[CI][NPU] Fix kimi_k2_6 16p in64k perf test and dsv4-flash testcases (#37760)

python/sglang/test/ascend/e2e/run_npu_testcase.sh

@@ -110,6 +110,9 @@ fi
 # set environment of cann
 . /usr/local/Ascend/cann/set_env.sh
 . /usr/local/Ascend/nnal/atb/set_env.sh
+# Adapt Deepseek-V4-Flash test cases with additional environment variables.
+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
 
 echo "Running test case ${test_case}"
 tc_name=${test_case##*/}

test/registered/npu/accuracy/deepseek_v4_flash/test_npu_deepseek_v4_flash_w8a8_8p_gpqa.py

@@ -39,8 +39,11 @@ DEEPSEEK_V4_FLASH_W8A8_DSPARK_8P_ENVS = {
     "SGLANG_DSPARK_FAST_KERNEL": "0",
     # deepep
     "DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
-    "HCCL_BUFFSIZE": "1400",
+    "DEEPEP_HCCL_BUFFSIZE": "2048",
     "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "64",
+    # war barrier
+    "SGLANG_ENABLE_WAR_BARRIER": "1",
+    "SGLANG_FORCE_COARSE_WAR_BARRIER": "1",
 }
 
 DEEPSEEK_V4_FLASH_W8A8_DSPARK_8P_OTHER_ARGS = [

test/registered/npu/performance/deepseek_v4_flash/test_npu_deepseek_v4_flash_w8a8_1p1d_16p_in8k_out1k_50ms.py

@@ -24,7 +24,13 @@ DEEPSEEK_V4_FLASH_W8A8_1P1D_PREFILL_ENVS = {
     "HCCL_SOCKET_IFNAME": "lo",
     "GLOO_SOCKET_IFNAME": "lo",
     "HCCL_OP_EXPANSION_MODE": "AIV",
+    # deepep
     "DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
+    "DEEPEP_HCCL_BUFFSIZE": "2048",
+    "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "35",
+    # 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",
@@ -36,13 +42,6 @@ DEEPSEEK_V4_FLASH_W8A8_1P1D_PREFILL_ENVS = {
     "SGLANG_OPT_USE_TILELANG_MHC_PRE": "False",
     "SGLANG_OPT_DEEPGEMM_HC_PRENORM": "False",
     "SGLANG_OPT_USE_TILELANG_MHC_POST": "False",
-    # ZBAL
-    "HCCL_BUFFSIZE": "8",
-    "SGLANG_ZBAL_LOCAL_MEM_SIZE": "62084",
-    "SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK": "0",
-    "ZBAL_NPU_ALLOC_CONF": "use_vmm_for_static_memory:True",
-    "SGLANG_ZBAL_BOOTSTRAP_URL": "tcp://127.0.0.1:24669",
-    "ZBAL_ENABLE_GRAPH": "1",
     # PD disagg
     "SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "60",
     # MTP
@@ -80,6 +79,7 @@ DEEPSEEK_V4_FLASH_W8A8_1P1D_DECODE_ENVS = {
     # MTP
     "SGLANG_ENABLE_SPEC_V2": "1",
     "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
+    "SGLANG_NPU_USE_MULTI_STREAM": "1",
 }
 
 # Prefill node launch arguments for DSV4-Flash PD-Sep.
@@ -102,13 +102,13 @@ DEEPSEEK_V4_FLASH_W8A8_1P1D_PREFILL_ARGS = [
     "--disaggregation-bootstrap-port",
     8998,
     "--mem-fraction-static",
-    0.62,
+    0.68,
     "--prefill-max-requests",
     6,
     "--max-prefill-tokens",
-    70000,
+    80000,
     "--chunked-prefill-size",
-    -1,
+    131072,
     "--max-running-requests",
     112,
     "--dp-size",
@@ -152,7 +152,7 @@ DEEPSEEK_V4_FLASH_W8A8_1P1D_DECODE_ARGS = [
     1,
     "--disable-radix-cache",
     "--chunked-prefill-size",
-    32768,
+    -1,
     "--disaggregation-mode",
     "decode",
     "--disaggregation-transfer-backend",
@@ -181,7 +181,6 @@ DEEPSEEK_V4_FLASH_W8A8_1P1D_DECODE_ARGS = [
     36,
     40,
     48,
-    56,
     # MTP (EAGLE) configuration.
     "--speculative-algorithm",
     "EAGLE",
@@ -219,7 +218,7 @@ class TestNPUDeepSeekV4FlashW8A81P1D16PIn8kOut1k50ms(
     num_prompts = 2400
     max_concurrency = 800
     random_range_ratio = 1
-    warmup_requests = 0
+    warmup_requests = 16
     request_rate = float("inf")
     seed = 1
     tpot = 50

test/registered/npu/performance/deepseek_v4_flash/test_npu_deepseek_v4_flash_w8a8_8p_in32k_out1k_50ms.py

@@ -15,27 +15,19 @@ 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",
-    "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "64",
+    "DEEPEP_HCCL_BUFFSIZE": "2048",
-    # zbal
+    "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "35",
-    "HCCL_BUFFSIZE": "8",
+    # war barrier
-    "SGLANG_ZBAL_LOCAL_MEM_SIZE": "61000",
+    "SGLANG_ENABLE_WAR_BARRIER": "1",
-    "SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK": "0",
+    "SGLANG_FORCE_COARSE_WAR_BARRIER": "1",
-    "ZBAL_NPU_ALLOC_CONF": "use_vmm_for_static_memory:True",
-    "SGLANG_ZBAL_BOOTSTRAP_URL": "tcp://127.0.0.1:24669",
-    "ZBAL_ENABLE_GRAPH": "1",
-    # dsv4
-    "IS_DEEPSEEK_V4": "1",
-    "SGLANG_DEBUG_LAYER_NORM": "1",
-    "SGLANG_DEBUG_FWD_INPUT": "1",
-    "USE_FUSED_HC_PRE_ASCENDC": "1",
-    "SGLANG_DSV4_NPU_FUSED_COMPRESSOR": "1",
-    "SGLANG_DSV4_NPU_FUSED_COMPRESSOR_PREFILL": "0",
     # skip gpu branch
     "SGLANG_OPT_FP8_WO_A_GEMM": "0",
     "SGLANG_OPT_USE_OVERLAP_STORE_CACHE": "False",
@@ -50,11 +42,6 @@ DEEPSEEK_V4_FLASH_W8A8_8P_ENVS = {
     # mtp
     "SGLANG_ENABLE_SPEC_V2": "1",
     "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
-    "SGLANG_NPU_PROFILING": "0",
-    "SGLANG_DEBUG_MTP_VERIFY": "0",
-    "SGLANG_DEBUG_MTP_VERIFY_LIMIT": "8",
-    "SGLANG_DEBUG_MTP_VERIFY_ROWS": "4",
-    "SGLANG_DISABLE_DRAFT_EXTEND_GRAPH": "1",
 }
 
 # Server launch arguments for DSV4-Flash W8A8 single-node 8p PD-mix.
@@ -68,12 +55,14 @@ DEEPSEEK_V4_FLASH_W8A8_8P_OTHER_ARGS = [
     "npu",
     "--prefill-max-requests",
     32,
+    "--max-prefill-tokens",
+    80000,
     "--attention-backend",
     "dsv4",
     "--watchdog-timeout",
     9000,
     "--mem-fraction-static",
-    0.7,
+    0.68,
     "--chunked-prefill-size",
     131072,
     "--max-running-requests",
@@ -89,7 +78,7 @@ DEEPSEEK_V4_FLASH_W8A8_8P_OTHER_ARGS = [
     "modelslim",
     "--enable-dp-lm-head",
     "--kv-cache-dtype",
-    "auto",
+    "bfloat16",
     "--skip-server-warmup",
     "--cuda-graph-bs",
     1,
@@ -126,7 +115,7 @@ class TestNPUDeepSeekV4FlashW8A88PIn32kOut1k50ms(TestNpuPerformanceTestCaseBase)
     num_prompts = 64
     max_concurrency = 64
     random_range_ratio = 1
-    warmup_requests = 0
+    warmup_requests = 16
     request_rate = float("inf")
     seed = 1
     tpot = 50

test/registered/npu/performance/deepseek_v4_flash/test_npu_deepseek_v4_flash_w8a8_8p_in8k_out1k_50ms.py

@@ -15,27 +15,18 @@ 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",
-    "USE_NPU_MOE_GATING_TOP_K": "1",
+    "HCCL_OP_EXPANSION_MODE": "AIV",
     "SGLANG_NPU_USE_MULTI_STREAM": "1",
     # deepep
     "DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
-    "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "128",
+    "DEEPEP_HCCL_BUFFSIZE": "2048",
-    # zbal
+    "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "35",
-    "HCCL_BUFFSIZE": "8",
+    # war barrier
-    "SGLANG_ZBAL_LOCAL_MEM_SIZE": "61000",
+    "SGLANG_ENABLE_WAR_BARRIER": "1",
-    "SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK": "0",
+    "SGLANG_FORCE_COARSE_WAR_BARRIER": "1",
-    "ZBAL_NPU_ALLOC_CONF": "use_vmm_for_static_memory:True",
-    "SGLANG_ZBAL_BOOTSTRAP_URL": "tcp://127.0.0.1:24669",
-    "ZBAL_ENABLE_GRAPH": "1",
-    # dsv4
-    "IS_DEEPSEEK_V4": "1",
-    "SGLANG_DEBUG_LAYER_NORM": "1",
-    "SGLANG_DEBUG_FWD_INPUT": "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_FP8_WO_A_GEMM": "0",
     "SGLANG_OPT_USE_OVERLAP_STORE_CACHE": "False",
@@ -50,11 +41,6 @@ DEEPSEEK_V4_FLASH_W8A8_8P_ENVS = {
     # mtp
     "SGLANG_ENABLE_SPEC_V2": "1",
     "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
-    "SGLANG_NPU_PROFILING": "0",
-    "SGLANG_DEBUG_MTP_VERIFY": "0",
-    "SGLANG_DEBUG_MTP_VERIFY_LIMIT": "8",
-    "SGLANG_DEBUG_MTP_VERIFY_ROWS": "4",
-    "SGLANG_DISABLE_DRAFT_EXTEND_GRAPH": "1",
 }
 
 # Server launch arguments for DSV4-Flash W8A8 single-node 8p PD-mix.
@@ -68,12 +54,14 @@ DEEPSEEK_V4_FLASH_W8A8_8P_OTHER_ARGS = [
     "npu",
     "--prefill-max-requests",
     160,
+    "--max-prefill-tokens",
+    80000,
     "--attention-backend",
     "dsv4",
     "--watchdog-timeout",
     9000,
     "--mem-fraction-static",
-    0.7,
+    0.68,
     "--chunked-prefill-size",
     131072,
     "--max-running-requests",
@@ -89,7 +77,7 @@ DEEPSEEK_V4_FLASH_W8A8_8P_OTHER_ARGS = [
     "modelslim",
     "--enable-dp-lm-head",
     "--kv-cache-dtype",
-    "auto",
+    "bfloat16",
     "--skip-server-warmup",
     "--cuda-graph-bs",
     1,
@@ -127,7 +115,7 @@ class TestNPUDeepSeekV4FlashW8A88PIn8kOut1k50ms(TestNpuPerformanceTestCaseBase):
     num_prompts = 160
     max_concurrency = 160
     random_range_ratio = 1
-    warmup_requests = 0
+    warmup_requests = 16
     request_rate = float("inf")
     seed = 1
     tpot = 50

test/registered/npu/performance/kimi_k2_6/test_npu_kimi_k2_6_w4a8_16p_in64k_out1k_100ms.py

@@ -48,7 +48,7 @@ OTHER_ARGS = [
     "--max-running-requests",
     32,
     "--chunked-prefill-size",
-    262144,
+    32768,
     "--context-length",
     75000,
     "--enable-multimodal",