float16, torch.bfloat16]:
|
||||
with graph_capture():
|
||||
with graph_capture() as graph_capture_context:
|
||||
# use integers so result matches NCCL exactly
|
||||
inp1 = torch.randint(1,
|
||||
16, (sz, ),
|
||||
@@ -62,7 +62,8 @@ def graph_allreduce(tp_size, pp_size, rank, distributed_init_port):
|
||||
device=torch.cuda.current_device())
|
||||
torch.cuda.synchronize()
|
||||
graph = torch.cuda.CUDAGraph()
|
||||
with torch.cuda.graph(graph):
|
||||
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
|
||||
|
||||
@@ -5,7 +5,7 @@ import pytest
|
||||
import torch
|
||||
|
||||
from vllm.distributed.communication_op import ( # noqa
|
||||
graph_mode, tensor_model_parallel_all_reduce)
|
||||
graph_capture, tensor_model_parallel_all_reduce)
|
||||
from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator
|
||||
from vllm.distributed.device_communicators.pynccl_wrapper import NCCLLibrary
|
||||
from vllm.distributed.parallel_state import (ensure_model_parallel_initialized,
|
||||
@@ -103,7 +103,7 @@ def multiple_tp_with_vllm_worker_fn():
|
||||
device = torch.device(f"cuda:{torch.distributed.get_rank()}")
|
||||
ensure_model_parallel_initialized(2, 2)
|
||||
tensor = torch.ones(16, 1024, 1024, dtype=torch.float32, device=device)
|
||||
with graph_mode():
|
||||
with graph_capture():
|
||||
# two tp groups can communicate independently
|
||||
if torch.distributed.get_rank() in [0, 1]:
|
||||
tensor = tensor_model_parallel_all_reduce(tensor)
|
||||
|
||||
Reference in New Issue
Block a user