[AMD] prepare for MI300x PR runner pool: registry mirror, runner routing, threshold tuning (#23156)

This commit is contained in:
Bingxu Chen
2026-04-21 00:58:23 -07:00
committed by GitHub
parent 74fdf9cd77
commit 09b1d10d59
14 changed files with 145 additions and 117 deletions
@@ -10,6 +10,7 @@ from sglang.test.test_utils import (
CustomTestCase,
auto_config_device,
get_benchmark_args,
is_in_amd_ci,
is_in_ci,
popen_launch_server,
run_benchmark,
@@ -70,7 +71,8 @@ class TestMultiTokenizer(CustomTestCase, MMLUMixin):
f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
)
self.assertLess(res["median_e2e_latency_ms"], 11000)
self.assertLess(res["median_ttft_ms"], 86)
# relax for mi300x
self.assertLess(res["median_ttft_ms"], 130 if is_in_amd_ci() else 86)
self.assertLess(res["median_itl_ms"], 10)