[benchmark]add lite-vl and x1 yaml (#4130)

benchmarks/yaml/eb45-128k-wint4-a800-tp8.yaml

@@ -6,3 +6,4 @@ tensor_parallel_size: 8
 max_num_batched_tokens: 4096
 max_num_partial_prefills: 3
 max_long_partial_prefills: 3
+quantization: wint4

benchmarks/yaml/eb45-128k-wint8-a800-tp8.yaml

@@ -6,3 +6,4 @@ tensor_parallel_size: 8
 max_num_batched_tokens: 4096
 max_num_partial_prefills: 3
 max_long_partial_prefills: 3
+quantization: wint8

benchmarks/yaml/eb45-32k-wint2-tp4.yaml

@@ -0,0 +1,5 @@
+max_model_len: 32768
+max_num_seqs: 256
+kv_cache_ratio: 0.75
+tensor_parallel_size: 4
+gpu_memory_utilization: 0.9

benchmarks/yaml/eb45-32k-wint4-tp4_decode.yaml

@@ -13,3 +13,4 @@ pd_comm_port: "2334"
 max_num_batched_tokens: 384
 max_num_partial_prefills: 3
 max_long_partial_prefills: 3
+quantization: wint4

benchmarks/yaml/eb45-32k-wint4-tp4_prefill.yaml

@@ -10,3 +10,4 @@ engine_worker_queue_port: 6677
 cache_transfer_protocol: "rdma,ipc"
 rdma_comm_ports: "7675,7676,7677,7678"
 pd_comm_port: "2333"
+quantization: wint4

benchmarks/yaml/eb45-vl-128k-wint4-h800-tp8.yaml

@@ -0,0 +1,11 @@
+enable_mm: True
+max_model_len: 131072
+max_num_seqs: 56
+gpu_memory_utilization: 0.8
+kv_cache_ratio: 0.8
+tensor_parallel_size: 8
+quantization: wint4
+limit_mm_per_prompt: '{"image": 100, "video": 100}'
+enable_chunked_prefill: True
+max_num_batched_tokens: 384
+reasoning_parser: ernie-45-vl

benchmarks/yaml/eb45-vl-32k-wint8-a800-tp8.yaml

@@ -1,7 +1,7 @@
 enable_mm: True
 max_model_len: 32768
 max_num_seqs: 36
-gpu_memory_utilization: 0.95
+gpu_memory_utilization: 0.9
 kv_cache_ratio: 0.8
 tensor_parallel_size: 8
 quantization: wint8

benchmarks/yaml/eb45-vl-32k-wint8-h800-tp8.yaml

@@ -1,7 +1,7 @@
 enable_mm: True
 max_model_len: 32768
 max_num_seqs: 36
-gpu_memory_utilization: 0.8
+gpu_memory_utilization: 0.85
 kv_cache_ratio: 0.8
 tensor_parallel_size: 8
 quantization: wint8

benchmarks/yaml/eb45-vl-lite-32k-bf16-a800-tp1.yaml

@@ -0,0 +1,9 @@
+enable_mm: True
+max_model_len: 32768
+max_num_seqs: 128
+gpu_memory_utilization: 0.9
+kv_cache_ratio: 0.71
+tensor_parallel_size: 1
+enable_chunked_prefill: True
+max_num_batched_tokens: 384
+reasoning_parser: ernie-45-vl

benchmarks/yaml/eb45-vl-lite-32k-wint4-a800-tp1.yaml

@@ -0,0 +1,10 @@
+enable_mm: True
+max_model_len: 32768
+max_num_seqs: 128
+gpu_memory_utilization: 0.9
+kv_cache_ratio: 0.71
+tensor_parallel_size: 1
+enable_chunked_prefill: True
+max_num_batched_tokens: 384
+quantization: wint4
+reasoning_parser: ernie-45-vl

benchmarks/yaml/eb45-vl-lite-32k-wint8-a800-tp1.yaml

@@ -0,0 +1,10 @@
+enable_mm: True
+max_model_len: 32768
+max_num_seqs: 128
+gpu_memory_utilization: 0.9
+kv_cache_ratio: 0.71
+tensor_parallel_size: 1
+enable_chunked_prefill: True
+max_num_batched_tokens: 384
+quantization: wint8
+reasoning_parser: ernie-45-vl

benchmarks/yaml/request_yaml/eb45-vl-128k.yaml

@@ -0,0 +1 @@
+max_tokens: 131071

benchmarks/yaml/request_yaml/eb45-vl-32k.yaml

@@ -0,0 +1 @@
+max_tokens: 12288

benchmarks/yaml/request_yaml/x1-128k.yaml

@@ -2,7 +2,7 @@ top_p: 0.95
 temperature: 0.6
 metadata:
   min_tokens: 1
-max_tokens: 65535
+max_tokens: 131071
 repetition_penalty: 1.0
 frequency_penalty: 0
 presence_penalty: 0

benchmarks/yaml/x1-a3b-128k-wint8-h800-tp1.yaml

@@ -0,0 +1,6 @@
+or_parallel_size: 1
+max_model_len: 131072
+max_num_seqs: 32
+quantization: wint8
+reasoning_parser: ernie_x1
+tool_call_parser: ernie_x1