minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

xpu: move prefill-only model tests to the nightly-xpu-1-gpu grid (#36814)

Browse Source
This commit is contained in:
ashwini rathi
2026-08-31 14:03:48 +08:00
committed by GitHub
parent 62c470697e
commit 10b67aa7a1
4 changed files with 4 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/xpu/test_xpu_classification.py
+1 -1
View File
@@ -14,7 +14,7 @@ from sglang.test.ci.ci_register import register_xpu_ci
from sglang.test.runners import HFRunner, SRTRunner
from sglang.test.test_utils import CustomTestCase, empty_gpu_cache
register_xpu_ci(est_time=120, suite="stage-b-test-1-gpu-xpu")
register_xpu_ci(est_time=120, suite="nightly-xpu-1-gpu", nightly=True)
MODEL_PATH = "jason9693/Qwen2.5-1.5B-apeach"
TP_SIZE = 1
test/registered/xpu/test_xpu_embedding.py
+1 -1
View File
@@ -17,7 +17,7 @@ from sglang.test.ci.ci_register import register_xpu_ci
from sglang.test.runners import DEFAULT_PROMPTS, HFRunner, SRTRunner
from sglang.test.test_utils import CustomTestCase, empty_gpu_cache, get_similarities
register_xpu_ci(est_time=180, suite="stage-b-test-1-gpu-xpu")
register_xpu_ci(est_time=180, suite="nightly-xpu-1-gpu", nightly=True)
MODEL_PATH = "Alibaba-NLP/gte-Qwen2-1.5B-instruct"
TP_SIZE = 1
test/registered/xpu/test_xpu_rerank.py
+1 -21
View File
@@ -23,18 +23,7 @@ from sglang.test.ci.ci_register import register_xpu_ci
from sglang.test.runners import TEST_RERANK_QUERY_DOCS, HFRunner, SRTRunner
from sglang.test.test_utils import CustomTestCase, empty_gpu_cache
def _xpu_total_gib() -> float:
if not torch.xpu.is_available():
return 0.0
return torch.xpu.get_device_properties(0).total_memory / (1024**3)
# fp32+Triton fits on B60 (22GiB) but hangs on B580 (~12GiB).
_LARGE_XPU_VRAM_GIB = 20.0
_HAS_LARGE_XPU = _xpu_total_gib() >= _LARGE_XPU_VRAM_GIB
register_xpu_ci(est_time=180, suite="stage-b-test-1-gpu-xpu")
register_xpu_ci(est_time=180, suite="nightly-xpu-1-gpu", nightly=True)
MODEL_PATH = "Qwen/Qwen3-Reranker-0.6B"
TP_SIZE = 1
@@ -168,10 +157,6 @@ class TestXPUDecoderRerank(CustomTestCase):
self._assert_close_scores(prompts)
# Ported from test/manual/prefill_only/test_cross_encoder_models.py.
# fp32+triton fits on B60 (22GiB) but OOMs on B580 (~12GiB); bf16+intel_xpu on
# this encoder model does not match HF (tracked separately), so on <20GiB XPU
# the class is skipped rather than shipping a knowingly-wrong config.
CROSS_ENCODER_MODEL_PATH = "BAAI/bge-reranker-v2-m3"
CROSS_ENCODER_TP_SIZE = 1
CROSS_ENCODER_SCORE_TOLERANCE = 1e-2
@@ -180,11 +165,6 @@ CROSS_ENCODER_ATTENTION_BACKEND = "triton"
CROSS_ENCODER_MEM_FRACTION_STATIC = 0.65
@unittest.skipUnless(
_HAS_LARGE_XPU,
"bge-reranker-v2-m3 fp32+triton OOMs on <20GiB XPU (B580); "
"bf16+intel_xpu on this encoder produces wrong scores.",
)
class TestXPUCrossEncoderRerank(CustomTestCase):
@classmethod
def setUpClass(cls):
test/registered/xpu/test_xpu_reward.py
+1 -1
View File
@@ -15,7 +15,7 @@ from sglang.test.ci.ci_register import register_xpu_ci
from sglang.test.runners import HFRunner, SRTRunner
from sglang.test.test_utils import CustomTestCase, empty_gpu_cache
register_xpu_ci(est_time=60, suite="stage-b-test-1-gpu-xpu")
register_xpu_ci(est_time=60, suite="nightly-xpu-1-gpu", nightly=True)
MODEL_PATH = "Skywork/Skywork-Reward-V2-Qwen3-0.6B"
TP_SIZE = 1