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, 20:45 UTC
[ci][distributed] fix flaky tests (#6806)
youkaichao committed
Jul 25, 2024 at 17:44 UTC
443c7cf4cf891e6957d4b31655e58cabceb5a2a7
1 file changed
+8
-1
tests/distributed/test_pipeline_parallel.py
+8
-1
@@ -1,3 +1,10 @@
1
+"""
2
+WARNING: This test runs in both single-node (4 GPUs) and multi-node
3
+ (2 node with 2 GPUs each) modes. If the test only uses 2 GPUs, it is
4
+ important to set the distributed backend to "mp" to avoid Ray scheduling
5
+ all workers in a node other than the head node, which can cause the test
6
+ to fail.
7
+"""
8
import os
9
10
import pytest
@@ -78,7 +85,7 @@ def test_pp_cudagraph(PP_SIZE, MODEL_NAME, ATTN_BACKEND):
85
"--pipeline-parallel-size",
86
str(PP_SIZE),
87
"--distributed-executor-backend",
81
- "ray",
88
+ "mp",
89
]
90
os.environ["VLLM_ATTENTION_BACKEND"] = ATTN_BACKEND
91