Skip to content

[TPU] Support multi-host inference #7457

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Aug 13, 2024
Merged

[TPU] Support multi-host inference #7457

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
15bdf1f
[TPU] Support multi-host inference
WoosukKwon Aug 13, 2024
c861231
Add comment
WoosukKwon Aug 13, 2024
4d9c028
Minor fix in doc
WoosukKwon Aug 13, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion docs/source/getting_started/tpu-installation.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,7 @@ vLLM supports Google Cloud TPUs using PyTorch XLA.
Requirements
------------

* Google Cloud TPU VM (single host)
* Google Cloud TPU VM (single & multi host)
* TPU versions: v5e, v5p, v4
* Python: 3.10

Expand Down
13 changes: 10 additions & 3 deletions vllm/distributed/device_communicators/tpu_communicator.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,4 @@
import ray
import torch
import torch.distributed as dist
from torch.distributed import ProcessGroup
Expand All @@ -18,9 +19,15 @@ def __init__(self, group: ProcessGroup):
return
self.disabled = False

local_rank = dist.get_rank(group)
world_size = dist.get_world_size(group)
pjrt.initialize_multiprocess(local_rank, world_size)
# NOTE(woosuk): When using TP > 1 on TPUs, every TPU on the same node
# must be used together. Therefore, the local rank and world size can
# be simply calculated as follows.
global_rank = dist.get_rank(group)
global_world_size = dist.get_world_size(group)
num_nodes = len(ray.nodes())
local_world_size = global_world_size // num_nodes
local_rank = global_rank % local_world_size
pjrt.initialize_multiprocess(local_rank, local_world_size)
xr._init_world_size_ordinal()

def all_reduce(self, x: torch.Tensor) -> torch.Tensor:
Expand Down