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, 17:00 UTC
[TPU] Align worker index with node boundary (#7932)
Woosuk Kwon committed
Sep 1, 2024 at 23:09 UTC
e2b2aa5a0fdd3e682dd1fbd62e2ba81b8aa054d2
1 file changed
+28
vllm/executor/ray_tpu_executor.py
+28
@@ -111,12 +111,40 @@ class RayTPUExecutor(TPUExecutor):
111
# Else, added to the list of workers.
112
self.workers.append(worker)
113
114
+ logger.debug("workers: %s", self.workers)
115
+ logger.debug("driver_dummy_worker: %s", self.driver_dummy_worker)
116
if self.driver_dummy_worker is None:
117
raise ValueError(
118
"Ray does not allocate any TPUs on the driver node. Consider "
119
"adjusting the Ray placement group or running the driver on a "
120
"TPU node.")
121
122
+ worker_ips = [
123
+ ray.get(worker.get_node_ip.remote()) # type: ignore[attr-defined]
124
+ for worker in self.workers
125
+ ]
126
+ ip_counts: Dict[str, int] = {}
127
+ for ip in worker_ips:
128
+ ip_counts[ip] = ip_counts.get(ip, 0) + 1
129
+
130
+ def sort_by_driver_then_worker_ip(worker):
131
+ """
132
+ Sort the workers based on 3 properties:
133
+ 1. If the worker is on the same node as the driver (vllm engine),
134
+ it should be placed first.
135
+ 2. Then, if the worker is on a node with fewer workers, it should
136
+ be placed first.
137
+ 3. Finally, if the work is on a node with smaller IP address, it
138
+ should be placed first.
139
+ """
140
+ ip = ray.get(worker.get_node_ip.remote())
141
+ return (ip != driver_ip, ip_counts[ip], ip)
142
+
143
+ # After sorting, the workers on the same node will be
144
+ # close to each other, and the workers on the driver
145
+ # node will be placed first.
146
+ self.workers = sorted(self.workers, key=sort_by_driver_then_worker_ip)
147
+
148
# Get the set of TPU IDs used on each node.
149
worker_node_and_gpu_ids = self._run_workers("get_node_and_gpu_ids",
150
use_dummy_driver=True)