From eae34a416c187c3a28832f8c1455022a003f6290 Mon Sep 17 00:00:00 2001
From: xiegegege <46314656+xiegegege@users.noreply.github.com>
Date: Tue, 25 Nov 2025 19:53:30 +0800
Subject: [PATCH] [benchmark]add qwen3-235b pd+ep yaml (#5225)

---
 benchmarks/yaml/eb45-32k-wint4-mtp-h800-tp4.yaml    |  2 +-
 .../yaml/qwen3-235b-32k-fp8-tp1-dp4_decode.yaml     | 13 +++++++++++++
 .../yaml/qwen3-235b-32k-fp8-tp1-dp4_prefill.yaml    | 13 +++++++++++++
 3 files changed, 27 insertions(+), 1 deletion(-)
 create mode 100644 benchmarks/yaml/qwen3-235b-32k-fp8-tp1-dp4_decode.yaml
 create mode 100644 benchmarks/yaml/qwen3-235b-32k-fp8-tp1-dp4_prefill.yaml

diff --git a/benchmarks/yaml/eb45-32k-wint4-mtp-h800-tp4.yaml b/benchmarks/yaml/eb45-32k-wint4-mtp-h800-tp4.yaml
index 974c2eaf7..c71c247ee 100644
--- a/benchmarks/yaml/eb45-32k-wint4-mtp-h800-tp4.yaml
+++ b/benchmarks/yaml/eb45-32k-wint4-mtp-h800-tp4.yaml
@@ -1,6 +1,6 @@
 max_model_len: 32768
 max_num_seqs: 96
-gpu_memory_utilization: 0.85
+gpu_memory_utilization: 0.8
 kv_cache_ratio: 0.71
 tensor_parallel_size: 4
 quantization: wint4
diff --git a/benchmarks/yaml/qwen3-235b-32k-fp8-tp1-dp4_decode.yaml b/benchmarks/yaml/qwen3-235b-32k-fp8-tp1-dp4_decode.yaml
new file mode 100644
index 000000000..28acf9994
--- /dev/null
+++ b/benchmarks/yaml/qwen3-235b-32k-fp8-tp1-dp4_decode.yaml
@@ -0,0 +1,13 @@
+max_model_len: 32768
+max_num_seqs: 32
+data_parallel_size: 4
+tensor_parallel_size: 1
+enable_expert_parallel: True
+enable_prefix_caching: False
+splitwise_role: decode
+cache_transfer_protocol: "rdma"
+rdma_comm_ports: "7671,7672,7673,7674"
+pd_comm_port: "2335"
+engine_worker_queue_port: "4582,4583,4584,4585"
+graph_optimization_config:
+  use_cudagraph: False
diff --git a/benchmarks/yaml/qwen3-235b-32k-fp8-tp1-dp4_prefill.yaml b/benchmarks/yaml/qwen3-235b-32k-fp8-tp1-dp4_prefill.yaml
new file mode 100644
index 000000000..d9e879e72
--- /dev/null
+++ b/benchmarks/yaml/qwen3-235b-32k-fp8-tp1-dp4_prefill.yaml
@@ -0,0 +1,13 @@
+max_model_len: 32768
+max_num_seqs: 32
+data_parallel_size: 4
+tensor_parallel_size: 1
+enable_expert_parallel: True
+enable_prefix_caching: False
+splitwise_role: prefill
+cache_transfer_protocol: "rdma"
+rdma_comm_ports: "7675,7676,7677,7678"
+pd_comm_port: "2334"
+engine_worker_queue_port: "4368,4369,4360,4361"
+graph_optimization_config:
+  use_cudagraph: False