A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 19:30 UTC
[TPU] Support multi-host inference (#7457)
Woosuk Kwon committed
Aug 13, 2024 at 16:31 UTC
a08df8322e2e5f27cb696bc2c0d8cdabe185878b
2 files changed
+11
-4
docs/source/getting_started/tpu-installation.rst
+1
-1
@@ -8,7 +8,7 @@ vLLM supports Google Cloud TPUs using PyTorch XLA.
8
Requirements
9
------------
10
11
-* Google Cloud TPU VM (single host)
11
+* Google Cloud TPU VM (single & multi host)
12
* TPU versions: v5e, v5p, v4
13
* Python: 3.10
14
vllm/distributed/device_communicators/tpu_communicator.py
+10
-3
@@ -1,3 +1,4 @@
1
+import ray
2
import torch
3
import torch.distributed as dist
4
from torch.distributed import ProcessGroup
@@ -18,9 +19,15 @@ class TpuCommunicator:
19
return
20
self.disabled = False
21
21
- local_rank = dist.get_rank(group)
22
- world_size = dist.get_world_size(group)
23
- pjrt.initialize_multiprocess(local_rank, world_size)
22
+ # NOTE(woosuk): When using TP > 1 on TPUs, every TPU on the same node
23
+ # must be used together. Therefore, the local rank and world size can
24
+ # be simply calculated as follows.
25
+ global_rank = dist.get_rank(group)
26
+ global_world_size = dist.get_world_size(group)
27
+ num_nodes = len(ray.nodes())
28
+ local_world_size = global_world_size // num_nodes
29
+ local_rank = global_rank % local_world_size
30
+ pjrt.initialize_multiprocess(local_rank, local_world_size)
31
xr._init_world_size_ordinal()
32
33
def all_reduce(self, x: torch.Tensor) -> torch.Tensor: