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

[NPU][CI] Add CANN 9.1.0 and Ascend a5 nightly suites (#38833)

Browse Source
This commit is contained in:
pllimax
2026-09-14 22:39:42 +08:00
committed by GitHub
parent 433c999dd0
commit 7465e42b7a
8 changed files with 323 additions and 135 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/npu/accuracy/deepseek_v4_flash/test_npu_deepseek_v4_flash_fp8_4p_gpqa_a5.py
+143
View File
@@ -0,0 +1,143 @@
import unittest
from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
BENCHMARK_TOOL_DEFAULT,
TestNpuAccuracyTestCaseBase,
)
from sglang.test.ascend.e2e.test_npu_performance_utils import (
DEEPSEEK_V4_FLASH_DEFAULT_MODEL_PATH,
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=7200,
suite="nightly-acc-4-npu-a5",
nightly=True,
)
DEEPSEEK_V4_FLASH_FP8_4P_ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"STREAMS_PER_DEVICE": "32",
"SGLANG_SET_CPU_AFFINITY": "1",
"TASK_QUEUE_ENABLE": "1",
"INF_NAN_MODE_FORCE_DISABLE": "1",
"SGLANG_DEFAULT_THINKING": "1",
"SGLANG_DSV4_REASONING_EFFORT": "max",
# HCCL deepep
"HCCL_BUFFSIZE": "1024",
"HCCL_SOCKET_IFNAME": "lo",
"GLOO_SOCKET_IFNAME": "lo",
# DeepEP long-sequence normal combine (ROUND * TOKENS >= chunked_prefill_size / tp * dp)
"DEEPEP_NORMAL_COMBINE_ENABLE_LONG_SEQ": "1",
"DEEPEP_NORMAL_LONG_SEQ_ROUND": "16",
"DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "2048",
# "DEEPEP_HYBRID_DEPLOYMENT": "1",
# dsv4
"IS_DEEPSEEK_V4": "1",
"USE_FUSED_HC_PRE_ASCENDC": "1",
"SGLANG_DSV4_NPU_FUSED_COMPRESSOR": "1",
"SGLANG_DSV4_NPU_FUSED_COMPRESSOR_PREFILL": "1",
# skip gpu branch
"SGLANG_OPT_USE_OVERLAP_STORE_CACHE": "False",
"FORCE_DRAFT_MODEL_NON_QUANT": "1",
"SGLANG_DSV4_FP4_EXPERTS": "True",
"SGLANG_OPT_FUSE_WQA_WKV": "0",
"SGLANG_OPT_BF16_FP32_GEMM_ALGO": "torch",
"SGLANG_OPT_USE_FUSED_HASH_TOPK": "False",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "False",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "False",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "False",
"SGLANG_OPT_FP8_WO_A_GEMM": "False",
# performance
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
"SGLANG_NPU_USE_MULTI_STREAM": "1",
"USE_NPU_MOE_GATING_TOP_K": "1",
"SGLANG_PROFILE_WITH_STACK": "False",
}
DEEPSEEK_V4_FLASH_FP8_4P_OTHER_ARGS = [
"--page-size",
128,
"--tp-size",
4,
"--trust-remote-code",
"--attention-backend",
"dsv4",
"--device",
"npu",
"--watchdog-timeout",
9000,
"--mem-fraction-static",
0.72,
"--max-running-requests",
64,
"--chunked-prefill-size",
131072,
"--max-prefill-tokens",
131072,
"--cuda-graph-bs-decode",
1,
2,
4,
8,
10,
16,
"--kv-cache-dtype",
"auto",
"--enable-dp-lm-head",
"--disable-radix-cache",
"--enable-dp-attention",
"--dp-size",
4,
"--reasoning-parser",
"deepseek-v4",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
2,
"--speculative-eagle-topk",
1,
"--speculative-num-draft-tokens",
3,
"--quantization",
"fp8",
"--moe-a2a-backend",
"deepep",
"--deepep-mode",
"auto",
]
DEEPSEEK_V4_FLASH_FP8_4P_GENERATION_CONFIG_MAX = {
"max_tokens": 125000,
"top_p": 1,
"temperature": 1,
"n": 1,
"extra_body": {
"chat_template_kwargs": {"thinking": True, "reasoning_effort": "max"}
},
}
class TestNPUDeepSeekV4FlashFP84PGPQA(TestNpuAccuracyTestCaseBase):
"""Test NPU accuracy for DeepSeek-V4-Flash FP8 4p EAGLE GPQA."""
benchmark_tool = BENCHMARK_TOOL_DEFAULT
model = DEEPSEEK_V4_FLASH_DEFAULT_MODEL_PATH
other_args = DEEPSEEK_V4_FLASH_FP8_4P_OTHER_ARGS
envs = DEEPSEEK_V4_FLASH_FP8_4P_ENVS
accuracy = 0.874
datasets = ["gpqa_diamond"]
few_shot_num = 0
generation_config = DEEPSEEK_V4_FLASH_FP8_4P_GENERATION_CONFIG_MAX
eval_batch_size = 128
stream = True
timeout = 6000
seed = 1
def test_npu_deepseek_v4_flash_fp8_4p_gpqa(self):
"""Run NPU accuracy test for DeepSeek-V4-Flash FP8 4p EAGLE GPQA."""
self.run_accuracy()
if __name__ == "__main__":
unittest.main()
test/registered/npu/accuracy/deepseek_v4_flash/test_npu_deepseek_v4_flash_w8a8_8p_gpqa.py
+6
View File
@@ -15,6 +15,12 @@ register_npu_ci(
nightly=True,
)
register_npu_ci(
est_time=7200,
suite="nightly-acc-16-npu-a3-cann910",
nightly=True,
)
DEEPSEEK_V4_FLASH_W8A8_DSPARK_8P_ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"STREAMS_PER_DEVICE": "32",
test/registered/npu/performance/deepseek_v4_flash/test_npu_deepseek_v4_flash_w8a8_8p_in32k_out1k_50ms.py
-132
View File
@@ -1,132 +0,0 @@
import unittest
from sglang.test.ascend.e2e.test_npu_performance_utils import (
AISBENCHMARK_DATASET_DEFAULT,
BENCHMARK_TOOL_DEFAULT,
DEEPSEEK_V4_FLASH_0731_W8A8_MODEL_PATH,
TestNpuPerformanceTestCaseBase,
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(est_time=1800, suite="nightly-perf-16-npu-a3", nightly=True)
# Environment variables for DSV4-Flash single-node PD-mix deployment.
DEEPSEEK_V4_FLASH_W8A8_8P_ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"STREAMS_PER_DEVICE": "32",
"INF_NAN_MODE_FORCE_DISABLE": "1",
"SGLANG_SET_CPU_AFFINITY": "1",
"HCCL_SOCKET_IFNAME": "lo",
"GLOO_SOCKET_IFNAME": "lo",
"HCCL_OP_EXPANSION_MODE": "AIV",
"USE_NPU_MOE_GATING_TOP_K": "1",
"SGLANG_NPU_USE_MULTI_STREAM": "1",
# deepep
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
"DEEPEP_HCCL_BUFFSIZE": "2048",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "35",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
# war barrier
"SGLANG_ENABLE_WAR_BARRIER": "1",
"SGLANG_FORCE_COARSE_WAR_BARRIER": "1",
# skip gpu branch
"SGLANG_OPT_FP8_WO_A_GEMM": "0",
"SGLANG_OPT_USE_OVERLAP_STORE_CACHE": "False",
"FORCE_DRAFT_MODEL_NON_QUANT": "1",
"SGLANG_DSV4_FP4_EXPERTS": "False",
"SGLANG_OPT_FUSE_WQA_WKV": "0",
"SGLANG_OPT_BF16_FP32_GEMM_ALGO": "torch",
"SGLANG_OPT_USE_FUSED_HASH_TOPK": "False",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "False",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "False",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "False",
# mtp
"SGLANG_ENABLE_SPEC_V2": "1",
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
}
# Server launch arguments for DSV4-Flash W8A8 single-node 8p PD-mix.
DEEPSEEK_V4_FLASH_W8A8_8P_OTHER_ARGS = [
"--page-size",
128,
"--tp-size",
16,
"--trust-remote-code",
"--device",
"npu",
"--prefill-max-requests",
32,
"--max-prefill-tokens",
80000,
"--attention-backend",
"dsv4",
"--watchdog-timeout",
9000,
"--mem-fraction-static",
0.68,
"--chunked-prefill-size",
131072,
"--max-running-requests",
64,
"--dp-size",
16,
"--enable-dp-attention",
"--moe-a2a-backend",
"deepep",
"--deepep-mode",
"auto",
"--quantization",
"modelslim",
"--enable-dp-lm-head",
"--kv-cache-dtype",
"bfloat16",
"--skip-server-warmup",
"--cuda-graph-bs-decode",
1,
2,
4,
8,
# MTP (EAGLE) configuration.
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
2,
"--speculative-eagle-topk",
1,
"--speculative-num-draft-tokens",
3,
"--ep-size",
16,
"--disable-radix-cache",
]
class TestNPUDeepSeekV4FlashW8A88PIn32kOut1k50ms(TestNpuPerformanceTestCaseBase):
"""Test NPU performance for DeepSeek-V4-Flash W8A8 8p in32k out1k."""
benchmark_tool = BENCHMARK_TOOL_DEFAULT
dataset_type = AISBENCHMARK_DATASET_DEFAULT
model = DEEPSEEK_V4_FLASH_0731_W8A8_MODEL_PATH
other_args = DEEPSEEK_V4_FLASH_W8A8_8P_OTHER_ARGS
envs = DEEPSEEK_V4_FLASH_W8A8_8P_ENVS
dataset_name = "random"
dataset_path = "/root/.cache/modelscope/hub/datasets/gsm8k_deepseekv4/cache0_32000/formal_run1_64_32000_cache0.json"
input_len = 32000
output_len = 1000
num_prompts = 64
max_concurrency = 64
random_range_ratio = 1
warmup_requests = 16
request_rate = float("inf")
seed = 1
tpot = 50
max_attempts = 3
output_token_throughput = 927
def test_npu_deepseek_v4_flash_w8a8_8p_in32k_out1k_50ms(self):
"""Run NPU performance test for DeepSeek-V4-Flash W8A8 8p in32k out1k."""
self.run_throughput()
if __name__ == "__main__":
unittest.main()
test/registered/npu/performance/deepseek_v4_flash/test_npu_deepseek_v4_flash_w8a8_8p_in8k_out1k_50ms.py
+1
View File
@@ -9,6 +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=1800, suite="nightly-perf-16-npu-a3", nightly=True)
register_npu_ci(est_time=1800, suite="nightly-perf-16-npu-a3-cann910", nightly=True)
# Environment variables for DSV4-Flash single-node PD-mix deployment.
DEEPSEEK_V4_FLASH_W8A8_8P_ENVS = {
test/run_suite.py
+3
View File
@@ -170,9 +170,12 @@ NIGHTLY_SUITES = {
"nightly-acc-2-npu-a3",
"nightly-acc-8-npu-a3",
"nightly-acc-16-npu-a3",
"nightly-acc-4-npu-a5",
"nightly-perf-2-npu-a3",
"nightly-perf-4-npu-a3",
"nightly-perf-16-npu-a3",
"nightly-perf-16-npu-a3-cann910",
"nightly-acc-16-npu-a3-cann910",
"full-4-npu-a2",
"full-1-npu-a3",
"full-2-npu-a3",