minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

NPU Nightly Pipeline Skip Test Case Adaptation and Recovery Testing (#26353)

Browse Source
This commit is contained in:
liuxianglong17
2026-05-29 09:29:25 +08:00
committed by GitHub
parent 1c2857b064
commit 6258947039
17 changed files with 152 additions and 119 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_qwen3_next.py
+8 -2
View File
@@ -20,7 +20,6 @@ register_npu_ci(
est_time=200,
suite="nightly-8-npu-a3",
nightly=True,
disabled="https://github.com/Ascend/sglang/issues/58",
)
@@ -56,7 +55,11 @@ class TestQwen3Next(CustomTestCase):
"--watchdog-timeout",
9000,
"--disable-radix-cache",
"--disable-cuda-graph",
"--cuda-graph-bs",
2,
4,
6,
8,
"--max-prefill-tokens",
28672,
"--max-total-tokens",
@@ -69,6 +72,9 @@ class TestQwen3Next(CustomTestCase):
-1,
],
env={
# In NPU scenarios, operators only support BF16 precision.
# This environment variable needs to be set for quantizing weights.
"SGLANG_DEEPEP_BF16_DISPATCH": "1",
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"STREAMS_PER_DEVICE": "32",
"HCCL_OP_EXPANSION_MODE": "AIV",
test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_low_latency_qwen3_next.py
+13 -3
View File
@@ -20,7 +20,6 @@ register_npu_ci(
est_time=200,
suite="nightly-8-npu-a3",
nightly=True,
disabled="https://github.com/Ascend/sglang/issues/58",
)
@@ -56,7 +55,11 @@ class TestQwen3Next(CustomTestCase):
"--watchdog-timeout",
9000,
"--disable-radix-cache",
"--disable-cuda-graph",
"--cuda-graph-bs",
2,
4,
6,
8,
"--chunked-prefill-size",
1024,
"--max-prefill-tokens",
@@ -69,11 +72,18 @@ class TestQwen3Next(CustomTestCase):
"low_latency",
],
env={
# The product of the following two environment variables must be greater than --max-prefill-tokens
# divide by dp size
"DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "3000",
"DEEPEP_NORMAL_LONG_SEQ_ROUND": "10",
# In NPU scenarios, operators only support BF16 precision.
# This environment variable needs to be set for quantizing weights.
"SGLANG_DEEPEP_BF16_DISPATCH": "1",
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"STREAMS_PER_DEVICE": "32",
"HCCL_OP_EXPANSION_MODE": "AIV",
"HCCL_ALGO": "level0:NA;level1:ring",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "20",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "160",
"HCCL_BUFFSIZE": "2048",
**os.environ,
},