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

enable ut test for xpu devices (#11712)

Browse Source 
Co-authored-by: jundu <jun.du@intel.com>
Co-authored-by: Gao, Pengfei <pengfei.gao@intel.com>
This commit is contained in:
DiweiSun
2026-02-03 11:15:14 -08:00
committed by GitHub
co-authored by jundu Gao, Pengfei
parent 0a6925639b
commit 495290aefd
20 changed files with 237 additions and 151 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/manual/test_expert_location_updater.py
+4 -1
View File
@@ -10,6 +10,7 @@ import torch.multiprocessing as mp
from torch.multiprocessing import Process
from sglang.srt.eplb import expert_location_updater
from sglang.srt.utils import get_device
from sglang.test.test_utils import CustomTestCase, find_available_port
from sglang.utils import is_in_ci
@@ -61,7 +62,7 @@ class TestExpertLocationUpdater(CustomTestCase):
def test_gpu(self):
if is_in_ci():
return
self._test_common(device="cuda")
self._test_common(device=get_device())
def _test_common(self, device):
infos = []
@@ -135,6 +136,8 @@ def _run_subprocess(
)
if device == "cuda":
torch.cuda.set_device(f"cuda:{rank}")
if device == "xpu":
torch.xpu.set_device(f"xpu:{rank}")
for info in infos:
_execute_test(info, rank=rank, num_gpus=num_gpus, device=device)
test/manual/test_forward_split_prefill.py
+2 -1
View File
@@ -20,6 +20,7 @@ from sglang.srt.model_executor.model_runner import ModelRunner
from sglang.srt.sampling.sampling_params import SamplingParams
from sglang.srt.server_args import PortArgs, ServerArgs
from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
from sglang.srt.utils import get_device
from sglang.srt.utils.hf_transformers_utils import get_tokenizer
from sglang.test.test_utils import DEFAULT_SMALL_MODEL_NAME_FOR_TEST, CustomTestCase
@@ -32,7 +33,7 @@ class TestForwardSplitPrefill(CustomTestCase):
"""Set up the test environment once for all tests."""
cls.model_path = DEFAULT_SMALL_MODEL_NAME_FOR_TEST
cls.tp_size = 1
cls.device = "cuda"
cls.device = get_device()
# Initialize server args
cls.server_args = ServerArgs(
test/manual/test_get_weights_by_name.py
+7 -5
View File
@@ -3,16 +3,18 @@ import unittest
import numpy as np
import requests
import torch
from transformers import AutoModelForCausalLM
import sglang as sgl
from sglang.srt.utils import get_device
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
empty_gpu_cache,
get_gpu_count,
is_in_ci,
popen_launch_server,
)
@@ -32,7 +34,7 @@ class TestGetWeightsByName(CustomTestCase):
def init_hf_model(self, model_name, tie_word_embeddings):
self.hf_model = AutoModelForCausalLM.from_pretrained(
model_name, torch_dtype="bfloat16", tie_word_embeddings=tie_word_embeddings
).to("cuda:0")
).to(get_device())
def init_backend(self, backend, dp, tp, model_name):
self.backend = backend
@@ -61,7 +63,7 @@ class TestGetWeightsByName(CustomTestCase):
def clean_up(self):
del self.hf_model
gc.collect()
torch.cuda.empty_cache()
empty_gpu_cache()
if self.backend == "Engine":
self.engine.shutdown()
else:
@@ -132,11 +134,11 @@ class TestGetWeightsByName(CustomTestCase):
("Runtime", 1, 1, DEFAULT_SMALL_MODEL_NAME_FOR_TEST),
("Engine", 1, 1, DEFAULT_MODEL_NAME_FOR_TEST),
]
if torch.cuda.device_count() >= 2:
if get_gpu_count() >= 2:
test_suits.append(("Engine", 1, 2, DEFAULT_SMALL_MODEL_NAME_FOR_TEST))
test_suits.append(("Runtime", 2, 1, DEFAULT_MODEL_NAME_FOR_TEST))
if torch.cuda.device_count() >= 4:
if get_gpu_count() >= 4:
test_suits.extend(
[
("Engine", 2, 2, DEFAULT_SMALL_MODEL_NAME_FOR_TEST),
test/manual/test_triton_moe_wna16.py
+17 -10
View File
@@ -7,6 +7,7 @@ from sglang.srt.layers.activation import SiluAndMul
from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_moe
from sglang.srt.layers.moe.topk import TopKConfig, select_experts
from sglang.srt.server_args import ServerArgs, set_global_server_args_for_scheduler
from sglang.srt.utils import get_device
NUM_EXPERTS = [8, 64]
TOP_KS = [2, 6]
@@ -159,10 +160,10 @@ def test_fused_moe_wn16(
weight_bits: int,
):
print(m, n, k, e, topk, dtype, group_size, has_zp, weight_bits)
a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
w2 = torch.randn((e, k, n), device="cuda", dtype=dtype) / 10
score = torch.randn((m, e), device="cuda", dtype=dtype)
a = torch.randn((m, k), device=get_device(), dtype=dtype) / 10
w1 = torch.randn((e, 2 * n, k), device=get_device(), dtype=dtype) / 10
w2 = torch.randn((e, k, n), device=get_device(), dtype=dtype) / 10
score = torch.randn((m, e), device=get_device(), dtype=dtype)
if weight_bits == 4:
pack_factor = 2
@@ -174,16 +175,22 @@ def test_fused_moe_wn16(
w1_ref = w1.clone()
w2_ref = w2.clone()
w1_qweight = torch.empty(
(e, 2 * n, k // pack_factor), device="cuda", dtype=torch.uint8
(e, 2 * n, k // pack_factor), device=get_device(), dtype=torch.uint8
)
w2_qweight = torch.empty((e, k, n // pack_factor), device="cuda", dtype=torch.uint8)
w1_scales = torch.empty((e, 2 * n, k // group_size), device="cuda", dtype=dtype)
w2_scales = torch.empty((e, k, n // group_size), device="cuda", dtype=dtype)
w2_qweight = torch.empty(
(e, k, n // pack_factor), device=get_device(), dtype=torch.uint8
)
w1_scales = torch.empty(
(e, 2 * n, k // group_size), device=get_device(), dtype=dtype
)
w2_scales = torch.empty((e, k, n // group_size), device=get_device(), dtype=dtype)
w1_qzeros = torch.empty(
(e, 2 * n // pack_factor, k // group_size), device="cuda", dtype=torch.uint8
(e, 2 * n // pack_factor, k // group_size),
device=get_device(),
dtype=torch.uint8,
)
w2_qzeros = torch.empty(
(e, k // pack_factor, n // group_size), device="cuda", dtype=torch.uint8
(e, k // pack_factor, n // group_size), device=get_device(), dtype=torch.uint8
)
for i in range(e * 2):