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

Push test case scripts from test repo to main upstream community repository (#31114)

Browse Source
This commit is contained in:
pllimax
2026-07-16 21:24:01 +08:00
committed by GitHub
parent 7d0fd5101d
commit 4ad418d2c3
14 changed files with 541 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/nightly-test-npu-e2e-single-node.yml
+2
View File
@@ -164,6 +164,8 @@ jobs:
source /usr/local/Ascend/cann/set_env.sh || true
source /usr/local/Ascend/nnal/atb/set_env.sh || true
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash || true
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/custom_transformer/bin/set_env.bash || true
# Set environment of cann
log_path="/root/.cache/tests/logs/log/${current_date}/${tc_name}/${HOSTNAME}"
.github/workflows/nightly-test-npu.yml
+20 -4
View File
@@ -239,10 +239,6 @@ jobs:
- name: glm4_7_flash_1p_aime25
runner: linux-aarch64-a3-2
test_case: test/registered/ascend/accuracy/glm4_7_flash/test_npu_glm4_7_flash_1p_aime25.py
# glm4_6v_flash accuracy tests
- name: glm4_6v_flash_1p_mmmu
runner: linux-aarch64-a3-2
test_case: test/registered/ascend/accuracy/glm4_6v_flash/test_npu_glm4_6v_flash_1p_mmmu.py
# qwen3_vl_8b_thinking accuracy tests
- name: qwen3_vl_8b_thinking_1p_mmmu
runner: linux-aarch64-a3-2
@@ -251,6 +247,11 @@ jobs:
- name: qwen3_vl_30b_a3b_thinking_1p_mmmu
runner: linux-aarch64-a3-2
test_case: test/registered/ascend/accuracy/qwen3_vl_30b_a3b_thinking/test_npu_qwen3_vl_30b_a3b_thinking_1p_mmmu.py
# deepseek_v4_flash performance tests
- name: deepseek_v4_flash_w8a8_8p_in8k_out1k_50ms
runner: linux-aarch64-a3-16
test_case: test/registered/ascend/performance/deepseek_v4_flash/test_npu_deepseek_v4_flash_w8a8_8p_in8k_out1k_50ms.py
test_type: 'perf'
uses: ./.github/workflows/nightly-test-npu-e2e-single-node.yml
with:
runner: ${{ matrix.test_config.runner }}
@@ -278,6 +279,21 @@ jobs:
test_case: test/registered/ascend/performance/glm5_1/test_npu_glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms_aime26.py
test_type: 'perf'
prefill_decode_deployment: 'separation'
# mimo_v2_flash performance tests
- name: test_npu_mimo_v2_flash_1p1d_12p_in16k_out1_ttft_5s
prefill_size: 1
decode_size: 1
router_size: 1
test_case: test/registered/ascend/performance/mimo_v2_flash/test_npu_mimo_v2_flash_1p1d_12p_in16k_out1_ttft_5s.py
test_type: 'perf'
prefill_decode_deployment: 'separation'
- name: test_npu_mimo_v2_flash_1p1d_12p_in16k_out1k_tpot_20ms
prefill_size: 1
decode_size: 1
router_size: 1
test_case: test/registered/ascend/performance/mimo_v2_flash/test_npu_mimo_v2_flash_1p1d_12p_in16k_out1k_tpot_20ms.py
test_type: 'perf'
prefill_decode_deployment: 'separation'
uses: ./.github/workflows/nightly-test-npu-e2e-multi-node.yml
with:
runner: linux-amd64-cpu-8
python/sglang/test/ascend/e2e/test_npu_performance_utils.py
+6
View File
@@ -154,6 +154,9 @@ GLM_5_1_W4A8_MODEL_PATH = "/root/.cache/modelscope/hub/models/Eco-Tech/GLM-5.1-w
MINIMAX_M2_5_W8A8_MODEL_PATH = (
"/root/.cache/modelscope/hub/models/Eco-Tech/MiniMax-M2.5-w8a8-QuaRot"
)
MIMO_V2_FLASH_MODEL_PATH = (
"/root/.cache/modelscope/hub/models/iridiumine/MiMo-V2-Flash-W8A8"
)
MINIMAX_M2_5_EAGLE3_MODEL_PATH = (
"/root/.cache/modelscope/hub/models/sgl-npu/MiniMax-M2.5-eagel-model-0318"
)
@@ -161,6 +164,9 @@ MINIMAX_M2_5_EAGLE3_MODEL_PATH = (
QWEN3_5_397B_W8A8_MODEL_PATH = (
"/root/.cache/modelscope/hub/models/Eco-Tech/Qwen3.5-397B-A17B-w8a8-mtp"
)
DEEPSEEK_V4_FLASH_W8A8_MTP_MODEL_PATH = (
"/root/.cache/modelscope/hub/models/Eco-Tech/DeepSeek-V4-Flash-w8a8-mtp"
)
QWEN3_5_397B_W4A8_MODEL_PATH = (
"/root/.cache/modelscope/hub/models/Eco-Tech/Qwen3.5-397B-A17B-w4a8-mtp"
)
test/registered/ascend/performance/deepseek_v4_flash/test_npu_deepseek_v4_flash_w8a8_8p_in8k_out1k_50ms.py
+127
View File
@@ -0,0 +1,127 @@
import unittest
from sglang.test.ascend.e2e.test_npu_performance_utils import (
AISBENCHMARK_DATASET_DEFAULT,
BENCHMARK_TOOL_DEFAULT,
DEEPSEEK_V4_FLASH_W8A8_MTP_MODEL_PATH,
TestNpuPerformanceTestCaseBase,
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="performance testcase",
)
# 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",
# deepep
"DEEPEP_HCCL_BUFFSIZE": "1000",
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
"DEEPEP_NORMAL_LONG_SEQ_ROUND": "16",
"DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "2048",
"DEEPEP_NORMAL_COMBINE_ENABLE_LONG_SEQ": "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 (EAGLE) related envs
"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",
"--attention-backend",
"dsv4",
"--watchdog-timeout",
9000,
"--mem-fraction-static",
0.7,
"--prefill-max-requests",
2,
"--disable-radix-cache",
"--chunked-prefill-size",
-1,
"--max-running-requests",
160,
"--dp-size",
16,
"--enable-dp-attention",
"--moe-a2a-backend",
"deepep",
"--deepep-mode",
"auto",
"--quantization",
"modelslim",
"--enable-dp-lm-head",
"--kv-cache-dtype",
"bfloat16",
"--cuda-graph-bs",
1,
2,
4,
8,
10,
# MTP (EAGLE) configuration.
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
2,
"--speculative-eagle-topk",
1,
"--speculative-num-draft-tokens",
3,
]
class TestNPUDeepSeekV4FlashW8A88PIn8kOut1k50ms(TestNpuPerformanceTestCaseBase):
"""Test NPU performance for DeepSeek-V4-Flash W8A8 8p in8k out1k."""
benchmark_tool = BENCHMARK_TOOL_DEFAULT
dataset_type = AISBENCHMARK_DATASET_DEFAULT
model = DEEPSEEK_V4_FLASH_W8A8_MTP_MODEL_PATH
other_args = DEEPSEEK_V4_FLASH_W8A8_8P_OTHER_ARGS
envs = DEEPSEEK_V4_FLASH_W8A8_8P_ENVS
dataset_name = "random"
input_len = 8000
output_len = 1000
num_prompts = 160
max_concurrency = 160
random_range_ratio = 1
warmup_requests = 0
request_rate = float("inf")
seed = 1
tpot = 50
output_token_throughput = 1708
def test_npu_deepseek_v4_flash_w8a8_8p_in8k_out1k_50ms(self):
"""Run NPU performance test for DeepSeek-V4-Flash W8A8 8p in8k out1k."""
self.run_throughput()
if __name__ == "__main__":
unittest.main()
test/registered/ascend/performance/glm5_1/test_npu_glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms_aime26.py
+1 -1
View File
@@ -25,7 +25,7 @@ GLM_5_1_PD_SEP_PREFILL_ENVS = {
"STREAMS_PER_DEVICE": "32",
"SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "1200",
"SGLANG_DISAGGREGATION_WAITING_TIMEOUT": "1200",
"HCCL_BUFFSIZE": "1200",
"DEEPEP_HCCL_BUFFSIZE": "1200",
"DEEPEP_NORMAL_LONG_SEQ_ROUND": "72",
"DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "1024",
"DEEPEP_NORMAL_COMBINE_ENABLE_LONG_SEQ": "1",
test/registered/ascend/performance/kimi_k2_6/test_npu_kimi_k2_6_w4a8_16p_in64k_out1k_100ms_aime25.py
+1 -1
View File
@@ -26,7 +26,7 @@ ENVS = {
"STREAMS_PER_DEVICE": "32",
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "64",
"HCCL_BUFFSIZE": "4400",
"DEEPEP_HCCL_BUFFSIZE": "4400",
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
"HCCL_SOCKET_IFNAME": NIC_NAME,
"GLOO_SOCKET_IFNAME": NIC_NAME,
test/registered/ascend/performance/mimo_v2_flash/test_npu_mimo_v2_flash_1p1d_12p_in16k_out1_ttft_5s.py
+191
View File
@@ -0,0 +1,191 @@
import unittest
from sglang.test.ascend.e2e.test_npu_performance_utils import (
AISBENCHMARK_DATASET_DEFAULT,
BENCHMARK_TOOL_DEFAULT,
MIMO_V2_FLASH_MODEL_PATH,
TestNpuPerfMultiNodePdSepTestCaseBase,
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="performance testcase",
)
PREFILL_ENVS = {
"SGLANG_SET_CPU_AFFINITY": "1",
"STREAMS_PER_DEVICE": "32",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "32",
"DEEPEP_HCCL_BUFFSIZE": "1024",
"HCCL_OP_EXPANSION_MODE": "AIV",
"HCCL_SOCKET_IFNAME": "lo",
"GLOO_SOCKET_IFNAME": "lo",
"SGLANG_NPU_PROFILING": "0",
"SGLANG_NPU_PROFILING_STAGE": "prefill",
"DEEPEP_NORMAL_LONG_SEQ_ROUND": "32",
"DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "3584",
"SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "3600",
"SGLANG_DISAGGREGATION_WAITING_TIMEOUT": "3600",
"SGLANG_ENABLE_SPEC_V2": "1",
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "0",
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
"SGLANG_DEEPEP_BF16_DISPATCH": "0",
"SGLANG_DISAGGREGATION_FORCE_QUERY_PREFILL_DP_RANK": "1",
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"HCCL_CONNECT_TIMEOUT": "1800",
"ASCEND_USE_FIA": "1",
}
DECODE_ENVS = {
"SGLANG_SET_CPU_AFFINITY": "1",
"STREAMS_PER_DEVICE": "32",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "128",
"DEEPEP_HCCL_BUFFSIZE": "800",
"HCCL_OP_EXPANSION_MODE": "AIV",
"HCCL_SOCKET_IFNAME": "lo",
"GLOO_SOCKET_IFNAME": "lo",
"SGLANG_NPU_PROFILING": "0",
"SGLANG_NPU_PROFILING_STAGE": "prefill",
"DEEPEP_NORMAL_LONG_SEQ_ROUND": "32",
"DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "3584",
"SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "3600",
"SGLANG_DISAGGREGATION_WAITING_TIMEOUT": "3600",
"SGLANG_ENABLE_SPEC_V2": "1",
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "0",
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
"SGLANG_DEEPEP_BF16_DISPATCH": "0",
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"HCCL_CONNECT_TIMEOUT": "1800",
"SGLANG_PROFILE_WITH_STACK": "True",
"ASCEND_USE_FIA": "1",
}
PREFILL_ARGS = [
"--attention-backend",
"ascend",
"--device",
"npu",
"--tp-size",
8,
"--nnodes",
"1",
"--node-rank",
"0",
"--chunked-prefill-size",
8192,
"--trust-remote-code",
"--max-running-requests",
64,
"--mem-fraction-static",
0.8,
"--swa-full-tokens-ratio",
0.3,
"--disaggregation-mode",
"prefill",
"--disaggregation-transfer-backend",
"ascend",
"--disable-radix-cache",
"--disable-cuda-graph",
"--disable-piecewise-cuda-graph",
"--dp-size",
"2",
]
DECODE_ARGS = [
"--attention-backend",
"ascend",
"--device",
"npu",
"--tp-size",
16,
"--nnodes",
"1",
"--node-rank",
"0",
"--trust-remote-code",
"--max-running-requests",
64,
"--mem-fraction-static",
0.8,
"--swa-full-tokens-ratio",
0.3,
"--cuda-graph-bs",
1,
2,
4,
8,
12,
16,
20,
24,
28,
32,
"--disaggregation-mode",
"decode",
"--disaggregation-transfer-backend",
"ascend",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
3,
"--speculative-eagle-topk",
1,
"--speculative-num-draft-tokens",
4,
"--enable-multi-layer-eagle",
"--disable-radix-cache",
"--dp-size",
"2",
"--enable-dp-attention",
"--enable-dp-lm-head",
"--moe-a2a-backend",
"deepep",
"--deepep-mode",
"low_latency",
]
ROUTER_ARGS = [
"--health-check-interval-secs",
"3600",
"--mini-lb",
]
MODEL_CONFIG = {
"model_path": MIMO_V2_FLASH_MODEL_PATH,
"prefill_args": PREFILL_ARGS,
"decode_args": DECODE_ARGS,
"prefill_envs": PREFILL_ENVS,
"decode_envs": DECODE_ENVS,
"router_args": ROUTER_ARGS,
"router_envs": {},
}
class TestNPUMimo_v2_flash_1P1D_16p_In16k_Out1_TTFT_5s(
TestNpuPerfMultiNodePdSepTestCaseBase
):
"""Test NPU performance for mimo_v2_flash 1P+1D 16p: input_len=16000, output_len=1, 0 cache, TTFT=5s"""
model_config = MODEL_CONFIG
benchmark_tool = BENCHMARK_TOOL_DEFAULT
dataset_type = AISBENCHMARK_DATASET_DEFAULT
dataset_name = "random"
max_concurrency = 64
num_prompts = 128
request_rate = 0.4
input_len = 16000
output_len = 1
random_range_ratio = 1
seed = 1
ttft = 5000
def test_npu_mimo_v2_flash_1p1d_16p_in16k_out1_ttft_5s(self):
"""Run NPU performance test for 1P+1D 16p with 16k input, 1 output, 0 cache, TTFT=5s"""
self.run_throughput()
if __name__ == "__main__":
unittest.main()
test/registered/ascend/performance/mimo_v2_flash/test_npu_mimo_v2_flash_1p1d_12p_in16k_out1k_tpot_20ms.py
+187
View File
@@ -0,0 +1,187 @@
import unittest
from sglang.test.ascend.e2e.test_npu_performance_utils import (
AISBENCHMARK_DATASET_DEFAULT,
BENCHMARK_TOOL_DEFAULT,
MIMO_V2_FLASH_MODEL_PATH,
TestNpuPerfMultiNodePdSepTestCaseBase,
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(
est_time=3600,
suite="",
nightly=True,
disabled="performance testcase",
)
PREFILL_ENVS = {
"SGLANG_SET_CPU_AFFINITY": "1",
"STREAMS_PER_DEVICE": "32",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "32",
"DEEPEP_HCCL_BUFFSIZE": "1024",
"HCCL_OP_EXPANSION_MODE": "AIV",
"HCCL_SOCKET_IFNAME": "lo",
"GLOO_SOCKET_IFNAME": "lo",
"SGLANG_NPU_PROFILING": "0",
"SGLANG_NPU_PROFILING_STAGE": "prefill",
"DEEPEP_NORMAL_LONG_SEQ_ROUND": "32",
"DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "3584",
"SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "3600",
"SGLANG_DISAGGREGATION_WAITING_TIMEOUT": "3600",
"SGLANG_ENABLE_SPEC_V2": "1",
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "0",
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
"SGLANG_DEEPEP_BF16_DISPATCH": "0",
"SGLANG_DISAGGREGATION_FORCE_QUERY_PREFILL_DP_RANK": "1",
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"HCCL_CONNECT_TIMEOUT": "1800",
"ASCEND_USE_FIA": "1",
}
DECODE_ENVS = {
"SGLANG_SET_CPU_AFFINITY": "1",
"STREAMS_PER_DEVICE": "32",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "128",
"DEEPEP_HCCL_BUFFSIZE": "800",
"HCCL_OP_EXPANSION_MODE": "AIV",
"HCCL_SOCKET_IFNAME": "lo",
"GLOO_SOCKET_IFNAME": "lo",
"SGLANG_NPU_PROFILING": "0",
"SGLANG_NPU_PROFILING_STAGE": "prefill",
"DEEPEP_NORMAL_LONG_SEQ_ROUND": "32",
"DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "3584",
"SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "3600",
"SGLANG_DISAGGREGATION_WAITING_TIMEOUT": "3600",
"SGLANG_ENABLE_SPEC_V2": "1",
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "0",
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
"SGLANG_DEEPEP_BF16_DISPATCH": "0",
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"HCCL_CONNECT_TIMEOUT": "1800",
"SGLANG_PROFILE_WITH_STACK": "True",
"ASCEND_USE_FIA": "1",
}
PREFILL_ARGS = [
"--attention-backend",
"ascend",
"--device",
"npu",
"--tp-size",
8,
"--nnodes",
"1",
"--node-rank",
"0",
"--chunked-prefill-size",
8192,
"--trust-remote-code",
"--max-running-requests",
64,
"--mem-fraction-static",
0.8,
"--swa-full-tokens-ratio",
0.3,
"--disaggregation-mode",
"prefill",
"--disaggregation-transfer-backend",
"ascend",
"--disable-radix-cache",
"--disable-cuda-graph",
"--disable-piecewise-cuda-graph",
"--dp-size",
"2",
]
DECODE_ARGS = [
"--attention-backend",
"ascend",
"--device",
"npu",
"--tp-size",
16,
"--nnodes",
"1",
"--node-rank",
"0",
"--trust-remote-code",
"--max-running-requests",
32,
"--mem-fraction-static",
0.8,
"--swa-full-tokens-ratio",
0.3,
"--cuda-graph-bs",
1,
2,
4,
8,
12,
16,
"--disaggregation-mode",
"decode",
"--disaggregation-transfer-backend",
"ascend",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
3,
"--speculative-eagle-topk",
1,
"--speculative-num-draft-tokens",
4,
"--enable-multi-layer-eagle",
"--disable-radix-cache",
"--dp-size",
"2",
"--enable-dp-attention",
"--enable-dp-lm-head",
"--moe-a2a-backend",
"deepep",
"--deepep-mode",
"low_latency",
]
ROUTER_ARGS = [
"--health-check-interval-secs",
"3600",
"--mini-lb",
]
MODEL_CONFIG = {
"model_path": MIMO_V2_FLASH_MODEL_PATH,
"prefill_args": PREFILL_ARGS,
"decode_args": DECODE_ARGS,
"prefill_envs": PREFILL_ENVS,
"decode_envs": DECODE_ENVS,
"router_args": ROUTER_ARGS,
"router_envs": {},
}
class TestNPUMimo_v2_flash_1P1D_16p_In16k_Out1k_TTFT_5s(
TestNpuPerfMultiNodePdSepTestCaseBase
):
"""Test NPU performance for mimo_v2_flash 1P+1D 16p: input_len=16000, output_len=1000, 0 cache, TPOT=20ms"""
model_config = MODEL_CONFIG
benchmark_tool = BENCHMARK_TOOL_DEFAULT
dataset_type = AISBENCHMARK_DATASET_DEFAULT
dataset_name = "random"
max_concurrency = 32
num_prompts = 128
request_rate = float("inf")
input_len = 16000
output_len = 1000
random_range_ratio = 1
seed = 1
tpot = 20
def test_npu_mimo_v2_flash_1p1d_16p_in16k_out1k_tpot_20ms(self):
"""Run NPU performance test for 1P+1D 16p with 16k input, 1k output, 0 cache, TPOT=20ms"""
self.run_throughput()
if __name__ == "__main__":
unittest.main()
test/registered/ascend/performance/minimax_m2_5/test_npu_minimax_m2_5_w8a8_4p_in64k_out1k_prefix90_50ms_gpqa.py
+1 -1
View File
@@ -30,7 +30,7 @@ MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_ENVS = {
"SGLANG_NPU_FUSED_MOE_MODE": "2",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "140000",
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
"HCCL_BUFFSIZE": "1024",
"DEEPEP_HCCL_BUFFSIZE": "1024",
"SGLANG_EXTERNAL_MODEL_PACKAGE": "custom_eagle3",
"PYTHONPATH": f"{MINIMAX_M2_5_EAGLE3_MODEL_PATH}:{os.environ.get('PYTHONPATH', '')}",
}
test/registered/ascend/performance/minimax_m2_5/test_npu_minimax_m2_5_w8a8_8p_in3k5_out1k5_50ms_gpqa.py
+1 -1
View File
@@ -26,7 +26,7 @@ MINIMAX_M2_5_HIGH_THROUGHPUT_ENVS = {
"HCCL_SOCKET_IFNAME": "lo",
"GLOO_SOCKET_IFNAME": "lo",
"TASK_QUEUE_ENABLE": "1",
"HCCL_BUFFSIZE": "1024",
"DEEPEP_HCCL_BUFFSIZE": "1024",
"ASCEND_USE_FIA": "1",
"SGLANG_SET_CPU_AFFINITY": "1",
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
test/registered/ascend/performance/qwen3_30b_a3b/test_npu_qwen3_30b_w8a8_1p_in3k5_out1k5_50ms_aime25.py
+1 -1
View File
@@ -29,7 +29,7 @@ QWEN3_30B_A3B_ENVS = {
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
"SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
"SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "200",
"HCCL_BUFFSIZE": "400",
"DEEPEP_HCCL_BUFFSIZE": "400",
}
QWEN3_30B_A3B_OTHER_ARGS = [
test/registered/ascend/performance/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_in128k_out1k_50ms.py
+1 -1
View File
@@ -18,7 +18,7 @@ register_npu_ci(
QWEN3_6_35B_A3B_128K_1K_ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"STREAMS_PER_DEVICE": "32",
"HCCL_BUFFSIZE": "1600",
"DEEPEP_HCCL_BUFFSIZE": "1600",
"HCCL_SOCKET_IFNAME": "lo",
"GLOO_SOCKET_IFNAME": "lo",
"HCCL_OP_EXPANSION_MODE": "AIV",
test/registered/ascend/performance/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_in64k_out1k_prefix90_50ms_aime26.py
+1 -1
View File
@@ -21,7 +21,7 @@ register_npu_ci(
QWEN3_6_35B_A3B_64K_PREFIX_ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"STREAMS_PER_DEVICE": "32",
"HCCL_BUFFSIZE": "300",
"DEEPEP_HCCL_BUFFSIZE": "300",
"HCCL_SOCKET_IFNAME": "lo",
"GLOO_SOCKET_IFNAME": "lo",
"HCCL_OP_EXPANSION_MODE": "AIV",
test/registered/ascend/performance/qwen3_next_80b_a3b_instruct/test_npu_qwen3_next_80b_w8a8_2p_in6k_out1k5_bs16_aime25.py
+1 -1
View File
@@ -33,7 +33,7 @@ QWEN3_NEXT_80B_A3B_ENVS = {
"SGLANG_WARMUP_TIMEOUT": "3600",
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
"FORCE_DRAFT_MODEL_NON_QUANT": "1",
"HCCL_BUFFSIZE": "2000",
"DEEPEP_HCCL_BUFFSIZE": "2000",
"ZBCCL_LOCAL_MEM_SIZE": "60416",
"SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK": "0",
"ZBCCL_BOOTSTRAP_URL": "tcp://127.0.0.1:24669",