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

[NPU] Improve the execution efficiency and maintainability of pr‑test‑npu (#33724)

Browse Source 
Co-authored-by: Even Zhou <even.y.zhou@outlook.com>
Co-authored-by: sglang-npu-bot <sglangnpu@163.com>
This commit is contained in:
Cherry_ming
2026-08-08 16:42:44 +08:00
committed by GitHub
co-authored by Even Zhou sglang-npu-bot
parent e732c0a9dc
commit dd5d82bead
60 changed files with 978 additions and 565 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/npu/accuracy/glm4_7_flash/test_npu_glm4_7_flash_1p_gsm8k.py
+1 -6
View File
@@ -6,12 +6,7 @@ from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
from sglang.test.ascend.e2e.test_npu_performance_utils import GLM_4_7_FLASH_MODEL_PATH
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="accuracy testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-acc-2-npu-a3")
ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
test/registered/npu/accuracy/glm5_top64_pruned/test_npu_glm5_top64_pruned_bf16_8p_gsm8k.py
+2 -7
View File
@@ -8,12 +8,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="accuracy testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-acc-16-npu-a3")
ENVS = {
"SGLANG_SET_CPU_AFFINITY": "1",
@@ -54,7 +49,7 @@ class TestNPUGLM5_Top64_Pruned_GSM8K(TestNpuAccuracyTestCaseBase):
model = GLM5_TOP64_PRUNED_GSM8K_MODEL_PATH
envs = ENVS
other_args = OTHER_ARGS
accuracy = 0.50
accuracy = 0.48
datasets = ["gsm8k"]
generation_config = {
"max_tokens": 2048,
test/registered/npu/accuracy/moonshotai_moonlight_16b_a3b/test_npu_moonlight_16b_a3b_bf16_1p_gsm8k.py
+1 -6
View File
@@ -8,12 +8,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="accuracy testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-acc-2-npu-a3")
MODEL_ENVS = {
"SGLANG_SET_CPU_AFFINITY": "1",
test/registered/npu/accuracy/qwen3_5_9b/test_npu_qwen3_5_9b_bf16_1p_gsm8k.py
+1 -6
View File
@@ -8,12 +8,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="stage-b-test-2-npu-a3",
nightly=True,
disabled="accuracy testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-acc-2-npu-a3")
QWEN3_5_9B_ENVS = {
"SGLANG_SET_CPU_AFFINITY": "1",
test/registered/npu/accuracy/qwen3_vl_30b_a3b/test_npu_qwen3_vl_30b_a3b_bf16_2p_gsm8k.py
+1 -6
View File
@@ -8,12 +8,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="stage-b-test-4-npu-a3",
nightly=True,
disabled="accuracy testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-acc-4-npu-a3")
QWEN3_VL_30B_A3B_ENVS = {
"SGLANG_SET_CPU_AFFINITY": "1",
test/registered/npu/accuracy/qwen3_vl_8b/test_npu_qwen3_vl_8b_bf16_2p_gsm8k.py
+1 -6
View File
@@ -8,12 +8,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="stage-b-test-4-npu-a3",
nightly=True,
disabled="accuracy testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-acc-4-npu-a3")
QWEN3_VL_8B_ENVS = {
"SGLANG_SET_CPU_AFFINITY": "1",
test/registered/npu/accuracy/qwen3_vl_8b_thinking/test_npu_qwen3_vl_8b_thinking_1p_mmmu.py
+1 -6
View File
@@ -9,12 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="performance testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-perf-2-npu-a3")
_is_pr_pipeline = os.environ.get("GITHUB_EVENT_NAME") == "pull_request"
test/registered/npu/basic_function/HiCache/test_npu_hicache_mha.py
+1 -1
View File
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-1-npu-a3")
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
test/registered/npu/basic_function/HiCache/test_npu_hicache_mla.py
+1 -1
View File
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-4-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-4-npu-a3")
register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
test/registered/npu/basic_function/backends/test_npu_sampling_backend.py
+1 -1
View File
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-1-npu-a3")
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
test/registered/npu/basic_function/dllm/test_npu_llada2_mini.py
+1 -1
View File
@@ -10,7 +10,7 @@ from sglang.test.test_utils import (
write_github_step_summary,
)
register_npu_ci(est_time=400, suite="stage-b-test-4-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-4-npu-a3")
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
test/registered/npu/basic_function/dp_attn/test_npu_dp_attention.py
+1 -1
View File
@@ -23,7 +23,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-4-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-4-npu-a3")
register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
test/registered/npu/basic_function/optimization_debug/test_npu_compile_graph_tp1_bf16.py
+1 -1
View File
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-1-npu-a3")
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
test/registered/npu/basic_function/optimization_debug/test_npu_piecewise_graph_prefill.py
+1 -1
View File
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
run_bench_one_batch,
)
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-1-npu-a3")
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
test/registered/npu/basic_function/parallel_strategy/data_parallelism/test_npu_load_balance_method.py
+1 -1
View File
@@ -18,7 +18,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=700, suite="stage-b-test-16-npu-a3", nightly=False)
register_npu_ci(est_time=700, suite="base-b-test-16-npu-a3")
register_npu_ci(est_time=700, suite="nightly-16-npu-a3", nightly=True)
test/registered/npu/basic_function/parallel_strategy/data_parallelism/test_npu_load_balance_method_pd_disaggregation.py
+1 -1
View File
@@ -23,7 +23,7 @@ from sglang.test.test_utils import (
popen_launch_pd_server,
)
register_npu_ci(est_time=3600, suite="stage-b-test-16-npu-a3", nightly=False)
register_npu_ci(est_time=3600, suite="base-b-test-16-npu-a3")
register_npu_ci(est_time=3600, suite="nightly-16-npu-a3", nightly=True)
load_balance_method_options = [
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep.py
+1 -1
View File
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-16-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-16-npu-a3")
register_npu_ci(est_time=400, suite="nightly-16-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_eplb_min_rebalancing_utilization_threshold.py
+1 -1
View File
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-8-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-8-npu-a3")
register_npu_ci(est_time=400, suite="nightly-8-npu-a3", nightly=True)
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_expert_distribution_recorder_mode.py
+1 -1
View File
@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-2-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-2-npu-a3")
register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_moe_dense_tp_size.py
+1 -1
View File
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-2-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-2-npu-a3")
register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_moe_runner_backend.py
+1 -1
View File
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-2-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-2-npu-a3")
register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
test/registered/npu/basic_function/pd_disaggregation/test_npu_pd_disaggregation.py
+1 -1
View File
@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
popen_with_error_check,
)
register_npu_ci(est_time=400, suite="stage-b-test-16-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-16-npu-a3")
register_npu_ci(est_time=400, suite="nightly-16-npu-a3", nightly=True)
test/registered/npu/basic_function/quant/test_npu_autoround_dense.py
+1 -1
View File
@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-1-npu-a3")
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
logger = logging.getLogger(__name__)
test/registered/npu/basic_function/quant/test_npu_autoround_moe.py
+1 -1
View File
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-1-npu-a3")
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
logger = logging.getLogger(__name__)
test/registered/npu/basic_function/quant/test_npu_gptq_moe.py
+1 -1
View File
@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-1-npu-a3")
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
logger = logging.getLogger(__name__)
test/registered/npu/basic_function/quant/test_npu_w4a4_quantization.py
+1 -1
View File
@@ -13,7 +13,7 @@ from sglang.test.ascend.test_ascend_utils import (
from sglang.test.ci.ci_register import register_npu_ci
from sglang.test.test_utils import CustomTestCase
register_npu_ci(est_time=400, suite="stage-b-test-4-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-4-npu-a3")
register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
test/registered/npu/basic_function/runtime_opts/test_npu_mla_fia_w8a8int8.py
+1 -1
View File
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-2-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-2-npu-a3")
register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
test/registered/npu/basic_function/runtime_opts/test_npu_mla_w8a8int8.py
+1 -1
View File
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-4-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-4-npu-a3")
register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
test/registered/npu/basic_function/runtime_opts/test_npu_tp1_bf16.py
+1 -1
View File
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-1-npu-a3")
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
test/registered/npu/basic_function/runtime_opts/test_npu_tp2_bf16.py
+1 -1
View File
@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-2-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-2-npu-a3")
register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
test/registered/npu/basic_function/runtime_opts/test_npu_tp2_fia_bf16.py
+1 -1
View File
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-2-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-2-npu-a3")
register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
test/registered/npu/basic_function/runtime_opts/test_npu_tp4_bf16.py
+1 -1
View File
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-4-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-4-npu-a3")
register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
test/registered/npu/basic_function/speculative_inference/test_npu_basic_sanity_eagle3.py
+1 -1
View File
@@ -17,7 +17,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=200, suite="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=200, suite="base-b-test-1-npu-a3")
register_npu_ci(est_time=200, suite="nightly-1-npu-a3", nightly=True)
test/registered/npu/basic_function/speculative_inference/test_npu_eagle3.py
+1 -1
View File
@@ -9,7 +9,7 @@ from sglang.test.ascend.test_ascend_utils import (
from sglang.test.ci.ci_register import register_npu_ci
from sglang.test.test_utils import CustomTestCase
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-1-npu-a3")
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
test/registered/npu/basic_function/speculative_inference/test_npu_speculative_attention_mode.py
+1 -5
View File
@@ -17,11 +17,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(
est_time=400,
suite="stage-b-test-4-npu-a3",
nightly=False,
)
register_npu_ci(est_time=400, suite="base-b-test-4-npu-a3")
register_npu_ci(
est_time=400,
suite="nightly-4-npu-a3",
test/registered/npu/basic_function/speculative_inference/test_npu_speculative_draft_attention_backend.py
+1 -1
View File
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-16-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-16-npu-a3")
register_npu_ci(est_time=400, suite="nightly-16-npu-a3", nightly=True)
MODEL_PATH = DEEPSEEK_R1_0528_W4A8_PER_CHANNEL_WEIGHTS_PATH
test/registered/npu/basic_function/speculative_inference/test_npu_speculative_moe_a2a_backend.py
+1 -1
View File
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-16-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-16-npu-a3")
register_npu_ci(est_time=400, suite="nightly-16-npu-a3", nightly=True)
test/registered/npu/basic_function/speculative_inference/test_npu_speculative_multi_npu.py
+1 -1
View File
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-4-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="base-b-test-4-npu-a3")
register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
_ASCEND_BACKEND = "ascend"
test/registered/npu/basic_function/speculative_inference/test_npu_speculative_token_map.py
+1 -1
View File
@@ -20,7 +20,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_npu_ci(est_time=300, suite="stage-b-test-4-npu-a3", nightly=False)
register_npu_ci(est_time=300, suite="base-b-test-4-npu-a3")
register_npu_ci(est_time=300, suite="nightly-4-npu-a3", nightly=True)
test/registered/npu/performance/deepseek_v4_flash/test_npu_deepseek_v4_flash_w8a8_8p_in8k_out1k_50ms.py
+1 -6
View File
@@ -8,12 +8,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="performance testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-perf-16-npu-a3")
# Environment variables for DSV4-Flash single-node PD-mix deployment.
DEEPSEEK_V4_FLASH_W8A8_8P_ENVS = {
test/registered/npu/performance/kimi_k2_6/test_npu_kimi_k2_6_w4a8_8p_in3k5_out1k5_20ms.py
+1 -6
View File
@@ -10,12 +10,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=1800,
suite="full-16-npu-a3",
nightly=True,
disabled="Currently it is executed by the npu performance workflow.",
)
register_npu_ci(est_time=1800, suite="base-c-test-perf-16-npu-a3")
KIMI_K2_6_ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
test/registered/npu/performance/minimax_m2_5/test_npu_minimax_m2_5_w8a8_4p_in64k_out1k_prefix90_50ms.py
+1 -6
View File
@@ -9,12 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="npu-performance",
nightly=True,
disabled="performance testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-perf-8-npu-a3")
MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
test/registered/npu/performance/qwen3-8b/test_npu_qwen3_8b_w8a8_1p_in3k5_out1k5_50ms.py
+1 -6
View File
@@ -9,12 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="performance testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-perf-2-npu-a3")
QWEN3_8B_ENVS = {
"SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "600",
test/registered/npu/performance/qwen3_235b_a22b/test_npu_qwen3_235b_w8a8_8p_in3k5_out1k5_50ms.py
+1 -6
View File
@@ -9,12 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="performance testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-perf-16-npu-a3")
QWEN3_235B_ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
test/registered/npu/performance/qwen3_30b_a3b/test_npu_qwen3_30b_w8a8_1p_in3k5_out1k5_50ms.py
+1 -6
View File
@@ -9,12 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="performance testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-perf-2-npu-a3")
QWEN3_30B_A3B_ENVS = {
"ASCEND_LAUNCH_BLOCKING": "0",
test/registered/npu/performance/qwen3_32b/test_npu_qwen3_32b_w8a8_2p_in3k5_out1k5_50ms.py
+1 -6
View File
@@ -9,12 +9,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="performance testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-perf-4-npu-a3")
QWEN3_32B_ENVS = {
"SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "600",
test/registered/npu/performance/qwen3_5_397b/test_npu_qwen3_5_397b_w4a8_8p_in3k5_out1k5_50ms.py
+1 -6
View File
@@ -8,12 +8,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="nightly-16-npu-a3",
nightly=True,
disabled="performance testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-perf-16-npu-a3")
QWEN3_5_397B_A17B_ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
test/registered/npu/performance/qwen3_6_27b/test_npu_qwen3_6_27b_1p_in1024x1024_30_out1024_50ms.py
+1 -6
View File
@@ -8,12 +8,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="full-2-npu-a3",
nightly=True,
disabled="performance case",
)
register_npu_ci(est_time=3600, suite="base-c-test-perf-2-npu-a3")
QWEN3_6_27B_1024_ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
test/registered/npu/performance/qwen3_6_27b/test_npu_qwen3_6_27b_w8a8_1p_in64k_out1k_50ms.py
+1 -6
View File
@@ -8,12 +8,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="performance testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-perf-2-npu-a3")
QWEN3_6_27B_64K_1K_ENVS = {
"STREAMS_PER_DEVICE": "32",
test/registered/npu/performance/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_in64k_out1k_prefix90_50ms.py
+1 -6
View File
@@ -8,12 +8,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="performance testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-perf-2-npu-a3")
QWEN3_6_35B_A3B_64K_PREFIX_ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
test/registered/npu/performance/qwen3_next_80b_a3b_instruct/test_npu_qwen3_next_80b_w8a8_2p_in6k_out1k5_bs16.py
+1 -6
View File
@@ -7,12 +7,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="performance testcase",
)
register_npu_ci(est_time=3600, suite="base-c-test-perf-4-npu-a3")
QWEN3_NEXT_80B_A3B_ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
test/registered/unit/layers/attention/test_vision_backend_selection.py
+1 -1
View File
@@ -11,7 +11,7 @@ from sglang.srt.layers.attention import vision
from sglang.test.ci.ci_register import register_cpu_ci, register_npu_ci
register_cpu_ci(est_time=2, suite="base-a-test-cpu")
register_npu_ci(est_time=2, suite="stage-b-test-1-npu-a3")
register_npu_ci(est_time=2, suite="base-b-test-1-npu-a3")
register_npu_ci(est_time=2, suite="nightly-1-npu-a3", nightly=True)
test/registered/unit/npu/attention/test_npu_ascend_backend.py
+1 -1
View File
@@ -12,7 +12,7 @@ import torch
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(est_time=5, suite="stage-a-unit-test-npu")
register_npu_ci(est_time=5, suite="base-a-test-1-npu-a2")
# Mock NPU-only modules before importing the source module.
for _ in (
test/registered/unit/npu/attention/test_npu_ascend_dsv4_backend.py
+1 -1
View File
@@ -12,7 +12,7 @@ import torch
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(est_time=4, suite="stage-a-unit-test-npu")
register_npu_ci(est_time=4, suite="base-a-test-1-npu-a2")
for mod in (
"torch_npu",
test/run_suite.py
+14 -7
View File
@@ -93,13 +93,20 @@ PER_COMMIT_SUITES = {
"extra-b-test-8-gpu-h200",
],
HWBackend.NPU: [
"base-a-test-1-gpu-small",
"stage-a-unit-test-npu",
"stage-b-test-1-npu-a3",
"stage-b-test-2-npu-a3",
"stage-b-test-4-npu-a3",
"stage-b-test-8-npu-a3",
"stage-b-test-16-npu-a3",
"base-a-test-1-npu-a2",
"base-b-test-1-npu-a3",
"base-b-test-2-npu-a3",
"base-b-test-4-npu-a3",
"base-b-test-8-npu-a3",
"base-b-test-16-npu-a3",
"base-c-test-acc-2-npu-a3",
"base-c-test-acc-4-npu-a3",
"base-c-test-acc-8-npu-a3",
"base-c-test-acc-16-npu-a3",
"base-c-test-perf-2-npu-a3",
"base-c-test-perf-4-npu-a3",
"base-c-test-perf-8-npu-a3",
"base-c-test-perf-16-npu-a3",
],
HWBackend.XPU: [
"stage-a-test-1-gpu-xpu",