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

[NPU] Set DEEPEP_HYBRID_DEPLOYMENT=1 for collocated DeepEP test cases (#38667)

Browse Source
This commit is contained in:
pllimax
2026-09-10 00:08:07 +08:00
committed by GitHub
parent ffe98a4279
commit 2b1c4e4c85
14 changed files with 15 additions and 0 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_w8a8_8p_gpqa.py
+1
View File
@@ -41,6 +41,7 @@ DEEPSEEK_V4_FLASH_W8A8_DSPARK_8P_ENVS = {
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
"DEEPEP_HCCL_BUFFSIZE": "2048",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "64",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
# war barrier
"SGLANG_ENABLE_WAR_BARRIER": "1",
"SGLANG_FORCE_COARSE_WAR_BARRIER": "1",
test/registered/npu/accuracy/glm5_2/test_npu_glm_5_2_w4a8_16p_gpqa.py
+1
View File
@@ -28,6 +28,7 @@ GLM_5_2_W4A8_16P_TWO_NODE_ENVS = {
"DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "1024",
"DEEPEP_NORMAL_LONG_SEQ_ROUND": "72",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "32",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
"HCCL_SOCKET_IFNAME": NIC_NAME,
"GLOO_SOCKET_IFNAME": NIC_NAME,
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_v3_2_w8a8.py
+1
View File
@@ -50,6 +50,7 @@ class TestDeepEpDeepseekV32(GSM8KAscendMixin, TestMMLU, CustomTestCase):
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"STREAMS_PER_DEVICE": "32",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "16",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
"HCCL_BUFFSIZE": "1600",
"HCCL_OP_EXPANSION_MODE": "AIV",
"SGLANG_NPU_USE_MLAPO": "0",
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_qwen3_480b.py
+1
View File
@@ -65,6 +65,7 @@ class TestDeepEpQwen(GSM8KAscendMixin, TestMMLU, CustomTestCase):
env = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "600",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
"HCCL_BUFFSIZE": "2100",
"HCCL_OP_EXPANSION_MODE": "AIV",
"TRANSFORMERS_VERBOSITY": "error",
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_qwen3_next.py
+1
View File
@@ -66,6 +66,7 @@ class TestQwen3Next(GSM8KAscendMixin, TestMMLU, CustomTestCase):
"HCCL_OP_EXPANSION_MODE": "AIV",
"HCCL_ALGO": "level0:NA;level1:ring",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "20",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
"HCCL_BUFFSIZE": "2000",
"GDN_ATTN_BACKEND_TRITON": "1",
**os.environ,
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_low_latency_deepseek_v3_2_w8a8.py
+1
View File
@@ -48,6 +48,7 @@ class TestDeepEpDeepseekV32(GSM8KAscendMixin, TestMMLU, CustomTestCase):
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"STREAMS_PER_DEVICE": "32",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "128",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
"HCCL_BUFFSIZE": "2048",
"HCCL_OP_EXPANSION_MODE": "AIV",
"TASK_QUEUE_ENABLE": "0",
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_low_latency_qwen3_480b.py
+1
View File
@@ -66,6 +66,7 @@ class TestDeepEpQwen(GSM8KAscendMixin, TestMMLU, CustomTestCase):
env = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "600",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
"HCCL_BUFFSIZE": "2100",
"HCCL_OP_EXPANSION_MODE": "AIV",
"TRANSFORMERS_VERBOSITY": "error",
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_low_latency_qwen3_next.py
+1
View File
@@ -70,6 +70,7 @@ class TestQwen3Next(GSM8KAscendMixin, TestMMLU, CustomTestCase):
"HCCL_OP_EXPANSION_MODE": "AIV",
"HCCL_ALGO": "level0:NA;level1:ring",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "160",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
"HCCL_BUFFSIZE": "2048",
"GDN_ATTN_BACKEND_TRITON": "1",
**os.environ,
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_eplb_min_rebalancing_utilization_threshold.py
+2
View File
@@ -80,6 +80,7 @@ class TestEplbMinRebalancingUtilizationThresholdBase(CustomTestCase):
"SGLANG_EXPERT_LOCATION_UPDATER_CANARY": "1",
"HCCL_BUFFSIZE": "1024",
"SGLANG_DEEPEP_BF16_DISPATCH": "1",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
"TRANSFORMERS_VERBOSITY": "error",
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
**os.environ,
@@ -155,6 +156,7 @@ class TestEplbMinRebalancingUtilizationThreshold095(
"SGLANG_EXPERT_LOCATION_UPDATER_CANARY": "1",
"HCCL_BUFFSIZE": "1024",
"SGLANG_NPU_DISABLE_ACL_FORMAT_WEIGHT": "1",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
"TRANSFORMERS_VERBOSITY": "error",
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
**os.environ,
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_expert_distribution_recorder_mode.py
+1
View File
@@ -61,6 +61,7 @@ class TestExpertDistributionRecorderModeStatic(CustomTestCase):
env={
"SGLANG_NPU_DISABLE_ACL_FORMAT_WEIGHT": "1",
"HCCL_BUFFSIZE": "1024",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
"SGLANG_EXPERT_DISTRIBUTION_RECORDER_DIR": f"{cls.path}",
"TRANSFORMERS_VERBOSITY": "error",
},
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_moe_dense_tp_size.py
+1
View File
@@ -54,6 +54,7 @@ class TestAscendMoeDenseTPSize(CustomTestCase):
],
env={
"HCCL_BUFFSIZE": "1024",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
},
)
test/registered/npu/performance/deepseek_v4_flash/test_npu_deepseek_v4_flash_w8a8_8p_in32k_out1k_50ms.py
+1
View File
@@ -25,6 +25,7 @@ DEEPSEEK_V4_FLASH_W8A8_8P_ENVS = {
"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",
test/registered/npu/performance/deepseek_v4_flash/test_npu_deepseek_v4_flash_w8a8_8p_in8k_out1k_50ms.py
+1
View File
@@ -24,6 +24,7 @@ DEEPSEEK_V4_FLASH_W8A8_8P_ENVS = {
"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",
test/registered/npu/performance/qwen3_5_397b/test_npu_qwen3_5_397b_w4a8_8p_in3k5_out1k5_50ms.py
+1
View File
@@ -17,6 +17,7 @@ QWEN3_5_397B_A17B_ENVS = {
"STREAMS_PER_DEVICE": "32",
"ASCEND_USE_FIA": "1",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "128",
"DEEPEP_HYBRID_DEPLOYMENT": "1",
"HCCL_BUFFSIZE": "0",
"DEEPEP_NORMAL_LONG_SEQ_ROUND": "6",
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",