From 0b984c374d634f04cd5ed47ac352b48498ab7a4d Mon Sep 17 00:00:00 2001 From: chx96642264 Date: Tue, 14 Jul 2026 21:18:38 +0800 Subject: [PATCH 1/6] adapt failed deepseek-r1 testcases --- .../test_npu_deepseek_r1_w4a8_1p1d_16p_in3k5_out1k5_50ms.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_1p1d_16p_in3k5_out1k5_50ms.py b/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_1p1d_16p_in3k5_out1k5_50ms.py index 5c5717e37776..24908dba9151 100644 --- a/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_1p1d_16p_in3k5_out1k5_50ms.py +++ b/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_1p1d_16p_in3k5_out1k5_50ms.py @@ -27,6 +27,7 @@ "SGLANG_USE_FIA_NZ": "1", "ENABLE_MOE_NZ": "1", "HCCL_BUFFSIZE": "4000", + "DEEPEP_HCCL_BUFFSIZE": "3200", "DEEP_NORMAL_MODE_USE_INT8_QUANT": "1", "TASK_QUEUE_ENABLE": "2", "HCCL_SOCKET_IFNAME": NIC_NAME, @@ -42,6 +43,7 @@ "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1", "SGLANG_ENABLE_SPEC_V2": "1", "HCCL_BUFFSIZE": "800", + "DEEPEP_HCCL_BUFFSIZE": "800", "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "78", "TASK_QUEUE_ENABLE": "1", "HCCL_SOCKET_IFNAME": NIC_NAME, @@ -154,7 +156,7 @@ "--tool-call-parser", "deepseekv3", "--max-total-tokens", - 145792, + 130000, ], "router_args": [], } From cda7b787ce673e8e0bd310c863170e529a4d595b Mon Sep 17 00:00:00 2001 From: chx96642264 Date: Tue, 14 Jul 2026 21:27:56 +0800 Subject: [PATCH 2/6] adapt failed deepseek-r1 testcases --- .../test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py b/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py index 97963b4c9372..d51406270803 100644 --- a/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py +++ b/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py @@ -22,8 +22,8 @@ "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "200", "HCCL_SOCKET_IFNAME": "lo", "GLOO_SOCKET_IFNAME": "lo", - "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "56", - "HCCL_BUFFSIZE": "2000", + "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "72", + "DEEPEP_HCCL_BUFFSIZE": "2000", "DEEPEP_NORMAL_LONG_SEQ_ROUND": "10", "DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "1024", "DEEP_NORMAL_MODE_USE_INT8_QUANT": "1", @@ -84,7 +84,7 @@ "--tool-call-parser", "deepseekv3", "--max-total-tokens", - 112500, + 100000, ] From c6315765f4af026ca00f44bfa60daa996d68a47e Mon Sep 17 00:00:00 2001 From: chx96642264 Date: Tue, 14 Jul 2026 21:46:50 +0800 Subject: [PATCH 3/6] adapt failed deepseek-r1 testcases --- .../test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py b/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py index d51406270803..7e50a5b7a181 100644 --- a/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py +++ b/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py @@ -58,7 +58,7 @@ 8188, "--disable-radix-cache", "--chunked-prefill-size", - -1, + 65536, "--max-prefill-tokens", 3000, "--moe-a2a-backend", From 009072b8d908d3bd2cb2528fbfeae24c6d64b535 Mon Sep 17 00:00:00 2001 From: chx96642264 Date: Wed, 15 Jul 2026 14:12:41 +0800 Subject: [PATCH 4/6] ds-r1, 8p, 18 --- .../test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py b/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py index 7e50a5b7a181..f2e92e1a2816 100644 --- a/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py +++ b/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py @@ -100,8 +100,8 @@ class TestNPUDeepSeekR1W4A8(TestAscendPerformanceTestCaseBase): input_len = 3500 output_len = 1500 random_range_ratio = 1 - seed = 1 tpot = 50 + seed = 1 output_token_throughput = 3457.1 def test_throughput(self): From 201502144f8fd0ade3d1978dfdac6a75bb060062 Mon Sep 17 00:00:00 2001 From: chx96642264 Date: Thu, 23 Jul 2026 09:27:17 +0800 Subject: [PATCH 5/6] ds-r1, 19 --- .../test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py b/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py index f2e92e1a2816..8396ba7108ca 100644 --- a/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py +++ b/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py @@ -72,7 +72,7 @@ "--speculative-algorithm", "NEXTN", "--speculative-num-steps", - 3, + 2, "--speculative-eagle-topk", 1, "--speculative-num-draft-tokens", From f4470d1b1fd3d803356bcf7a1e3d8a14010693bf Mon Sep 17 00:00:00 2001 From: chx96642264 Date: Fri, 31 Jul 2026 17:25:10 +0800 Subject: [PATCH 6/6] 0731 --- .../test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py b/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py index 8396ba7108ca..1b0ea6fbe5d5 100644 --- a/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py +++ b/test/registered/ascend/performance/deepseek_r1/test_npu_deepseek_r1_w4a8_8p_in3k5_out1k5_50ms.py @@ -19,11 +19,12 @@ "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True", "STREAMS_PER_DEVICE": "32", "SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1", - "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "200", + "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "50", "HCCL_SOCKET_IFNAME": "lo", "GLOO_SOCKET_IFNAME": "lo", "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "72", "DEEPEP_HCCL_BUFFSIZE": "2000", + "HCCL_BUFFSIZE": "300", "DEEPEP_NORMAL_LONG_SEQ_ROUND": "10", "DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "1024", "DEEP_NORMAL_MODE_USE_INT8_QUANT": "1", @@ -49,6 +50,7 @@ 4, 8, 12, + 13, 14, "--mem-fraction-static", 0.9, @@ -84,7 +86,7 @@ "--tool-call-parser", "deepseekv3", "--max-total-tokens", - 100000, + 90000, ]