@@ -0,0 +1,324 @@
|
||||
test_name: "DeepSeek-V4-Pro-w4a8-PD-no-prefix"
|
||||
model: "Eco-Tech/DeepSeek-V4-Pro-w4a8-mtp"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 16
|
||||
dp_size_local: 8
|
||||
dp_rank_start: 0
|
||||
tp_size: 2
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 16
|
||||
dp_size_local: 8
|
||||
dp_rank_start: 8
|
||||
tp_size: 2
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "10"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_CONNECT_TIMEOUT: "6000"
|
||||
HCCL_BUFFSIZE: "1024"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_CONNECT_TIMEOUT: "1200"
|
||||
HCCL_BUFFSIZE: "1800"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "150000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v4"
|
||||
- --tool-call-parser
|
||||
- "deepseek_v4"
|
||||
- --enable-auto-tool-choice
|
||||
- --reasoning-parser
|
||||
- "deepseek_v4"
|
||||
- --model-loader-extra-config
|
||||
- '{"enable_multithread_load": true, "num_threads": 128}'
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.94"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --block-size
|
||||
- "128"
|
||||
- --enforce-eager
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 1,"method": "mtp"}'
|
||||
- --additional-config
|
||||
- '{"enable_cpu_binding": true, "enable_dsa_cp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeHybridConnector", "kv_role": "kv_producer", "kv_port": "30100", "engine_id": "1", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 8}, "decode": {"dp_size": 16, "tp_size": 2}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "150000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v4"
|
||||
- --tool-call-parser
|
||||
- "deepseek_v4"
|
||||
- --enable-auto-tool-choice
|
||||
- --reasoning-parser
|
||||
- "deepseek_v4"
|
||||
- --model-loader-extra-config
|
||||
- '{"enable_multithread_load": true, "num_threads": 128}'
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.94"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --block-size
|
||||
- "128"
|
||||
- --enforce-eager
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 1,"method": "mtp"}'
|
||||
- --additional-config
|
||||
- '{"enable_cpu_binding": true, "enable_dsa_cp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeHybridConnector", "kv_role": "kv_producer", "kv_port": "30100", "engine_id": "1", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 8}, "decode": {"dp_size": 16, "tp_size": 2}}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "150000"
|
||||
- --max-num-batched-tokens
|
||||
- "120"
|
||||
- --max-num-seqs
|
||||
- "30"
|
||||
- --async-scheduling
|
||||
- --block-size
|
||||
- "128"
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v4"
|
||||
- --tool-call-parser
|
||||
- "deepseek_v4"
|
||||
- --enable-auto-tool-choice
|
||||
- --reasoning-parser
|
||||
- "deepseek_v4"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --model-loader-extra-config
|
||||
- '{"enable_multithread_load": true, "num_threads": 128}'
|
||||
- --trust-remote-code
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3,"method": "mtp"}'
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeHybridConnector", "kv_role": "kv_consumer", "kv_port": "30800", "engine_id": "8", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 8}, "decode": {"dp_size": 16, "tp_size": 2}}}'
|
||||
- --additional-config
|
||||
- '{"ascend_compilation_config":{"enable_npugraph_ex":true,"enable_static_kernel":false},"enable_cpu_binding":true,"multistream_overlap_shared_expert":true,"recompute_scheduler_enable":true}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "150000"
|
||||
- --max-num-batched-tokens
|
||||
- "120"
|
||||
- --max-num-seqs
|
||||
- "30"
|
||||
- --async-scheduling
|
||||
- --block-size
|
||||
- "128"
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v4"
|
||||
- --tool-call-parser
|
||||
- "deepseek_v4"
|
||||
- --enable-auto-tool-choice
|
||||
- --reasoning-parser
|
||||
- "deepseek_v4"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --model-loader-extra-config
|
||||
- '{"enable_multithread_load": true, "num_threads": 128}'
|
||||
- --trust-remote-code
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3,"method": "mtp"}'
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeHybridConnector", "kv_role": "kv_consumer", "kv_port": "30800", "engine_id": "8", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 8}, "decode": {"dp_size": 16, "tp_size": 2}}}'
|
||||
- --additional-config
|
||||
- '{"ascend_compilation_config":{"enable_npugraph_ex":true,"enable_static_kernel":false},"enable_cpu_binding":true,"multistream_overlap_shared_expert":true,"recompute_scheduler_enable":true}'
|
||||
|
||||
benchmarks:
|
||||
perf_TPOT50_3_5_1_5:
|
||||
case_type: performance
|
||||
dataset_path_local: vllm-ascend/GSM8K-in3500-bs4096-prefix0-ds
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 1024
|
||||
max_out_len: 1500
|
||||
batch_size: 256
|
||||
request_rate: 3
|
||||
baseline: 3822.66
|
||||
threshold: 0.95
|
||||
perf_TPOT20_3_5_1_5:
|
||||
case_type: performance
|
||||
dataset_path_local: vllm-ascend/GSM8K-in3500-bs4096-prefix0-ds
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 16
|
||||
max_out_len: 1500
|
||||
batch_size: 4
|
||||
request_rate: 0
|
||||
baseline: 170.83
|
||||
threshold: 0.95
|
||||
perf_TPOT50_64_1:
|
||||
case_type: performance
|
||||
dataset_path_local: GSM8K_prefix0_in65536_bs400_deepseek
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 64
|
||||
max_out_len: 1024
|
||||
batch_size: 16
|
||||
request_rate: 0.25
|
||||
baseline: 223.8
|
||||
threshold: 0.95
|
||||
@@ -0,0 +1,346 @@
|
||||
test_name: "DeepSeek-V4-Pro-w4a8-prefix-cache-PD"
|
||||
model: "Eco-Tech/DeepSeek-V4-Pro-w4a8-mtp"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 16
|
||||
dp_size_local: 8
|
||||
dp_rank_start: 0
|
||||
tp_size: 2
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 16
|
||||
dp_size_local: 8
|
||||
dp_rank_start: 8
|
||||
tp_size: 2
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "10"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_CONNECT_TIMEOUT: "6000"
|
||||
HCCL_BUFFSIZE: "1024"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "4096"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_CONNECT_TIMEOUT: "1200"
|
||||
HCCL_BUFFSIZE: "1800"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "204800"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v4"
|
||||
- --tool-call-parser
|
||||
- "deepseek_v4"
|
||||
- --enable-auto-tool-choice
|
||||
- --reasoning-parser
|
||||
- "deepseek_v4"
|
||||
- --model-loader-extra-config
|
||||
- '{"enable_multithread_load": true, "num_threads": 128}'
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --block-size
|
||||
- "32"
|
||||
- --enforce-eager
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 1,"method": "mtp"}'
|
||||
- --additional-config
|
||||
- '{"enable_cpu_binding": true, "enable_dsa_cp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeHybridConnector", "kv_role": "kv_producer", "kv_port": "30100", "engine_id": "1", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 8}, "decode": {"dp_size": 16, "tp_size": 2}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "204800"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v4"
|
||||
- --tool-call-parser
|
||||
- "deepseek_v4"
|
||||
- --enable-auto-tool-choice
|
||||
- --reasoning-parser
|
||||
- "deepseek_v4"
|
||||
- --model-loader-extra-config
|
||||
- '{"enable_multithread_load": true, "num_threads": 128}'
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --block-size
|
||||
- "32"
|
||||
- --enforce-eager
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 1,"method": "mtp"}'
|
||||
- --additional-config
|
||||
- '{"enable_cpu_binding": true, "enable_dsa_cp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeHybridConnector", "kv_role": "kv_producer", "kv_port": "30100", "engine_id": "1", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 8}, "decode": {"dp_size": 16, "tp_size": 2}}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "204800"
|
||||
- --max-num-batched-tokens
|
||||
- "120"
|
||||
- --max-num-seqs
|
||||
- "30"
|
||||
- --async-scheduling
|
||||
- --block-size
|
||||
- "32"
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v4"
|
||||
- --tool-call-parser
|
||||
- "deepseek_v4"
|
||||
- --enable-auto-tool-choice
|
||||
- --reasoning-parser
|
||||
- "deepseek_v4"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --model-loader-extra-config
|
||||
- '{"enable_multithread_load": true, "num_threads": 128}'
|
||||
- --trust-remote-code
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3,"method": "mtp"}'
|
||||
- --gpu-memory-utilization
|
||||
- "0.96"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeHybridConnector", "kv_role": "kv_consumer", "kv_port": "30800", "engine_id": "8", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 8}, "decode": {"dp_size": 16, "tp_size": 2}}}'
|
||||
- --additional-config
|
||||
- '{"ascend_compilation_config":{"enable_npugraph_ex":true,"enable_static_kernel":false},"enable_cpu_binding":true,"multistream_overlap_shared_expert":true,"recompute_scheduler_enable":true}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "204800"
|
||||
- --max-num-batched-tokens
|
||||
- "120"
|
||||
- --max-num-seqs
|
||||
- "30"
|
||||
- --async-scheduling
|
||||
- --block-size
|
||||
- "32"
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v4"
|
||||
- --tool-call-parser
|
||||
- "deepseek_v4"
|
||||
- --enable-auto-tool-choice
|
||||
- --reasoning-parser
|
||||
- "deepseek_v4"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --model-loader-extra-config
|
||||
- '{"enable_multithread_load": true, "num_threads": 128}'
|
||||
- --trust-remote-code
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3,"method": "mtp"}'
|
||||
- --gpu-memory-utilization
|
||||
- "0.96"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeHybridConnector", "kv_role": "kv_consumer", "kv_port": "30800", "engine_id": "8", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 8}, "decode": {"dp_size": 16, "tp_size": 2}}}'
|
||||
- --additional-config
|
||||
- '{"ascend_compilation_config":{"enable_npugraph_ex":true,"enable_static_kernel":false},"enable_cpu_binding":true,"multistream_overlap_shared_expert":true,"recompute_scheduler_enable":true}'
|
||||
|
||||
benchmarks:
|
||||
perf_TPOT50_64_1_prefix90_warmup:
|
||||
case_type: performance
|
||||
dataset_path_local: vllm-ascend/GSM8K-in65536-bs512-prefix90-ds
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 4
|
||||
max_out_len: 1
|
||||
batch_size: 4
|
||||
request_rate: 0
|
||||
baseline: 0
|
||||
threshold: 0.95
|
||||
perf_TPOT50_64_1_prefix90:
|
||||
case_type: performance
|
||||
dataset_path_local: vllm-ascend/GSM8K-in65536-bs512-prefix90-ds
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 384
|
||||
max_out_len: 1024
|
||||
batch_size: 96
|
||||
request_rate: 2.5
|
||||
baseline: 1951.6
|
||||
threshold: 0.95
|
||||
perf_TPOT50_128k_1_prefix_warmup:
|
||||
case_type: performance
|
||||
dataset_path_local: vllm-ascend/GSM8K-in128k-bs512-prefix90-ds
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 4
|
||||
max_out_len: 1
|
||||
batch_size: 4
|
||||
request_rate: 0
|
||||
baseline: 0
|
||||
threshold: 0.95
|
||||
perf_TPOT50_128k_1_prefix90:
|
||||
case_type: performance
|
||||
dataset_path_local: vllm-ascend/GSM8K-in128k-bs512-prefix90-ds
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 192
|
||||
max_out_len: 1024
|
||||
batch_size: 48
|
||||
request_rate: 1
|
||||
baseline: 869.13
|
||||
threshold: 0.95
|
||||
acc_aime2025:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/aime2025
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
|
||||
temperature: 1.0
|
||||
top_p: 1.0
|
||||
thinking: "true"
|
||||
max_out_len: 65536
|
||||
batch_size: 32
|
||||
baseline: 90
|
||||
threshold: 10
|
||||
@@ -0,0 +1,201 @@
|
||||
test_name: "DeepSeek-V4-Flash-w8a8-PD-no-prefix"
|
||||
model: "Eco-Tech/DeepSeek-V4-Flash-w8a8-mtp"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0 ]
|
||||
decoder: [ 1 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 16
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 0
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "120"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "10"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1024"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "4096"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1500"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "1048576"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --no-enable-prefix-caching
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --model-loader-extra-config
|
||||
- '{"enable_multithread_load": true, "num_threads": 128}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 1,"method": "mtp"}'
|
||||
- --trust-remote-code
|
||||
- --block-size
|
||||
- "32"
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v4"
|
||||
- --tool-call-parser
|
||||
- "deepseek_v4"
|
||||
- --enable-auto-tool-choice
|
||||
- --reasoning-parser
|
||||
- "deepseek_v4"
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enforce-eager
|
||||
- --additional-config
|
||||
- '{"enable_cpu_binding": true, "enable_dsa_cp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeHybridConnector", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 4}, "decode": {"dp_size": 16, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "1048576"
|
||||
- --max-num-batched-tokens
|
||||
- "240"
|
||||
- --max-num-seqs
|
||||
- "60"
|
||||
- --async-scheduling
|
||||
- --block-size
|
||||
- "32"
|
||||
- --no-enable-prefix-caching
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --model-loader-extra-config
|
||||
- '{"enable_multithread_load": true, "num_threads": 128}'
|
||||
- --trust-remote-code
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v4"
|
||||
- --tool-call-parser
|
||||
- "deepseek_v4"
|
||||
- --enable-auto-tool-choice
|
||||
- --reasoning-parser
|
||||
- "deepseek_v4"
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3,"method": "mtp"}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeHybridConnector", "kv_role": "kv_consumer", "kv_port": "30100", "engine_id": "1", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 4}, "decode": {"dp_size": 16, "tp_size": 1}}}'
|
||||
- --additional-config
|
||||
- '{"ascend_compilation_config":{"enable_npugraph_ex":true,"enable_static_kernel":false},"enable_cpu_binding":true,"multistream_overlap_shared_expert":true,"recompute_scheduler_enable":true}'
|
||||
|
||||
benchmarks:
|
||||
perf_TPOT50_3_5_1_5:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs4096-prefix0-ds
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 4096
|
||||
max_out_len: 1500
|
||||
batch_size: 1024
|
||||
request_rate: 12
|
||||
baseline: 15569.48
|
||||
threshold: 0.95
|
||||
perf_TPOT20_3_5_1_5:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs4096-prefix0-ds
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 400
|
||||
max_out_len: 1500
|
||||
batch_size: 100
|
||||
request_rate: 12
|
||||
baseline: 4071.471
|
||||
threshold: 0.95
|
||||
perf_TPOT20_3_5_1_5_top_k:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs4096-prefix0-ds
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
temperature: 0.6
|
||||
top_k: -1
|
||||
top_p: 0.95
|
||||
num_prompts: 400
|
||||
max_out_len: 1500
|
||||
batch_size: 100
|
||||
request_rate: 12
|
||||
baseline: 4071.471
|
||||
threshold: 0.95
|
||||
@@ -0,0 +1,265 @@
|
||||
test_name: "DeepSeek_V3.1T_MTP1_128K_1K_PD"
|
||||
model: "Eco-Tech/DeepSeek-V3.1-Terminus-w8a8-mtp-QuaRot"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 0
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 16
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "400"
|
||||
LD_PRELOAD: "/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "10"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
TASK_QUEUE_ENABLE: 1
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_USE_V1: "1"
|
||||
LD_LIBRARY_PATH: "/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH"
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1200"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "135000"
|
||||
- --max-num-batched-tokens
|
||||
- "1024"
|
||||
- --max-num-seqs
|
||||
- "1"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.93"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":false,"enable_shared_expert_dp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "135000"
|
||||
- --max-num-batched-tokens
|
||||
- "1024"
|
||||
- --max-num-seqs
|
||||
- "1"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.93"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":false,"enable_shared_expert_dp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30100", "engine_id": "1", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "135000"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "1"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.94"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"recompute_scheduler_enable":true,"multistream_overlap_shared_expert": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "135000"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "1"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.94"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"recompute_scheduler_enable":true,"multistream_overlap_shared_expert": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
benchmarks:
|
||||
perf_wram_up:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in1024-bs210
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 100
|
||||
max_out_len: 1
|
||||
batch_size: 100
|
||||
request_rate: 0
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
perf_1:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in128k-bs512-prefix90-ds
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 1
|
||||
max_out_len: 1024
|
||||
batch_size: 1
|
||||
request_rate: 0
|
||||
baseline: 7
|
||||
threshold: 0.97
|
||||
|
||||
@@ -0,0 +1,255 @@
|
||||
test_name: "DeepSeek_V3.1T_MTP1_3_5K_1_5K_PD"
|
||||
model: "Eco-Tech/DeepSeek-V3.1-Terminus-w8a8-mtp-QuaRot"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 0
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 16
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
ASCEND_LOG_DEVICE_FLUSH_TIMEOUT: "0"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "400"
|
||||
LD_PRELOAD: "/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "10"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
TASK_QUEUE_ENABLE: 1
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_USE_V1: "1"
|
||||
LD_LIBRARY_PATH: "/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH"
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1200"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_shared_expert_dp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_shared_expert_dp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30100", "engine_id": "1", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "40"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.94"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"recompute_scheduler_enable":true,"multistream_overlap_shared_expert": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "40"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.94"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"recompute_scheduler_enable":true,"multistream_overlap_shared_expert": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
benchmarks:
|
||||
perf_1:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs2800
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 2400
|
||||
max_out_len: 1500
|
||||
batch_size: 700
|
||||
request_rate: 0
|
||||
baseline: 9800
|
||||
threshold: 0.97
|
||||
|
||||
@@ -0,0 +1,257 @@
|
||||
test_name: "DeepSeek_V3.1T_MTP1_PD"
|
||||
model: "Eco-Tech/DeepSeek-V3.1-Terminus-w8a8-mtp-QuaRot"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 0
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 16
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
ASCEND_LOG_DEVICE_FLUSH_TIMEOUT: "0"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "400"
|
||||
LD_PRELOAD: "/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "10"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
TASK_QUEUE_ENABLE: 1
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_USE_V1: "1"
|
||||
LD_LIBRARY_PATH: "/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH"
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1200"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_shared_expert_dp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_shared_expert_dp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30100", "engine_id": "1", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "40"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.94"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"recompute_scheduler_enable":true,"multistream_overlap_shared_expert": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "40"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.94"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"recompute_scheduler_enable":true,"multistream_overlap_shared_expert": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
benchmarks:
|
||||
acc:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/gsm8k
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_chat_prompt
|
||||
max_out_len: 32768
|
||||
batch_size: 64
|
||||
baseline: 96.29
|
||||
threshold: 1
|
||||
temperature: 0.6
|
||||
top_p: 0.95
|
||||
ignore_eos: false
|
||||
thinking: true
|
||||
|
||||
@@ -0,0 +1,257 @@
|
||||
test_name: "DeepSeek_V3.1T_MTP3_PD"
|
||||
model: "Eco-Tech/DeepSeek-V3.1-Terminus-w8a8-mtp-QuaRot"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 0
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 16
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
ASCEND_LOG_DEVICE_FLUSH_TIMEOUT: "0"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "400"
|
||||
LD_PRELOAD: "/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "10"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
TASK_QUEUE_ENABLE: 1
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_USE_V1: "1"
|
||||
LD_LIBRARY_PATH: "/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH"
|
||||
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1200"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 3}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_shared_expert_dp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 3}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_shared_expert_dp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30100", "engine_id": "1", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "40"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.94"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 3}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"recompute_scheduler_enable":true,"multistream_overlap_shared_expert": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "40"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.94"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 3}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"recompute_scheduler_enable":true,"multistream_overlap_shared_expert": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
benchmarks:
|
||||
acc:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/gsm8k
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_chat_prompt
|
||||
max_out_len: 32768
|
||||
batch_size: 64
|
||||
baseline: 96.29
|
||||
threshold: 1
|
||||
temperature: 0.6
|
||||
top_p: 0.95
|
||||
ignore_eos: false
|
||||
thinking: true
|
||||
@@ -0,0 +1,262 @@
|
||||
test_name: "DeepSeek_V3.1T_layerwise_PD"
|
||||
model: "Eco-Tech/DeepSeek-V3.1-Terminus-w8a8-mtp-QuaRot"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 0
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 16
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
SERVER_PORT: "${PORT}"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
ASCEND_LOG_DEVICE_FLUSH_TIMEOUT: "0"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "400"
|
||||
LD_PRELOAD: "/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "10"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
TASK_QUEUE_ENABLE: 1
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_USE_V1: "1"
|
||||
PYTHONHASHSEED: "0"
|
||||
ASCEND_CONNECT_TIMEOUT: "10000"
|
||||
ASCEND_TRANSFER_TIMEOUT: "10000"
|
||||
LD_LIBRARY_PATH: "/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH"
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1100"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeLayerwiseConnector", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeLayerwiseConnector", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "28"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[3,4,8,16,24,32,48,56]}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"multistream_overlap_shared_expert": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeLayerwiseConnector", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "28"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 1}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[3,4,8,16,24,32,48,56]}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"multistream_overlap_shared_expert": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeLayerwiseConnector", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
benchmarks:
|
||||
acc:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/gpqa
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: gpqa/gpqa_gen_0_shot_cot_chat_prompt
|
||||
max_out_len: 32768
|
||||
batch_size: 64
|
||||
baseline: 80.3
|
||||
threshold: 5
|
||||
temperature: 0.6
|
||||
top_p: 0.95
|
||||
ignore_eos: false
|
||||
thinking: true
|
||||
@@ -0,0 +1,280 @@
|
||||
test_name: "DeepSeek_V3.2T_MTP2_PD"
|
||||
model: "vllm-ascend/DeepSeek-V3.2-W8A8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 1
|
||||
dp_rank_start: 0
|
||||
tp_size: 16
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 1
|
||||
dp_rank_start: 1
|
||||
tp_size: 16
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
ASCEND_LOG_DEVICE_FLUSH_TIMEOUT: "0"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "400"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "10"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
TASK_QUEUE_ENABLE: 1
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_USE_V1: "1"
|
||||
ASCEND_TRANSPORT_PRINT: "1"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
LD_LIBRARY_PATH: "/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
VLLM_NIXL_ABORT_REQUEST_TIMEOUT: "300000"
|
||||
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "360"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1100"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
DYNAMIC_EPLB: "true"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v32"
|
||||
- --reasoning-parser
|
||||
- "deepseek_v3"
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 2}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 16}, "decode": {"dp_size": 8, "tp_size": 4}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v32"
|
||||
- --reasoning-parser
|
||||
- "deepseek_v3"
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 2}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 16}, "decode": {"dp_size": 8, "tp_size": 4}}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "100"
|
||||
- --max-num-seqs
|
||||
- "40"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.93"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v32"
|
||||
- --reasoning-parser
|
||||
- "deepseek_v3"
|
||||
- --async-scheduling
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 2}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[3,6,9,12,15,18,21,24,27,30,33,36,39,42]}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true,"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":2000,"algorithm_execution_interval":50, "eplb_policy_type":3, "num_redundant_experts":0}}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 16}, "decode": {"dp_size": 8, "tp_size": 4}}}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "100"
|
||||
- --max-num-seqs
|
||||
- "40"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.93"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v32"
|
||||
- --reasoning-parser
|
||||
- "deepseek_v3"
|
||||
- --async-scheduling
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 2}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[3,6,9,12,15,18,21,24,27,30,33,36,39,42]}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true,"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":2000,"algorithm_execution_interval":50, "eplb_policy_type":3, "num_redundant_experts":0}}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 16}, "decode": {"dp_size": 8, "tp_size": 4}}}'
|
||||
|
||||
benchmarks:
|
||||
acc:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/aime2025
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
|
||||
max_out_len: 65536
|
||||
batch_size: 32
|
||||
baseline: 89.9
|
||||
threshold: 10
|
||||
temperature: 1.0
|
||||
top_p: 0.95
|
||||
ignore_eos: false
|
||||
thinking: true
|
||||
|
||||
@@ -0,0 +1,276 @@
|
||||
test_name: "DeepSeek_V3.2T_MTP3_PD"
|
||||
model: "vllm-ascend/DeepSeek-V3.2-W8A8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
ASCEND_LOG_DEVICE_FLUSH_TIMEOUT: "0"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "400"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "10"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
TASK_QUEUE_ENABLE: 1
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_USE_V1: "1"
|
||||
ASCEND_TRANSPORT_PRINT: "1"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
LD_LIBRARY_PATH: "/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
VLLM_NIXL_ABORT_REQUEST_TIMEOUT: "300000"
|
||||
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "360"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1200"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
DYNAMIC_EPLB: "true"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "8000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v32"
|
||||
- --reasoning-parser
|
||||
- "deepseek_v3"
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 3}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 8}, "decode": {"dp_size": 8, "tp_size": 4}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "8000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v32"
|
||||
- --reasoning-parser
|
||||
- "deepseek_v3"
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 3}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 8}, "decode": {"dp_size": 8, "tp_size": 4}}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "8000"
|
||||
- --max-num-batched-tokens
|
||||
- "100"
|
||||
- --max-num-seqs
|
||||
- "40"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v32"
|
||||
- --reasoning-parser
|
||||
- "deepseek_v3"
|
||||
- --async-scheduling
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 3}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64]}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true,"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":2000,"algorithm_execution_interval":50, "eplb_policy_type":3, "num_redundant_experts":0}}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 8}, "decode": {"dp_size": 8, "tp_size": 4}}}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "8000"
|
||||
- --max-num-batched-tokens
|
||||
- "100"
|
||||
- --max-num-seqs
|
||||
- "40"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --tokenizer-mode
|
||||
- "deepseek_v32"
|
||||
- --reasoning-parser
|
||||
- "deepseek_v3"
|
||||
- --async-scheduling
|
||||
- --speculative-config
|
||||
- '{"method": "deepseek_mtp", "num_speculative_tokens": 3}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64]}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true,"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":2000,"algorithm_execution_interval":50, "eplb_policy_type":3, "num_redundant_experts":0}}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 4, "tp_size": 8}, "decode": {"dp_size": 8, "tp_size": 4}}}'
|
||||
|
||||
benchmarks:
|
||||
perf_1:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_in2048_bs4000_deepseek3_2
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 128
|
||||
max_out_len: 2048
|
||||
batch_size: 32
|
||||
request_rate: 0
|
||||
baseline: 1460.4861
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,335 @@
|
||||
test_name: "GLM-5.1-744B-128k-1k-90-50-PD"
|
||||
model: "Eco-Tech/GLM-5.1-w8a8c8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "202752"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true,"enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true,"c8_enable_reshape_optim":true}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30000",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "202752"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true,"enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true,"c8_enable_reshape_optim":true}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30000",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "202752"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true,"enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --enable-prefix-caching
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30200",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "202752"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true,"enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --enable-prefix-caching
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30200",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
|
||||
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in131072-bs500-prefix90-glm51
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 256
|
||||
max_out_len: 1024
|
||||
batch_size: 64
|
||||
request_rate: 1.52
|
||||
baseline: 1294.76
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,334 @@
|
||||
test_name: "GLM-5.1-744B-3.5k-1.5k-0-20"
|
||||
model: "Eco-Tech/GLM-5.1-w8a8c8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0, 1]
|
||||
decoder: [2, 3]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false, "c8_enable_reshape_optim":false}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --no-enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30001",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false, "c8_enable_reshape_optim":false}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --no-enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30001",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30203",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false}'
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30203",
|
||||
"engine_id": "3",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
|
||||
benchmarks:
|
||||
perf_TPOT20:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs2800
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 160
|
||||
max_out_len: 1500
|
||||
batch_size: 40
|
||||
request_rate: 6.87
|
||||
baseline: 1456.52
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,336 @@
|
||||
test_name: "GLM-5.1-744B-3.5k-1.5k-0-50"
|
||||
model: "Eco-Tech/GLM-5.1-w8a8c8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0, 1]
|
||||
decoder: [2, 3]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 0
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 16
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false, "c8_enable_reshape_optim":false}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --no-enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30001",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 32,
|
||||
"tp_size": 1
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false, "c8_enable_reshape_optim":false}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --no-enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30001",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 32,
|
||||
"tp_size": 1
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false, "finegrained_tp_config": {"lmhead_tensor_parallel_size":16}}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30203",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 32,
|
||||
"tp_size": 1
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false, "finegrained_tp_config": {"lmhead_tensor_parallel_size":16}}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30203",
|
||||
"engine_id": "3",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 32,
|
||||
"tp_size": 1
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
|
||||
benchmarks:
|
||||
perf_TPOT50:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs2800
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 1920
|
||||
max_out_len: 1500
|
||||
batch_size: 480
|
||||
request_rate: 6.5
|
||||
baseline: 8240.61
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,343 @@
|
||||
test_name: "GLM-5.1-744B-64k-1k-90-50"
|
||||
model: "Eco-Tech/GLM-5.1-w8a8c8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "c8_enable_reshape_optim":true}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30001",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "c8_enable_reshape_optim":true}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30001",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --enable-auto-tool-choice
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30200",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30200",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
|
||||
benchmarks:
|
||||
acc_aime2025:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/aime2025
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
|
||||
max_out_len: 32768
|
||||
batch_size: 32
|
||||
baseline: 93.33
|
||||
threshold: 10
|
||||
temperature: 1.0
|
||||
top_p: 0.95
|
||||
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in65536-bs1500-prefix90-glm51
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 512
|
||||
max_out_len: 1024
|
||||
batch_size: 128
|
||||
request_rate: 3.15
|
||||
baseline: 2722.60
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,339 @@
|
||||
test_name: "GLM-5.2-w4a8c8-128k-1k-90-50-PD"
|
||||
model: "Eco-Tech/GLM-5.2-w4a8c8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "202752"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true,"enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true,"c8_enable_reshape_optim":true}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30000",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "202752"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true,"enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true,"c8_enable_reshape_optim":true}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30000",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "202752"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true,"enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --enable-prefix-caching
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30200",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "202752"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true,"enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --enable-prefix-caching
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30200",
|
||||
"engine_id": "3",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
|
||||
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in131072-bs500-prefix90-glm51
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 256
|
||||
max_out_len: 1024
|
||||
batch_size: 64
|
||||
request_rate: 1.52
|
||||
baseline: 1294.76
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,334 @@
|
||||
test_name: "GLM-5.2-w4a8c8-3.5k-1.5k-0-20"
|
||||
model: "Eco-Tech/GLM-5.2-w4a8c8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0, 1]
|
||||
decoder: [2, 3]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false, "c8_enable_reshape_optim":false}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --no-enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30001",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false, "c8_enable_reshape_optim":false}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --no-enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30001",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30203",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false}'
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30203",
|
||||
"engine_id": "3",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
|
||||
benchmarks:
|
||||
perf_TPOT20:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs2800
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 160
|
||||
max_out_len: 1500
|
||||
batch_size: 40
|
||||
request_rate: 6.87
|
||||
baseline: 1456.52
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,336 @@
|
||||
test_name: "Eco-Tech/GLM-5.2-w4a8c8-3.5k-1.5k-0-50"
|
||||
model: "Eco-Tech/GLM-5.2-w4a8c8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0, 1]
|
||||
decoder: [2, 3]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 0
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 16
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false, "c8_enable_reshape_optim":false}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --no-enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30001",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 32,
|
||||
"tp_size": 1
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false, "c8_enable_reshape_optim":false}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --no-enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30001",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 32,
|
||||
"tp_size": 1
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false, "finegrained_tp_config": {"lmhead_tensor_parallel_size":16}}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30203",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 32,
|
||||
"tp_size": 1
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "18432"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": false, "finegrained_tp_config": {"lmhead_tensor_parallel_size":16}}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30203",
|
||||
"engine_id": "3",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 32,
|
||||
"tp_size": 1
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
|
||||
benchmarks:
|
||||
perf_TPOT50:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs2800
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 1920
|
||||
max_out_len: 1500
|
||||
batch_size: 480
|
||||
request_rate: 6.5
|
||||
baseline: 8240.61
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,347 @@
|
||||
test_name: "GLM-5.2-w4a8c8-64k-1k-90-50"
|
||||
model: "Eco-Tech/GLM-5.2-w4a8c8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0, 1]
|
||||
decoder: [2, 3]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "c8_enable_reshape_optim":true}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30001",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable" : false,"multistream_overlap_shared_expert": true, "enable_dsa_cp":true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "c8_enable_reshape_optim":true}'
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --trust-remote-code
|
||||
- --enable-prefix-caching
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30001",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --enable-auto-tool-choice
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30200",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --hf-overrides
|
||||
- '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --max-num-batched-tokens
|
||||
- "164"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp","enforce_eager":true}'
|
||||
- --additional-config
|
||||
- '{"fuse_muls_add":true,"recompute_scheduler_enable":true,"multistream_overlap_shared_expert":true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30200",
|
||||
"engine_id": "3",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
|
||||
benchmarks:
|
||||
acc_aime2025:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/aime2025
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
|
||||
max_out_len: 32768
|
||||
batch_size: 32
|
||||
baseline: 93.33
|
||||
threshold: 10
|
||||
temperature: 1.0
|
||||
top_p: 0.95
|
||||
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in65536-bs1500-prefix90-glm51
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 512
|
||||
max_out_len: 1024
|
||||
batch_size: 128
|
||||
request_rate: 3.15
|
||||
baseline: 2722.60
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,334 @@
|
||||
test_name: "GLM_5_1_PD-32k-0.5k-64-0"
|
||||
model: "Eco-Tech/GLM-5.1-w8a8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0, 1]
|
||||
decoder: [2, 3]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
HCCL_BUFFSIZE: "256"
|
||||
VLLM_TORCH_PROFILER_WITH_STACK: "0"
|
||||
ASCEND_AGGREGATE_ENABLE: "1"
|
||||
ASCEND_TRANSPORT_PRINT: "1"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
VLLM_NIXL_ABORT_REQUEST_TIMEOUT: "300000"
|
||||
VLLM_ENGINE_READY_TIMEOUT_S: "7200"
|
||||
HCCL_CONNECT_TIMEOUT: "1200"
|
||||
HCCL_INTRA_PCIE_ENABLE: "1"
|
||||
HCCL_INTRA_ROCE_ENABLE: "0"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"mtp"}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "65536"
|
||||
- --additional-config
|
||||
- '{"enable_dsa_cp": true}'
|
||||
- --max-num-batched-tokens
|
||||
- "4096"
|
||||
- --no-enable-prefix-caching
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "32121",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"mtp"}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "65536"
|
||||
- --additional-config
|
||||
- '{"enable_dsa_cp": true}'
|
||||
- --max-num-batched-tokens
|
||||
- "4096"
|
||||
- --no-enable-prefix-caching
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "32121",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method": "mtp"}'
|
||||
- "--seed"
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "150000"
|
||||
- --max-num-batched-tokens
|
||||
- "128"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --no-enable-prefix-caching
|
||||
- --quantization
|
||||
- ascend
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- glm47
|
||||
- --reasoning-parser
|
||||
- glm45
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "32123",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method": "mtp"}'
|
||||
- "--seed"
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "150000"
|
||||
- --max-num-batched-tokens
|
||||
- "128"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --no-enable-prefix-caching
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "32123",
|
||||
"engine_id": "3",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}}'
|
||||
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix0_in32768_bs100_glm
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 64
|
||||
max_out_len: 512
|
||||
batch_size: 16
|
||||
request_rate: 0
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,330 @@
|
||||
test_name: "GLM_5_1_PD-32k-0.5k-80-90"
|
||||
model: "Eco-Tech/GLM-5.1-w8a8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0, 1]
|
||||
decoder: [2, 3]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
HCCL_BUFFSIZE: "256"
|
||||
VLLM_TORCH_PROFILER_WITH_STACK: "0"
|
||||
ASCEND_AGGREGATE_ENABLE: "1"
|
||||
ASCEND_TRANSPORT_PRINT: "1"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
VLLM_NIXL_ABORT_REQUEST_TIMEOUT: "300000"
|
||||
VLLM_ENGINE_READY_TIMEOUT_S: "7200"
|
||||
HCCL_CONNECT_TIMEOUT: "1200"
|
||||
HCCL_INTRA_PCIE_ENABLE: "1"
|
||||
HCCL_INTRA_ROCE_ENABLE: "0"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"mtp"}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "65536"
|
||||
- --additional-config
|
||||
- '{"enable_dsa_cp": true}'
|
||||
- --max-num-batched-tokens
|
||||
- "4096"
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "32120",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"mtp"}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "65536"
|
||||
- --additional-config
|
||||
- '{"enable_dsa_cp": true}'
|
||||
- --max-num-batched-tokens
|
||||
- "4096"
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "32120",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method": "mtp"}'
|
||||
- "--seed"
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "150000"
|
||||
- --max-num-batched-tokens
|
||||
- "128"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "32122",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method": "mtp"}'
|
||||
- "--seed"
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "150000"
|
||||
- --max-num-batched-tokens
|
||||
- "128"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "32122",
|
||||
"engine_id": "3",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}}'
|
||||
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix90_in32768_bs100_glm
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 80
|
||||
max_out_len: 512
|
||||
batch_size: 20
|
||||
request_rate: 0
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,344 @@
|
||||
#131072-1024-90-32
|
||||
#65536-1024-90-64
|
||||
test_name: "GLM_5_1_PD-in64k_in128k_bs16_bs8"
|
||||
model: "Eco-Tech/GLM-5.1-w8a8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0, 1]
|
||||
decoder: [2, 3]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 2
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
HCCL_BUFFSIZE: "900"
|
||||
VLLM_TORCH_PROFILER_WITH_STACK: "0"
|
||||
ASCEND_AGGREGATE_ENABLE: "1"
|
||||
ASCEND_TRANSPORT_PRINT: "1"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
VLLM_NIXL_ABORT_REQUEST_TIMEOUT: "300000"
|
||||
VLLM_ENGINE_READY_TIMEOUT_S: "7200"
|
||||
HCCL_CONNECT_TIMEOUT: "1200"
|
||||
HCCL_INTRA_PCIE_ENABLE: "1"
|
||||
HCCL_INTRA_ROCE_ENABLE: "0"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"mtp"}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "140000"
|
||||
- --additional-config
|
||||
- '{"enable_dsa_cp": true}'
|
||||
- --max-num-batched-tokens
|
||||
- "4096"
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "32120",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"mtp"}'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "140000"
|
||||
- --additional-config
|
||||
- '{"enable_dsa_cp": true}'
|
||||
- --max-num-batched-tokens
|
||||
- "4096"
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "64"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --enforce-eager
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "32120",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"mtp"}'
|
||||
- "--seed"
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "140000"
|
||||
- --max-num-batched-tokens
|
||||
- "128"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,64,96,128]}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "32122",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
}
|
||||
}}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --speculative-config
|
||||
- '{"num_speculative_tokens": 3, "method":"mtp"}'
|
||||
- "--seed"
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "140000"
|
||||
- --max-num-batched-tokens
|
||||
- "128"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,64,96,128]}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true}'
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --async-scheduling
|
||||
- --enable-auto-tool-choice
|
||||
- --tool-call-parser
|
||||
- "glm47"
|
||||
- --reasoning-parser
|
||||
- "glm45"
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "32122",
|
||||
"engine_id": "3",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 8
|
||||
}
|
||||
}}'
|
||||
|
||||
benchmarks:
|
||||
perf_1:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix90_in65536_bs100_glm
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 64
|
||||
max_out_len: 1024
|
||||
batch_size: 16
|
||||
request_rate: 0
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
perf_2:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix90_in131072_bs100_glm
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 32
|
||||
max_out_len: 1024
|
||||
batch_size: 8
|
||||
request_rate: 0
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
|
||||
@@ -0,0 +1,269 @@
|
||||
test_name: "multi-node-kimi-k2.5-w4a8-ep-128k-1k-TPOT50"
|
||||
model: "Eco-Tech/Kimi-K2.5-W4A8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
SERVER_PORT: "${PORT}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1536"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
HCCL_BUFFSIZE: "800"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "1"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --no-enable-prefix-caching
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --additional-config
|
||||
- '{"enable_shared_expert_dp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"use_ascend_direct": true, "prefill": {"dp_size": 4, "tp_size": 4}, "decode": {"dp_size": 8, "tp_size": 4}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "1"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --no-enable-prefix-caching
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --additional-config
|
||||
- '{"enable_shared_expert_dp": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30100", "engine_id": "1", "kv_connector_extra_config": {"use_ascend_direct": true, "prefill": {"dp_size": 4, "tp_size": 4}, "decode": {"dp_size": 8, "tp_size": 4}}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133720"
|
||||
- --max-num-batched-tokens
|
||||
- "32"
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.91"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3", "model": "lightseekorg/kimi-k2.5-eagle3", "num_speculative_tokens": 3}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"use_ascend_direct": true, "prefill": {"dp_size": 4, "tp_size": 4}, "decode": {"dp_size": 8, "tp_size": 4}}}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133720"
|
||||
- --max-num-batched-tokens
|
||||
- "32"
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.91"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3", "model": "lightseekorg/kimi-k2.5-eagle3", "num_speculative_tokens": 3}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"use_ascend_direct": true, "prefill": {"dp_size": 4, "tp_size": 4}, "decode": {"dp_size": 8, "tp_size": 4}}}'
|
||||
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix90_in131072_bs1000_kimi
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
trust_remote_code: true
|
||||
num_prompts: 512
|
||||
max_out_len: 1024
|
||||
batch_size: 128
|
||||
request_rate: 3
|
||||
baseline: 1
|
||||
threshold: 0.95
|
||||
acc_aime2025:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/aime2025
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
|
||||
max_out_len: 65536
|
||||
temperature: 0.0
|
||||
top_p: 1
|
||||
top_k: -1
|
||||
repetition_penalty: 1.0
|
||||
batch_size: 32
|
||||
baseline: 95
|
||||
threshold: 10
|
||||
acc_GPQA:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/gpqa
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: gpqa/gpqa_gen_0_shot_cot_chat_prompt
|
||||
max_out_len: 65536
|
||||
batch_size: 32
|
||||
baseline: 87.6
|
||||
threshold: 5
|
||||
@@ -0,0 +1,246 @@
|
||||
test_name: "multi-node-kimi-k2.5-w4a8-ep-16k-1k-TPOT50"
|
||||
model: "Eco-Tech/Kimi-K2.5-W4A8"
|
||||
num_nodes: 4
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0, 1 ]
|
||||
decoder: [ 2, 3 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 0
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
- node_index: 3
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 16
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
SERVER_PORT: "${PORT}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1800"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --max-model-len
|
||||
- "32768"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --no-enable-prefix-caching
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --enforce-eager
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3", "model":"lightseekorg/kimi-k2.5-eagle3", "num_speculative_tokens": 3}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --max-model-len
|
||||
- "32768"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --no-enable-prefix-caching
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --enforce-eager
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3", "model":"lightseekorg/kimi-k2.5-eagle3", "num_speculative_tokens": 3}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30100", "engine_id": "1", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "48"
|
||||
- --max-model-len
|
||||
- "32768"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --no-enable-prefix-caching
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY", "cudagraph_capture_sizes":[4,8,16,32,48,64,80,96,112,128,144,160]}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"multistream_overlap_shared_expert": false}'
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3", "model":"lightseekorg/kimi-k2.5-eagle3", "num_speculative_tokens": 3}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 3
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "48"
|
||||
- --max-model-len
|
||||
- "32768"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --no-enable-prefix-caching
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY", "cudagraph_capture_sizes":[4,8,16,32,48,64,80,96,112,128,144,160]}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"multistream_overlap_shared_expert": false}'
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3", "model":"lightseekorg/kimi-k2.5-eagle3", "num_speculative_tokens": 3}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix0_in16384_bs1024_kimi25
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
trust_remote_code: true
|
||||
num_prompts: 1024
|
||||
max_out_len: 1024
|
||||
batch_size: 256
|
||||
request_rate: 3
|
||||
baseline: 2099.9
|
||||
threshold: 0.95
|
||||
@@ -0,0 +1,160 @@
|
||||
test_name: "Kimi-K2.6-W4A8-128k-1k-TPOT50-PD"
|
||||
model: "Eco-Tech/Kimi-K2.6-w4a8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0 ]
|
||||
decoder: [ 1 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
SERVER_PORT: "${PORT}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "120"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "512"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "800"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 1}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_producer","kv_port": "30000","engine_id": "0","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2,"tp_size": 8},"decode": {"dp_size": 4,"tp_size": 4}}}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 15}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true, "lmhead_tensor_parallel_size":16}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_consumer","kv_port": "30100","engine_id": "1","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2,"tp_size": 8},"decode": {"dp_size": 4,"tp_size": 4}}}'
|
||||
|
||||
benchmarks:
|
||||
acc:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/aime2025
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
|
||||
max_out_len: 32768
|
||||
batch_size: 8
|
||||
baseline: 93.33
|
||||
threshold: 10
|
||||
temperature: 1.0
|
||||
top_p: 1
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix90_in131072_bs1000_kimi
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 40
|
||||
max_out_len: 1024
|
||||
batch_size: 10
|
||||
request_rate: 0.3
|
||||
baseline: 220.503
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,150 @@
|
||||
test_name: "Kimi-K2.6-W4A8-254k-1k-TPOT20-PD"
|
||||
model: "Eco-Tech/Kimi-K2.6-w4a8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0 ]
|
||||
decoder: [ 1 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 1
|
||||
dp_size_local: 1
|
||||
dp_rank_start: 0
|
||||
tp_size: 16
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 1
|
||||
dp_size_local: 1
|
||||
dp_rank_start: 0
|
||||
tp_size: 16
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
SERVER_PORT: "${PORT}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "120"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
VLLM_SERVER_DEV_MODE: "1"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "512"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "800"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "262144"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "1"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 1}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_producer","kv_port": "30000","engine_id": "0","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 1,"tp_size": 16},"decode": {"dp_size": 1,"tp_size": 16}}}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "262144"
|
||||
- --max-num-batched-tokens
|
||||
- "64"
|
||||
- --max-num-seqs
|
||||
- "1"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 7}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true, "lmhead_tensor_parallel_size":16}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_consumer","kv_port": "30100","engine_id": "1","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 1,"tp_size": 16},"decode": {"dp_size": 1,"tp_size": 16}}}'
|
||||
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in254k-bs16-prefix99-kimi
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 1
|
||||
max_out_len: 1024
|
||||
batch_size: 1
|
||||
request_rate: 0
|
||||
baseline: 50.9973
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,152 @@
|
||||
test_name: "Kimi-K2.6-W4A8-3.5k-1.5k-TPOT50-PD"
|
||||
model: "Eco-Tech/Kimi-K2.6-w4a8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0 ]
|
||||
decoder: [ 1 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
SERVER_PORT: "${PORT}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "120"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
VLLM_SERVER_DEV_MODE: "1"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "512"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "800"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "20000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --no-enable-prefix-caching
|
||||
- --gpu-memory-utilization
|
||||
- "0.94"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 1}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_producer","kv_port": "30000","engine_id": "0","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2,"tp_size": 8},"decode": {"dp_size": 4,"tp_size": 4}}}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "20000"
|
||||
- --max-num-batched-tokens
|
||||
- "2048"
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --trust-remote-code
|
||||
- --no-enable-prefix-caching
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 7}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true, "lmhead_tensor_parallel_size":16}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_consumer","kv_port": "30100","engine_id": "1","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2,"tp_size": 8},"decode": {"dp_size": 4,"tp_size": 4}}}'
|
||||
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs4096-prefix0-kimi
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 512
|
||||
max_out_len: 1500
|
||||
batch_size: 128
|
||||
request_rate: 2.5
|
||||
baseline: 3206.8964
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,160 @@
|
||||
test_name: "Kimi-K2.7-W4A8-128k-1k-TPOT50-PD"
|
||||
model: "Eco-Tech/Kimi-K2.7-Code-w4a8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0 ]
|
||||
decoder: [ 1 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
SERVER_PORT: "${PORT}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "120"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "512"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "800"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 1}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_producer","kv_port": "30000","engine_id": "0","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2,"tp_size": 8},"decode": {"dp_size": 4,"tp_size": 4}}}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "133120"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 15}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true, "lmhead_tensor_parallel_size":16}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_consumer","kv_port": "30100","engine_id": "1","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2,"tp_size": 8},"decode": {"dp_size": 4,"tp_size": 4}}}'
|
||||
|
||||
benchmarks:
|
||||
acc:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/aime2025
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
|
||||
max_out_len: 32768
|
||||
batch_size: 8
|
||||
baseline: 93.33
|
||||
threshold: 10
|
||||
temperature: 1.0
|
||||
top_p: 1
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix90_in131072_bs1000_kimi
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 40
|
||||
max_out_len: 1024
|
||||
batch_size: 10
|
||||
request_rate: 0.3
|
||||
baseline: 220.503
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,150 @@
|
||||
test_name: "Kimi-K2.7-W4A8-254k-1k-TPOT20-PD"
|
||||
model: "Eco-Tech/Kimi-K2.7-Code-w4a8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0 ]
|
||||
decoder: [ 1 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 1
|
||||
dp_size_local: 1
|
||||
dp_rank_start: 0
|
||||
tp_size: 16
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 1
|
||||
dp_size_local: 1
|
||||
dp_rank_start: 0
|
||||
tp_size: 16
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
SERVER_PORT: "${PORT}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "120"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
VLLM_SERVER_DEV_MODE: "1"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "512"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "800"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "262144"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "1"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 1}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_producer","kv_port": "30000","engine_id": "0","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 1,"tp_size": 16},"decode": {"dp_size": 1,"tp_size": 16}}}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "262144"
|
||||
- --max-num-batched-tokens
|
||||
- "64"
|
||||
- --max-num-seqs
|
||||
- "1"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 7}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true, "lmhead_tensor_parallel_size":16}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_consumer","kv_port": "30100","engine_id": "1","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 1,"tp_size": 16},"decode": {"dp_size": 1,"tp_size": 16}}}'
|
||||
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in254k-bs16-prefix99-kimi
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 1
|
||||
max_out_len: 1024
|
||||
batch_size: 1
|
||||
request_rate: 0
|
||||
baseline: 50.9973
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,152 @@
|
||||
test_name: "Kimi-K2.7-W4A8-3.5k-1.5k-TPOT50-PD"
|
||||
model: "Eco-Tech/Kimi-K2.7-Code-w4a8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0 ]
|
||||
decoder: [ 1 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
SERVER_PORT: "${PORT}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "120"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
VLLM_SERVER_DEV_MODE: "1"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "512"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "800"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "20000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --no-enable-prefix-caching
|
||||
- --gpu-memory-utilization
|
||||
- "0.94"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 1}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_producer","kv_port": "30000","engine_id": "0","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2,"tp_size": 8},"decode": {"dp_size": 4,"tp_size": 4}}}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "20000"
|
||||
- --max-num-batched-tokens
|
||||
- "2048"
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --trust-remote-code
|
||||
- --no-enable-prefix-caching
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 7}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true, "lmhead_tensor_parallel_size":16}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_consumer","kv_port": "30100","engine_id": "1","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2,"tp_size": 8},"decode": {"dp_size": 4,"tp_size": 4}}}'
|
||||
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs4096-prefix0-kimi
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 512
|
||||
max_out_len: 1500
|
||||
batch_size: 128
|
||||
request_rate: 2.5
|
||||
baseline: 3206.8964
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,150 @@
|
||||
test_name: "Kimi-K2.7-W4A8-64k-1k-TPOT50-PD"
|
||||
model: "Eco-Tech/Kimi-K2.7-Code-w4a8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0 ]
|
||||
decoder: [ 1 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
SERVER_PORT: "${PORT}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "120"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
VLLM_SERVER_DEV_MODE: "1"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "512"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "800"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "68000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.94"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 1}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_producer","kv_port": "30000","engine_id": "0","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2,"tp_size": 8},"decode": {"dp_size": 4,"tp_size": 4}}}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --allowed-local-media-path
|
||||
- "/"
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --safetensors-load-strategy
|
||||
- 'prefetch'
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "68000"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.92"
|
||||
- --speculative-config
|
||||
- '{"method": "dflash","model": "z-lab/Kimi-K2.5-DFlash", "num_speculative_tokens": 15}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true, "lmhead_tensor_parallel_size":16}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1","kv_role": "kv_consumer","kv_port": "30100","engine_id": "1","kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2,"tp_size": 8},"decode": {"dp_size": 4,"tp_size": 4}}}'
|
||||
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix90_in65536_bs1000_kimi
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 60
|
||||
max_out_len: 1024
|
||||
batch_size: 15
|
||||
request_rate: 0.4
|
||||
baseline: 347.4475
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,221 @@
|
||||
test_name: "MiniMax-PD-in32k-bs4-1"
|
||||
model: "Eco-Tech/MiniMax-M2.5-w8a8-QuaRot"
|
||||
num_nodes: 3
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0]
|
||||
decoder: [1, 2]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 4
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
OMP_NUM_THREADS: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
HCCL_BUFFSIZE: "1200"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- ${PORT}
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --max-model-len
|
||||
- "40960"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enforce-eager
|
||||
- --no-enable-prefix_caching
|
||||
- --speculative-config
|
||||
- '{"method":"eagle3","model":"vllm-ascend/MiniMax-M2.5-eagle-model-0318","num_speculative_tokens":3}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": 30000,
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 4
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 4
|
||||
}
|
||||
}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "0"
|
||||
HCCL_BUFFSIZE: "512"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- ${PORT}
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --max-model-len
|
||||
- "40960"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix_caching
|
||||
- --speculative-config
|
||||
- '{"method":"eagle3","model":"vllm-ascend/MiniMax-M2.5-eagle-model-0318","num_speculative_tokens":3}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": 30100,
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 4
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 4
|
||||
}
|
||||
}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "0"
|
||||
HCCL_BUFFSIZE: "512"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- ${PORT}
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --max-model-len
|
||||
- "40960"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix_caching
|
||||
- --speculative-config
|
||||
- '{"method":"eagle3","model":"vllm-ascend/MiniMax-M2.5-eagle-model-0318","num_speculative_tokens":3}'
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": 30200,
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 4
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 4
|
||||
}
|
||||
}}'
|
||||
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix0_in32768_bs1000_Minimax
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 4
|
||||
max_out_len: 512
|
||||
batch_size: 1
|
||||
request_rate: 1
|
||||
baseline: 28.77
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,235 @@
|
||||
test_name: "Minimax_m2.7_in3_5_tpot50"
|
||||
model: "vllm-ascend/MiniMax-M2.7-w8a8-QuaRot"
|
||||
num_nodes: 3
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0 ]
|
||||
decoder: [ 1, 2 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
OMP_NUM_THREADS: "1"
|
||||
LD_PRELOAD: "/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD"
|
||||
LD_LIBRARY_PATH: "/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
PYTHONHASHSEED: "0"
|
||||
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1024"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "2048"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "0"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "300"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --no-enable-prefix-caching
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3","model":"Eco-Tech/MiniMax-M2.7-eagle-model-short","num_speculative_tokens":3}'
|
||||
- --enforce-eager
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_cpu_binding": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "55880",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}} }'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "75"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --no-enable-prefix-caching
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3","model":"Eco-Tech/MiniMax-M2.7-eagle-model-short","num_speculative_tokens":3}'
|
||||
- --async-scheduling
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_cpu_binding": true, "ascend_compilation_config":{"enable_npugraph_ex": true, "enable_static_kernel": false}}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "56900",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 4
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}}'
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "75"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --no-enable-prefix-caching
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3","model":"Eco-Tech/MiniMax-M2.7-eagle-model-short","num_speculative_tokens":3}'
|
||||
- --async-scheduling
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_cpu_binding": true, "ascend_compilation_config":{"enable_npugraph_ex": true, "enable_static_kernel": false}}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "56900",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 4,
|
||||
"tp_size": 4
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
}
|
||||
}}'
|
||||
benchmarks:
|
||||
perf_1:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_in3500_bs4000_minimax
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 800
|
||||
max_out_len: 1500
|
||||
batch_size: 200
|
||||
request_rate: 0
|
||||
baseline: 7603.536
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,236 @@
|
||||
test_name: "Minimax_m2.7_in3_5_tpot50"
|
||||
model: "vllm-ascend/MiniMax-M2.7-w8a8-QuaRot"
|
||||
num_nodes: 3
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0 ]
|
||||
decoder: [ 1, 2 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 16
|
||||
dp_size_local: 8
|
||||
dp_rank_start: 0
|
||||
tp_size: 2
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 16
|
||||
dp_size_local: 8
|
||||
dp_rank_start: 8
|
||||
tp_size: 2
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
OMP_NUM_THREADS: "1"
|
||||
LD_PRELOAD: "/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD"
|
||||
LD_LIBRARY_PATH: "/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
PYTHONHASHSEED: "0"
|
||||
|
||||
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1024"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "2048"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "0"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "300"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --no-enable-prefix-caching
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3","model":"Eco-Tech/MiniMax-M2.7-eagle-model-short","num_speculative_tokens":3}'
|
||||
- --enforce-eager
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_cpu_binding": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "55880",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 16,
|
||||
"tp_size": 2
|
||||
}} }'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "75"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --no-enable-prefix-caching
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3","model":"Eco-Tech/MiniMax-M2.7-eagle-model-short","num_speculative_tokens":3}'
|
||||
- --async-scheduling
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_cpu_binding": true, "ascend_compilation_config":{"enable_npugraph_ex": true, "enable_static_kernel": false}}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "56900",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 16,
|
||||
"tp_size": 2
|
||||
}
|
||||
}}'
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --max-model-len
|
||||
- "40000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "75"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --no-enable-prefix-caching
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3","model":"Eco-Tech/MiniMax-M2.7-eagle-model-short","num_speculative_tokens":3}'
|
||||
- --async-scheduling
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_cpu_binding": true, "ascend_compilation_config":{"enable_npugraph_ex": true, "enable_static_kernel": false}}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "56900",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 16,
|
||||
"tp_size": 2
|
||||
}
|
||||
}}'
|
||||
benchmarks:
|
||||
perf_1:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_in3500_bs4000_minimax
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 3440
|
||||
max_out_len: 1500
|
||||
batch_size: 860
|
||||
request_rate: 0
|
||||
baseline: 15006.5581
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,173 @@
|
||||
test_name: "Minimax_m2.7_in3_5_tpot50"
|
||||
model: "vllm-ascend/MiniMax-M2.7-w8a8-QuaRot"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0 ]
|
||||
decoder: [ 1 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
OMP_NUM_THREADS: "1"
|
||||
LD_PRELOAD: "/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD"
|
||||
LD_LIBRARY_PATH: "/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
PYTHONHASHSEED: "0"
|
||||
|
||||
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "1024"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "2048"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "0"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "48"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3","model":"Eco-Tech/MiniMax-M2.7-eagle-model-short","num_speculative_tokens":3}'
|
||||
- --enforce-eager
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_cpu_binding": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "55880",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
}} }'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "48"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --speculative-config
|
||||
- '{"method": "eagle3","model":"Eco-Tech/MiniMax-M2.7-eagle-model-short","num_speculative_tokens":3}'
|
||||
- --async-scheduling
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable":true,"enable_cpu_binding": true, "ascend_compilation_config":{"enable_npugraph_ex": true, "enable_static_kernel": false}}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "56900",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"use_ascend_direct": true,
|
||||
"prefill": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
}
|
||||
}}'
|
||||
|
||||
benchmarks:
|
||||
perf_1:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix90_in65536_bs1000_Minimax
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 72
|
||||
max_out_len: 1024
|
||||
batch_size: 18
|
||||
request_rate: 0
|
||||
baseline: 284.9
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,210 @@
|
||||
test_name: "QWEN3_235B_PD"
|
||||
model: "vllm-ascend/Qwen3-235B-A22B-W8A8"
|
||||
num_nodes: 3
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0 ]
|
||||
decoder: [ 1, 2 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 1
|
||||
dp_size_local: 1
|
||||
dp_rank_start: 0
|
||||
tp_size: 16
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 0
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 16
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "400"
|
||||
HCCL_ENTRY_LOG_ENABLE: 1
|
||||
LD_PRELOAD: "/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "10"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_USE_V1: "1"
|
||||
LD_LIBRARY_PATH: "/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_TRANSPORT_PRINT: "0"
|
||||
ASCEND_AGGREGATE_ENABLE: "1"
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
ASCEND_RT_VISIBLE_DEVICES: "0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15"
|
||||
HCCL_BUFFSIZE: "256"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
DISABLE_L2_CACHE: "1"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_TORCH_PROFILER_WITH_STACK: 0
|
||||
EXPERT_MAP_RECORD: "true"
|
||||
DYNAMIC_EPLB: "true"
|
||||
HCCL_BUFFSIZE: "1024"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- "7100"
|
||||
- --data-parallel-size
|
||||
- "1"
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --data-parallel-address
|
||||
- "12321"
|
||||
- --tensor-parallel-size
|
||||
- "16"
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "36864"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "8"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --additional-config
|
||||
- '{"ascend_scheduler_config":{"enabled":true,"enable_chunked_prefill":true}}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 1, "tp_size": 16}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "36864"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "28"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY","cudagraph_capture_sizes":[6,12,18]}'
|
||||
- --additional-config
|
||||
- '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"dynamic_eplb":true,"num_iterations_eplb_update":4000,"num_wait_worker_iterations":200,"init_redundancy_expert":0,"gate_eplb":false}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "1", "kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 1, "tp_size": 16}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "36864"
|
||||
- --max-num-batched-tokens
|
||||
- "256"
|
||||
- --max-num-seqs
|
||||
- "28"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY","cudagraph_capture_sizes":[6,12,18]}'
|
||||
- --additional-config
|
||||
- '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"dynamic_eplb":true,"num_iterations_eplb_update":4000,"num_wait_worker_iterations":200,"init_redundancy_expert":0,"gate_eplb":false}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "1", "kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 1, "tp_size": 16}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
benchmarks:
|
||||
acc:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/gsm8k
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_chat_prompt
|
||||
max_out_len: 32768
|
||||
batch_size: 64
|
||||
baseline: 96.74
|
||||
threshold: 1
|
||||
temperature: 0.6
|
||||
top_p: 0.95
|
||||
ignore_eos: false
|
||||
thinking: true
|
||||
|
||||
@@ -0,0 +1,206 @@
|
||||
test_name: "QWEN3_235B_PD_3_5K_1_5k"
|
||||
model: "vllm-ascend/Qwen3-235B-A22B-W8A8"
|
||||
num_nodes: 3
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [ 0 ]
|
||||
decoder: [ 1, 2 ]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 0
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 16
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_RPC_TIMEOUT: "3600000"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "204"
|
||||
HCCL_CONNECT_TIMEOUT: "400"
|
||||
HCCL_ENTRY_LOG_ENABLE: 1
|
||||
LD_PRELOAD: "/usr/lib/aarch64-linux-gnu/libjemalloc.so.2:$LD_PRELOAD"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "10"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
VLLM_USE_V1: "1"
|
||||
LD_LIBRARY_PATH: "/usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake:$LD_LIBRARY_PATH"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_TRANSPORT_PRINT: "0"
|
||||
ASCEND_AGGREGATE_ENABLE: "1"
|
||||
|
||||
env_prefill: &env_prefill
|
||||
<<: *env_common
|
||||
HCCL_BUFFSIZE: "256"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
|
||||
env_decode: &env_decode
|
||||
<<: *env_common
|
||||
VLLM_TORCH_PROFILER_WITH_STACK: 0
|
||||
EXPERT_MAP_RECORD: "true"
|
||||
DYNAMIC_EPLB: "true"
|
||||
HCCL_BUFFSIZE: "1024"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_prefill
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "6000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --enforce-eager
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --additional-config
|
||||
- '{"ascend_scheduler_config": {"enabled": true,"enable_chunked_prefill":true}}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_producer", "kv_port": "30000", "engine_id": "0", "kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "6000"
|
||||
- --max-num-batched-tokens
|
||||
- "4096"
|
||||
- --max-num-seqs
|
||||
- "15"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY","cudagraph_capture_sizes":[5,10,15]}'
|
||||
- --additional-config
|
||||
- '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"dynamic_eplb":true,"num_iterations_eplb_update":4000,"num_wait_worker_iterations":200,"init_redundancy_expert":0,"gate_eplb":false}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_decode
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "6000"
|
||||
- --max-num-batched-tokens
|
||||
- "4096"
|
||||
- --max-num-seqs
|
||||
- "15"
|
||||
- --trust-remote-code
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-enable-prefix-caching
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY","cudagraph_capture_sizes":[5,10,15]}'
|
||||
- --additional-config
|
||||
- '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"dynamic_eplb":true,"num_iterations_eplb_update":4000,"num_wait_worker_iterations":200,"init_redundancy_expert":0,"gate_eplb":false}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
|
||||
|
||||
benchmarks:
|
||||
perf_1:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs8000-qwen3
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 760
|
||||
max_out_len: 1500
|
||||
batch_size: 190
|
||||
request_rate: 0
|
||||
baseline: 3583.8747
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,254 @@
|
||||
test_name: "Qwen3.5-397B-w8a8-PD"
|
||||
model: "Eco-Tech/Qwen3.5-397B-A17B-w8a8-mtp"
|
||||
num_nodes: 3
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0]
|
||||
decoder: [1, 2]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 2
|
||||
dp_size_local: 2
|
||||
dp_rank_start: 0
|
||||
tp_size: 8
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 0
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 12321
|
||||
dp_size: 32
|
||||
dp_size_local: 16
|
||||
dp_rank_start: 16
|
||||
tp_size: 1
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
VLLM_ENGINE_READY_TIMEOUT_S: "7200"
|
||||
VLLM_RPC_TIMEOUT: "3600"
|
||||
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "30000"
|
||||
HCCL_EXEC_TIMEOUT: "2000"
|
||||
HCCL_CONNECT_TIMEOUT: "1200"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_NIXL_ABORT_REQUEST_TIMEOUT: "300000"
|
||||
HCCL_BUFFSIZE: "1536"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
VLLM_TORCH_PROFILER_WITH_STACK: "0"
|
||||
VLLM_USE_V1: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --api-server-count
|
||||
- "1"
|
||||
- --distributed-executor-backend
|
||||
- "mp"
|
||||
- --no-enable-prefix-caching
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "67000"
|
||||
- --max-num-batched-tokens
|
||||
- "16384"
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --enforce-eager
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --speculative-config
|
||||
- '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30000",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 32,
|
||||
"tp_size": 1
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --distributed-executor-backend
|
||||
- "mp"
|
||||
- --no-enable-prefix-caching
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "34816"
|
||||
- --max-num-batched-tokens
|
||||
- "128"
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY","cudagraph_capture_sizes":[4,8,16,24,32,40,48,56,64]}'
|
||||
- --speculative-config
|
||||
- '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true, "enable_cpu_binding": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30200",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 32,
|
||||
"tp_size": 1
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --distributed-executor-backend
|
||||
- "mp"
|
||||
- --no-enable-prefix-caching
|
||||
- --enable-expert-parallel
|
||||
- --seed
|
||||
- "1024"
|
||||
- --max-model-len
|
||||
- "34816"
|
||||
- --max-num-batched-tokens
|
||||
- "128"
|
||||
- --trust-remote-code
|
||||
- --max-num-seqs
|
||||
- "16"
|
||||
- --gpu-memory-utilization
|
||||
- "0.95"
|
||||
- --quantization
|
||||
- "ascend"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --compilation-config
|
||||
- '{"cudagraph_mode": "FULL_DECODE_ONLY","cudagraph_capture_sizes":[4,8,16,24,32,40,48,56,64]}'
|
||||
- --speculative-config
|
||||
- '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true, "enable_cpu_binding": true}'
|
||||
- --kv-transfer-config
|
||||
- '{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30200",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 32,
|
||||
"tp_size": 1
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
|
||||
benchmarks:
|
||||
acc_aime2025:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/aime2025
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
|
||||
max_out_len: 32768
|
||||
batch_size: 32
|
||||
baseline: 90
|
||||
threshold: 10
|
||||
temperature: 0.6
|
||||
top_p: 0.95
|
||||
top_k: 20
|
||||
min_p: 0.0
|
||||
presence_penalty: 0.0
|
||||
repetition_penalty: 1.0
|
||||
@@ -0,0 +1,271 @@
|
||||
test_name: "Qwen3.5-397B-w4a8_128k-1k-90-50"
|
||||
model: "Eco-Tech/Qwen3.5-397B-A17B-w4a8-mtp"
|
||||
num_nodes: 3
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0, 1]
|
||||
decoder: [2]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 8
|
||||
dp_rank_start: 0
|
||||
tp_size: 2
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_BUFFSIZE: "1024"
|
||||
VLLM_USE_V1: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "0"
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --max-num-seqs
|
||||
- "24"
|
||||
- --max-model-len
|
||||
- "140000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --seed
|
||||
- "1024"
|
||||
- --enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --enforce-eager
|
||||
- --speculative-config
|
||||
- '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3, "enforce_eager": true}'
|
||||
- --additional-config
|
||||
- '{"enable_cpu_binding": true}'
|
||||
- --profiler-config
|
||||
- '{"profiler": "torch", "torch_profiler_dir": "./profiling", "torch_profiler_with_stack": false}'
|
||||
- --mm-processor-cache-gb
|
||||
- "0"
|
||||
- --mm-encoder-tp-mode
|
||||
- "data"
|
||||
- --distributed-executor-backend
|
||||
- "mp"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --kv-transfer-config
|
||||
- '{
|
||||
"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_buffer_device": "npu",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "55010",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --max-num-seqs
|
||||
- "24"
|
||||
- --max-model-len
|
||||
- "140000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --seed
|
||||
- "1024"
|
||||
- --enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --enforce-eager
|
||||
- --speculative-config
|
||||
- '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3, "enforce_eager": true}'
|
||||
- --additional-config
|
||||
- '{"enable_cpu_binding": true}'
|
||||
- --profiler-config
|
||||
- '{"profiler": "torch", "torch_profiler_dir": "./profiling", "torch_profiler_with_stack": false}'
|
||||
- --mm-processor-cache-gb
|
||||
- "0"
|
||||
- --mm-encoder-tp-mode
|
||||
- "data"
|
||||
- --distributed-executor-backend
|
||||
- "mp"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --kv-transfer-config
|
||||
- '{
|
||||
"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_buffer_device": "npu",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "55010",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_SKIP_LOW_BLOCKS: "0"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --max-num-seqs
|
||||
- "24"
|
||||
- --max-model-len
|
||||
- "140000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --seed
|
||||
- "1024"
|
||||
- --enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --compilation-config
|
||||
- '{"cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,68,72,76,80,84,88,92,96], "cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true, "enable_cpu_binding": true}'
|
||||
- --profiler-config
|
||||
- '{"profiler": "torch", "torch_profiler_dir": "./profiling", "torch_profiler_with_stack": false}'
|
||||
- --mm-processor-cache-gb
|
||||
- "0"
|
||||
- --mm-encoder-tp-mode
|
||||
- "data"
|
||||
- --distributed-executor-backend
|
||||
- "mp"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --kv-transfer-config
|
||||
- '{
|
||||
"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_buffer_device": "npu",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "55011",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
benchmarks:
|
||||
acc_aime2025:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/aime2025
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
|
||||
max_out_len: 32768
|
||||
batch_size: 32
|
||||
baseline: 90
|
||||
threshold: 10
|
||||
temperature: 0.6
|
||||
top_p: 0.95
|
||||
top_k: 20
|
||||
min_p: 0.0
|
||||
presence_penalty: 0.0
|
||||
repetition_penalty: 1.0
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path_local: vllm-ascend/GSM8K_prefix90_in131072_bs100_qwen
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 256
|
||||
max_out_len: 1024
|
||||
batch_size: 64
|
||||
request_rate: 0
|
||||
baseline: 657.0699
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,195 @@
|
||||
test_name: "Qwen3.5-397B-w4a8_3.5k-1.5k-0-50"
|
||||
model: "Eco-Tech/Qwen3.5-397B-A17B-w4a8-mtp"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0]
|
||||
decoder: [1]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 8
|
||||
dp_rank_start: 0
|
||||
tp_size: 2
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 8
|
||||
dp_rank_start: 0
|
||||
tp_size: 2
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_BUFFSIZE: "1024"
|
||||
VLLM_USE_V1: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --max-num-seqs
|
||||
- "24"
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --seed
|
||||
- "1024"
|
||||
- --no-enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --enforce-eager
|
||||
- --speculative-config
|
||||
- '{"method": "qwen3_5_mtp", "num_speculative_tokens": 1, "enforce_eager": true}'
|
||||
- --additional-config
|
||||
- '{"enable_cpu_binding": true}'
|
||||
- --profiler-config
|
||||
- '{"profiler": "torch", "torch_profiler_dir": "./profiling", "torch_profiler_with_stack": false}'
|
||||
- --mm-processor-cache-gb
|
||||
- "0"
|
||||
- --mm-encoder-tp-mode
|
||||
- "data"
|
||||
- --distributed-executor-backend
|
||||
- "mp"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --kv-transfer-config
|
||||
- '{
|
||||
"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_buffer_device": "npu",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "55010",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_SKIP_LOW_BLOCKS: "0"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --max-num-seqs
|
||||
- "24"
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --seed
|
||||
- "1024"
|
||||
- --no-enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --compilation-config
|
||||
- '{"cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,68,72,76,80,84,88,92,96], "cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true, "enable_cpu_binding": true}'
|
||||
- --profiler-config
|
||||
- '{"profiler": "torch", "torch_profiler_dir": "./profiling", "torch_profiler_with_stack": false}'
|
||||
- --mm-processor-cache-gb
|
||||
- "0"
|
||||
- --mm-encoder-tp-mode
|
||||
- "data"
|
||||
- --distributed-executor-backend
|
||||
- "mp"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --kv-transfer-config
|
||||
- '{
|
||||
"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_buffer_device": "npu",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "55010",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
benchmarks:
|
||||
perf_50:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs2800
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 512
|
||||
max_out_len: 1500
|
||||
batch_size: 128
|
||||
request_rate: 5
|
||||
baseline: 4372.606
|
||||
threshold: 0.97
|
||||
perf_20:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs2800
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 512
|
||||
max_out_len: 1500
|
||||
batch_size: 128
|
||||
request_rate: 5
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,256 @@
|
||||
test_name: "Qwen3.5-397B-w4a8_64k-1k-90-50"
|
||||
model: "Eco-Tech/Qwen3.5-397B-A17B-w4a8-mtp"
|
||||
num_nodes: 3
|
||||
npu_per_node: 16
|
||||
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0, 1]
|
||||
decoder: [2]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 0
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 4
|
||||
dp_rank_start: 4
|
||||
tp_size: 4
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
- node_index: 2
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 8
|
||||
dp_rank_start: 0
|
||||
tp_size: 2
|
||||
dp_address: "${NODE_2_IP}"
|
||||
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_BUFFSIZE: "1024"
|
||||
VLLM_USE_V1: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --seed
|
||||
- "1024"
|
||||
- --enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --enforce-eager
|
||||
- --speculative-config
|
||||
- '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3, "enforce_eager": true}'
|
||||
- --additional-config
|
||||
- '{"enable_cpu_binding": true}'
|
||||
- --profiler-config
|
||||
- '{"profiler": "torch", "torch_profiler_dir": "./profiling", "torch_profiler_with_stack": false}'
|
||||
- --mm-processor-cache-gb
|
||||
- "0"
|
||||
- --mm-encoder-tp-mode
|
||||
- "data"
|
||||
- --distributed-executor-backend
|
||||
- "mp"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --kv-transfer-config
|
||||
- '{
|
||||
"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_buffer_device": "npu",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "55010",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
}
|
||||
}
|
||||
}'
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --gpu-memory-utilization
|
||||
- "0.8"
|
||||
- --seed
|
||||
- "1024"
|
||||
- --enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --enforce-eager
|
||||
- --speculative-config
|
||||
- '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3, "enforce_eager": true}'
|
||||
- --additional-config
|
||||
- '{"enable_cpu_binding": true}'
|
||||
- --profiler-config
|
||||
- '{"profiler": "torch", "torch_profiler_dir": "./profiling", "torch_profiler_with_stack": false}'
|
||||
- --mm-processor-cache-gb
|
||||
- "0"
|
||||
- --mm-encoder-tp-mode
|
||||
- "data"
|
||||
- --distributed-executor-backend
|
||||
- "mp"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --kv-transfer-config
|
||||
- '{
|
||||
"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_buffer_device": "npu",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "55010",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
- node_index: 2
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_SKIP_LOW_BLOCKS: "0"
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --max-num-seqs
|
||||
- "32"
|
||||
- --max-model-len
|
||||
- "70000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --seed
|
||||
- "1024"
|
||||
- --enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --compilation-config
|
||||
- '{"cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,68,72,76,80,84,88,92,96,100,104,108,112,116,120,124,128], "cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true, "enable_cpu_binding": true}'
|
||||
- --profiler-config
|
||||
- '{"profiler": "torch", "torch_profiler_dir": "./profiling", "torch_profiler_with_stack": false}'
|
||||
- --mm-processor-cache-gb
|
||||
- "0"
|
||||
- --mm-encoder-tp-mode
|
||||
- "data"
|
||||
- --distributed-executor-backend
|
||||
- "mp"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --kv-transfer-config
|
||||
- '{
|
||||
"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_buffer_device": "npu",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "55011",
|
||||
"engine_id": "2",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 4
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix90_in65536_bs1000_qwen
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 1000
|
||||
max_out_len: 1024
|
||||
batch_size: 250
|
||||
request_rate: 1
|
||||
baseline: 1041.18
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,188 @@
|
||||
#长序列只验功能，不做性能要求
|
||||
test_name: "Qwen3.5-397B-w4a8_986k-function"
|
||||
model: "Eco-Tech/Qwen3.5-397B-A17B-w4a8-mtp"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
|
||||
routing:
|
||||
type: "disaggregated_prefill"
|
||||
groups:
|
||||
prefiller: [0]
|
||||
decoder: [1]
|
||||
|
||||
config:
|
||||
- node_index: 0
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 8
|
||||
dp_rank_start: 0
|
||||
tp_size: 2
|
||||
dp_address: "${NODE_0_IP}"
|
||||
|
||||
|
||||
- node_index: 1
|
||||
port_start: 7100
|
||||
dp_rpc_port: 13321
|
||||
dp_size: 8
|
||||
dp_size_local: 8
|
||||
dp_rank_start: 0
|
||||
tp_size: 2
|
||||
dp_address: "${NODE_1_IP}"
|
||||
|
||||
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: "${PORT}"
|
||||
ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
OMP_PROC_BIND: "false"
|
||||
OMP_NUM_THREADS: "1"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
HCCL_BUFFSIZE: "1024"
|
||||
VLLM_USE_V1: "1"
|
||||
|
||||
templates:
|
||||
- node_index: 0
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --max-num-seqs
|
||||
- "24"
|
||||
- --max-model-len
|
||||
- "1010000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --seed
|
||||
- "1024"
|
||||
- --enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --enforce-eager
|
||||
- --speculative-config
|
||||
- '{"method": "qwen3_5_mtp", "num_speculative_tokens": 1, "enforce_eager": true}'
|
||||
- --additional-config
|
||||
- '{"enable_cpu_binding": true}'
|
||||
- --profiler-config
|
||||
- '{"profiler": "torch", "torch_profiler_dir": "./profiling", "torch_profiler_with_stack": false}'
|
||||
- --mm-processor-cache-gb
|
||||
- "0"
|
||||
- --mm-encoder-tp-mode
|
||||
- "data"
|
||||
- --distributed-executor-backend
|
||||
- "mp"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --hf-overrides
|
||||
- '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}'
|
||||
- --kv-transfer-config
|
||||
- '{
|
||||
"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_buffer_device": "npu",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "55010",
|
||||
"engine_id": "0",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
- node_index: 1
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd_template:
|
||||
- --host
|
||||
- "0.0.0.0"
|
||||
- --port
|
||||
- $SERVER_PORT
|
||||
- --data-parallel-size
|
||||
- ${DP_SIZE}
|
||||
- --data-parallel-rank
|
||||
- ${DP_RANK}
|
||||
- --data-parallel-address
|
||||
- ${DP_ADDRESS}
|
||||
- --data-parallel-rpc-port
|
||||
- ${DP_RPC_PORT}
|
||||
- --tensor-parallel-size
|
||||
- ${TP_SIZE}
|
||||
- --max-num-seqs
|
||||
- "24"
|
||||
- --max-model-len
|
||||
- "1010000"
|
||||
- --max-num-batched-tokens
|
||||
- "8192"
|
||||
- --gpu-memory-utilization
|
||||
- "0.9"
|
||||
- --seed
|
||||
- "1024"
|
||||
- --enable-prefix-caching
|
||||
- --async-scheduling
|
||||
- --compilation-config
|
||||
- '{"cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,68,72,76,80,84,88,92,96], "cudagraph_mode":"FULL_DECODE_ONLY"}'
|
||||
- --speculative-config
|
||||
- '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3}'
|
||||
- --additional-config
|
||||
- '{"recompute_scheduler_enable": true, "enable_cpu_binding": true}'
|
||||
- --profiler-config
|
||||
- '{"profiler": "torch", "torch_profiler_dir": "./profiling", "torch_profiler_with_stack": false}'
|
||||
- --mm-processor-cache-gb
|
||||
- "0"
|
||||
- --mm-encoder-tp-mode
|
||||
- "data"
|
||||
- --distributed-executor-backend
|
||||
- "mp"
|
||||
- --no-disable-hybrid-kv-cache-manager
|
||||
- --hf-overrides
|
||||
- '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}'
|
||||
- --kv-transfer-config
|
||||
- '{
|
||||
"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_buffer_device": "npu",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "55011",
|
||||
"engine_id": "1",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
}
|
||||
}
|
||||
}'
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs2800
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 4
|
||||
max_out_len: 1500
|
||||
batch_size: 2
|
||||
request_rate: 0
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
112
tests/e2e/weekly/multi_node/internal_dp/config/DeepSeek-V3.yaml
Normal file
112
tests/e2e/weekly/multi_node/internal_dp/config/DeepSeek-V3.yaml
Normal file
@@ -0,0 +1,112 @@
|
||||
# For disaggregated mode, set is_disaggregated: true, and set the following parameters:
|
||||
# Prefiller_index: the hosts index of the node running prefiller
|
||||
# Decoder_index: the hosts index of the node running decoder
|
||||
# Suppose we have **4 nodes** running a 2P1D setup (2 Prefillers + 1 Decoder):
|
||||
# ┌───────────────┬───────────────┬───────────────┬───────────────┐
|
||||
# │ node0 │ node1 │ node2 │ node3 │
|
||||
# │ Prefiller #1 │ Prefiller #2 │ Decoder │ Decoder │
|
||||
# └───────────────┴───────────────┴───────────────┴───────────────┘
|
||||
# For the prefiller nodes. the hosts should be node0 and node1
|
||||
# For the decoder nodes. we only have 1 decoder node(dp+tp+ep across node2 and node3. Where node3 is running with headless mode)
|
||||
# So the prefiller_host_index is [0, 1], and the decoder_host_index is [2]
|
||||
test_name: "test DeepSeek-V3 disaggregated_prefill"
|
||||
model: "vllm-ascend/DeepSeek-V3-W8A8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: true
|
||||
OMP_PROC_BIND: false
|
||||
OMP_NUM_THREADS: 1
|
||||
TASK_QUEUE_ENABLE: 1
|
||||
HCCL_BUFFSIZE: 1024
|
||||
SERVER_PORT: 8080
|
||||
NUMEXPR_MAX_THREADS: 128
|
||||
DISAGGREGATED_PREFILL_PROXY_SCRIPT: "examples/disaggregated_prefill_v1/load_balance_proxy_server_example.py"
|
||||
# For None kubernetes deployment, list the IPs of all nodes used in order as follow
|
||||
# cluster_hosts: [10.0.0.103, 10.0.0.139]
|
||||
disaggregated_prefill:
|
||||
enabled: true
|
||||
prefiller_host_index: [0]
|
||||
decoder_host_index: [1]
|
||||
|
||||
deployment:
|
||||
-
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve "vllm-ascend/DeepSeek-V3-W8A8"
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 2
|
||||
--data-parallel-size-local 2
|
||||
--tensor-parallel-size 8
|
||||
--seed 1024
|
||||
--enforce-eager
|
||||
--enable-expert-parallel
|
||||
--max-num-seqs 16
|
||||
--max-model-len 8192
|
||||
--max-num-batched-tokens 8192
|
||||
--quantization ascend
|
||||
--trust-remote-code
|
||||
--no-enable-prefix-caching
|
||||
--gpu-memory-utilization 0.9
|
||||
--kv-transfer-config
|
||||
'{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30000",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
-
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve "vllm-ascend/DeepSeek-V3-W8A8"
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 2
|
||||
--data-parallel-size-local 2
|
||||
--tensor-parallel-size 8
|
||||
--seed 1024
|
||||
--quantization ascend
|
||||
--max-num-seqs 16
|
||||
--max-model-len 8192
|
||||
--max-num-batched-tokens 8192
|
||||
--enable-expert-parallel
|
||||
--trust-remote-code
|
||||
--no-enable-prefix-caching
|
||||
--gpu-memory-utilization 0.9
|
||||
--kv-transfer-config
|
||||
'{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30200",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
}
|
||||
}
|
||||
}'
|
||||
benchmarks:
|
||||
acc:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/gsm8k-lite
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_chat_prompt
|
||||
max_out_len: 4096
|
||||
batch_size: 512
|
||||
baseline: 95
|
||||
threshold: 10
|
||||
@@ -0,0 +1,198 @@
|
||||
test_name: "weekly test DeepSeek-V3.2-W8A8-EP disaggregated_prefill"
|
||||
model: "vllm-ascend/DeepSeek-V3.2-W8A8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
env_common: &env_common
|
||||
OMP_PROC_BIND: false
|
||||
OMP_NUM_THREADS: 10
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_USE_V1: "1"
|
||||
HCCL_BUFFSIZE: "360"
|
||||
ASCEND_AGGREGATE_ENABLE: "1"
|
||||
ASCEND_TRANSPORT_PRINT: "1"
|
||||
VLLM_NIXL_ABORT_REQUEST_TIMEOUT: "300000"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
SERVER_PORT: 8080
|
||||
|
||||
disaggregated_prefill:
|
||||
enabled: true
|
||||
prefiller_host_index: [0]
|
||||
decoder_host_index: [1]
|
||||
|
||||
deployment:
|
||||
-
|
||||
envs:
|
||||
<<: *env_common
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: 1
|
||||
server_cmd: >
|
||||
vllm serve vllm-ascend/DeepSeek-V3.2-W8A8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 1
|
||||
--data-parallel-start-rank 0
|
||||
--data-parallel-size-local 1
|
||||
--data-parallel-address $LOCAL_IP
|
||||
--tensor-parallel-size 16
|
||||
--enable-expert-parallel
|
||||
--seed 1024
|
||||
--max-model-len 133000
|
||||
--max-num-batched-tokens 8192
|
||||
--trust-remote-code
|
||||
--max-num-seqs 16
|
||||
--gpu-memory-utilization 0.95
|
||||
--quantization ascend
|
||||
--enforce-eager
|
||||
--no-enable-prefix-caching
|
||||
--tokenizer-mode deepseek_v32
|
||||
--reasoning-parser deepseek_v3
|
||||
--speculative-config '{"num_speculative_tokens": 2, "method":"deepseek_mtp"}'
|
||||
--kv-transfer-config
|
||||
'{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30000",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 1,
|
||||
"tp_size": 16
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
-
|
||||
envs:
|
||||
<<: *env_common
|
||||
OMP_PROC_BIND: false
|
||||
OMP_NUM_THREADS: 10
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_USE_V1: "1"
|
||||
HCCL_BUFFSIZE: "1100"
|
||||
ASCEND_AGGREGATE_ENABLE: "1"
|
||||
ASCEND_TRANSPORT_PRINT: "1"
|
||||
ACL_OP_INIT_MODE: "1"
|
||||
ASCEND_A3_ENABLE: "1"
|
||||
VLLM_NIXL_ABORT_REQUEST_TIMEOUT: "300000"
|
||||
TASK_QUEUE_ENABLE: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
DYNAMIC_EPLB: "true"
|
||||
server_cmd: >
|
||||
vllm serve vllm-ascend/DeepSeek-V3.2-W8A8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 8
|
||||
--data-parallel-size-local 8
|
||||
--data-parallel-start-rank 0
|
||||
--data-parallel-address $LOCAL_IP
|
||||
--data-parallel-rpc-port 13389
|
||||
--tensor-parallel-size 2
|
||||
--enable-expert-parallel
|
||||
--seed 1024
|
||||
--max-model-len 133000
|
||||
--max-num-batched-tokens 100
|
||||
--speculative-config '{"num_speculative_tokens": 2, "method":"deepseek_mtp"}'
|
||||
--compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[3,6,9,12,15,18,21,24,27,30,33,36,39,42]}'
|
||||
--additional-config '{"recompute_scheduler_enable": true,"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":2000,"algorithm_execution_interval":50, "eplb_policy_type":3, "num_redundant_experts":0}}'
|
||||
--trust-remote-code
|
||||
--max-num-seqs 14
|
||||
--gpu-memory-utilization 0.92
|
||||
--no-enable-prefix-caching
|
||||
--quantization ascend
|
||||
--tokenizer-mode deepseek_v32
|
||||
--reasoning-parser deepseek_v3
|
||||
--kv-transfer-config
|
||||
'{"kv_connector": "MooncakeConnectorV1",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30200",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 1,
|
||||
"tp_size": 16
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 8,
|
||||
"tp_size": 2
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
|
||||
benchmarks:
|
||||
perf_16384_1024_num64:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix0_in16384_bs1200_deepseek
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 64
|
||||
max_out_len: 1024
|
||||
batch_size: 16
|
||||
request_rate: 1
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
|
||||
perf_16384_1024_num256:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix0_in16384_bs1200_deepseek
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 256
|
||||
max_out_len: 1024
|
||||
batch_size: 64
|
||||
request_rate: 1
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
|
||||
perf_32768_512_num32:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix0_in32768_bs1000_deepseek
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 32
|
||||
max_out_len: 512
|
||||
batch_size: 8
|
||||
request_rate: 1
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
|
||||
perf_32768_512_num128:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix0_in32768_bs1000_deepseek
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 128
|
||||
max_out_len: 512
|
||||
batch_size: 32
|
||||
request_rate: 0.7
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
|
||||
perf_65536_1024_num4:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix0_in65536_bs400_deepseek
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 4
|
||||
max_out_len: 1024
|
||||
batch_size: 1
|
||||
request_rate: 0.1
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
|
||||
perf_65536_1024_num16:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_prefix0_in65536_bs400_deepseek
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 16
|
||||
max_out_len: 1024
|
||||
batch_size: 4
|
||||
request_rate: 1
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,102 @@
|
||||
test_name: "test GLM-4.7-W8A8C8 PD separation with mooncake layerwise connector"
|
||||
model: "vllm-ascend/GLM-4.7-W8A8C8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
env_common: &env_common
|
||||
VLLM_USE_MODELSCOPE: true
|
||||
OMP_PROC_BIND: false
|
||||
HCCL_BUFFSIZE: 512
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
SERVER_PORT: 8080
|
||||
DISAGGREGATED_PREFILL_PROXY_SCRIPT: "examples/disaggregated_prefill_v1/load_balance_proxy_layerwise_server_example.py"
|
||||
|
||||
disaggregated_prefill:
|
||||
enabled: true
|
||||
prefiller_host_index: [0]
|
||||
decoder_host_index: [1]
|
||||
|
||||
deployment:
|
||||
-
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve "vllm-ascend/GLM-4.7-W8A8C8"
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 2
|
||||
--data-parallel-size-local 2
|
||||
--data-parallel-address $LOCAL_IP
|
||||
--data-parallel-rpc-port 13389
|
||||
--tensor-parallel-size 8
|
||||
--seed 1024
|
||||
--enforce-eager
|
||||
--enable-expert-parallel
|
||||
--max-num-seqs 32
|
||||
--max-model-len 8192
|
||||
--max-num-batched-tokens 8192
|
||||
--trust-remote-code
|
||||
--no-enable-prefix-caching
|
||||
--gpu-memory-utilization 0.9
|
||||
--additional-config '{"enable_shared_expert_dp": true}'
|
||||
--kv-transfer-config
|
||||
'{"kv_connector": "MooncakeLayerwiseConnector",
|
||||
"kv_role": "kv_producer",
|
||||
"kv_port": "30000",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
}
|
||||
}
|
||||
}'
|
||||
-
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve "vllm-ascend/GLM-4.7-W8A8C8"
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 2
|
||||
--data-parallel-size-local 2
|
||||
--data-parallel-address $MASTER_IP
|
||||
--data-parallel-rpc-port 13389
|
||||
--tensor-parallel-size 8
|
||||
--seed 1024
|
||||
--max-num-seqs 16
|
||||
--max-model-len 8192
|
||||
--max-num-batched-tokens 8192
|
||||
--enable-expert-parallel
|
||||
--trust-remote-code
|
||||
--no-enable-prefix-caching
|
||||
--gpu-memory-utilization 0.9
|
||||
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
--kv-transfer-config
|
||||
'{"kv_connector": "MooncakeLayerwiseConnector",
|
||||
"kv_role": "kv_consumer",
|
||||
"kv_port": "30200",
|
||||
"kv_connector_extra_config": {
|
||||
"prefill": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
},
|
||||
"decode": {
|
||||
"dp_size": 2,
|
||||
"tp_size": 8
|
||||
}
|
||||
}
|
||||
}'
|
||||
|
||||
benchmarks:
|
||||
acc:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/gsm8k-lite
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_chat_prompt
|
||||
max_out_len: 4096
|
||||
batch_size: 8
|
||||
baseline: 95
|
||||
threshold: 10
|
||||
@@ -0,0 +1,104 @@
|
||||
#198k长序列只验证功能
|
||||
test_name: "multi-node-GLM-5.1-W8A8C8-MTP-A3_198k_function"
|
||||
model: "Eco-Tech/GLM-5.1-w8a8c8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
env_common: &env_common
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
OMP_NUM_THREADS: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ENGINE_READY_TIMEOUT_S: "3000"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
SERVER_PORT: 8077
|
||||
|
||||
deployment:
|
||||
- envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve Eco-Tech/GLM-5.1-w8a8c8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 8
|
||||
--data-parallel-size-local 4
|
||||
--data-parallel-start-rank 0
|
||||
--data-parallel-address $LOCAL_IP
|
||||
--enable-expert-parallel
|
||||
--data-parallel-rpc-port 12980
|
||||
--tensor-parallel-size 4
|
||||
--hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
--seed 1024
|
||||
--tool-call-parser glm47
|
||||
--reasoning-parser glm45
|
||||
--enable-auto-tool-choice
|
||||
--max-num-seqs 6
|
||||
--max-model-len 202752
|
||||
--max-num-batched-tokens 4096
|
||||
--trust-remote-code
|
||||
--gpu-memory-utilization 0.92
|
||||
--quantization ascend
|
||||
--enable-chunked-prefill
|
||||
--enable-prefix-caching
|
||||
--async-scheduling
|
||||
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
--additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
|
||||
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
|
||||
-
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve Eco-Tech/GLM-5.1-w8a8c8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 8
|
||||
--data-parallel-size-local 4
|
||||
--data-parallel-start-rank 4
|
||||
--headless
|
||||
--data-parallel-address $MASTER_IP
|
||||
--enable-expert-parallel
|
||||
--data-parallel-rpc-port 12980
|
||||
--tensor-parallel-size 4
|
||||
--hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
--seed 1024
|
||||
--tool-call-parser glm47
|
||||
--reasoning-parser glm45
|
||||
--enable-auto-tool-choice
|
||||
--max-num-seqs 6
|
||||
--max-model-len 202752
|
||||
--max-num-batched-tokens 4096
|
||||
--trust-remote-code
|
||||
--gpu-memory-utilization 0.9
|
||||
--quantization ascend
|
||||
--enable-chunked-prefill
|
||||
--enable-prefix-caching
|
||||
--async-scheduling
|
||||
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
--additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
|
||||
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs2800
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 8
|
||||
max_out_len: 1500
|
||||
batch_size: 2
|
||||
request_rate: 0
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
acc_aime2025:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/aime2025
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
|
||||
max_out_len: 129024
|
||||
batch_size: 32
|
||||
baseline: 93.33
|
||||
threshold: 10
|
||||
temperature: 1.0
|
||||
top_p: 0.95
|
||||
@@ -0,0 +1,105 @@
|
||||
test_name: "multi-node-GLM-5.1-W8A8C8-MTP-A3_3.5k_1.5k_0_50"
|
||||
model: "Eco-Tech/GLM-5.1-w8a8c8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
env_common: &env_common
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
OMP_NUM_THREADS: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ENGINE_READY_TIMEOUT_S: "3000"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
SERVER_PORT: 8077
|
||||
|
||||
# TODO: need to identify why TP and mtp+1 divisibility rules break on dual-node case
|
||||
|
||||
deployment:
|
||||
- envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve Eco-Tech/GLM-5.1-w8a8c8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 8
|
||||
--data-parallel-size-local 4
|
||||
--data-parallel-start-rank 0
|
||||
--data-parallel-address $LOCAL_IP
|
||||
--enable-expert-parallel
|
||||
--data-parallel-rpc-port 12981
|
||||
--tensor-parallel-size 4
|
||||
--hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
--seed 1024
|
||||
--tool-call-parser glm47
|
||||
--reasoning-parser glm45
|
||||
--enable-auto-tool-choice
|
||||
--max-num-seqs 32
|
||||
--max-model-len 133120
|
||||
--max-num-batched-tokens 4096
|
||||
--trust-remote-code
|
||||
--gpu-memory-utilization 0.92
|
||||
--quantization ascend
|
||||
--enable-chunked-prefill
|
||||
--no-enable-prefix-caching
|
||||
--async-scheduling
|
||||
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
--additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
|
||||
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
|
||||
-
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve Eco-Tech/GLM-5.1-w8a8c8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 8
|
||||
--data-parallel-size-local 4
|
||||
--data-parallel-start-rank 4
|
||||
--headless
|
||||
--data-parallel-address $MASTER_IP
|
||||
--enable-expert-parallel
|
||||
--data-parallel-rpc-port 12981
|
||||
--tensor-parallel-size 4
|
||||
--hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
--seed 1024
|
||||
--tool-call-parser glm47
|
||||
--reasoning-parser glm45
|
||||
--enable-auto-tool-choice
|
||||
--max-num-seqs 32
|
||||
--max-model-len 133120
|
||||
--max-num-batched-tokens 4096
|
||||
--trust-remote-code
|
||||
--gpu-memory-utilization 0.95
|
||||
--quantization ascend
|
||||
--enable-chunked-prefill
|
||||
--no-enable-prefix-caching
|
||||
--async-scheduling
|
||||
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
--additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
|
||||
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
|
||||
benchmarks:
|
||||
perf_TPOT50:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_in3500_bs400
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 256
|
||||
max_out_len: 1500
|
||||
batch_size: 64
|
||||
request_rate: 0
|
||||
baseline: 1267.2573
|
||||
threshold: 0.97
|
||||
perf_TPOT20:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_in3500_bs400
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 32
|
||||
max_out_len: 1500
|
||||
batch_size: 8
|
||||
request_rate: 0
|
||||
baseline: 345.13
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,115 @@
|
||||
test_name: "multi-node-GLM-5.1-W8A8C8-MTP-A3_64k/128k"
|
||||
model: "Eco-Tech/GLM-5.1-w8a8c8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
env_common: &env_common
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
OMP_NUM_THREADS: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ENGINE_READY_TIMEOUT_S: "3000"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
SERVER_PORT: 8077
|
||||
|
||||
# TODO: need to identify why TP and mtp+1 divisibility rules break on dual-node case
|
||||
|
||||
deployment:
|
||||
- envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve Eco-Tech/GLM-5.1-w8a8c8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 8
|
||||
--data-parallel-size-local 4
|
||||
--data-parallel-start-rank 0
|
||||
--data-parallel-address $LOCAL_IP
|
||||
--enable-expert-parallel
|
||||
--data-parallel-rpc-port 12981
|
||||
--tensor-parallel-size 4
|
||||
--hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
--seed 1024
|
||||
--tool-call-parser glm47
|
||||
--reasoning-parser glm45
|
||||
--enable-auto-tool-choice
|
||||
--max-num-seqs 6
|
||||
--max-model-len 133120
|
||||
--max-num-batched-tokens 4096
|
||||
--trust-remote-code
|
||||
--gpu-memory-utilization 0.92
|
||||
--quantization ascend
|
||||
--enable-chunked-prefill
|
||||
--enable-prefix-caching
|
||||
--async-scheduling
|
||||
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
--additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
|
||||
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
|
||||
- envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve Eco-Tech/GLM-5.1-w8a8c8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 8
|
||||
--data-parallel-size-local 4
|
||||
--data-parallel-start-rank 4
|
||||
--headless
|
||||
--data-parallel-address $MASTER_IP
|
||||
--enable-expert-parallel
|
||||
--data-parallel-rpc-port 12981
|
||||
--tensor-parallel-size 4
|
||||
--hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
--seed 1024
|
||||
--tool-call-parser glm47
|
||||
--reasoning-parser glm45
|
||||
--enable-auto-tool-choice
|
||||
--max-num-seqs 6
|
||||
--max-model-len 133120
|
||||
--max-num-batched-tokens 4096
|
||||
--trust-remote-code
|
||||
--gpu-memory-utilization 0.9
|
||||
--quantization ascend
|
||||
--enable-chunked-prefill
|
||||
--enable-prefix-caching
|
||||
--async-scheduling
|
||||
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
--additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
|
||||
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
|
||||
benchmarks:
|
||||
perf_64k:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in65536-bs1500-prefix90-glm51
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 128
|
||||
max_out_len: 1024
|
||||
batch_size: 32
|
||||
request_rate: 0
|
||||
baseline: 539.11
|
||||
threshold: 0.97
|
||||
perf_128k_warmup:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in131072-bs500-prefix90-glm51
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 1
|
||||
max_out_len: 1
|
||||
batch_size: 1
|
||||
request_rate: 0
|
||||
baseline: 0
|
||||
threshold: 0.97
|
||||
perf_128k:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in131072-bs500-prefix90-glm51
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 256
|
||||
max_out_len: 1024
|
||||
batch_size: 64
|
||||
request_rate: 0
|
||||
baseline: 290.8308
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,104 @@
|
||||
#198k长序列只验证功能
|
||||
test_name: "multi-node-GLM-5.2-W4A8C8-MTP-A3_198k_function"
|
||||
model: "Eco-Tech/GLM-5.2-w4a8c8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
env_common: &env_common
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
OMP_NUM_THREADS: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ENGINE_READY_TIMEOUT_S: "3000"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
SERVER_PORT: 8077
|
||||
|
||||
deployment:
|
||||
- envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve Eco-Tech/GLM-5.2-w4a8c8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 8
|
||||
--data-parallel-size-local 4
|
||||
--data-parallel-start-rank 0
|
||||
--data-parallel-address $LOCAL_IP
|
||||
--enable-expert-parallel
|
||||
--data-parallel-rpc-port 12980
|
||||
--tensor-parallel-size 4
|
||||
--hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
--seed 1024
|
||||
--tool-call-parser glm47
|
||||
--reasoning-parser glm45
|
||||
--enable-auto-tool-choice
|
||||
--max-num-seqs 6
|
||||
--max-model-len 202752
|
||||
--max-num-batched-tokens 4096
|
||||
--trust-remote-code
|
||||
--gpu-memory-utilization 0.92
|
||||
--quantization ascend
|
||||
--enable-chunked-prefill
|
||||
--enable-prefix-caching
|
||||
--async-scheduling
|
||||
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
--additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
|
||||
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
|
||||
-
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve Eco-Tech/GLM-5.2-w4a8c8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 8
|
||||
--data-parallel-size-local 4
|
||||
--data-parallel-start-rank 4
|
||||
--headless
|
||||
--data-parallel-address $MASTER_IP
|
||||
--enable-expert-parallel
|
||||
--data-parallel-rpc-port 12980
|
||||
--tensor-parallel-size 4
|
||||
--hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
--seed 1024
|
||||
--tool-call-parser glm47
|
||||
--reasoning-parser glm45
|
||||
--enable-auto-tool-choice
|
||||
--max-num-seqs 6
|
||||
--max-model-len 202752
|
||||
--max-num-batched-tokens 4096
|
||||
--trust-remote-code
|
||||
--gpu-memory-utilization 0.9
|
||||
--quantization ascend
|
||||
--enable-chunked-prefill
|
||||
--enable-prefix-caching
|
||||
--async-scheduling
|
||||
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
--additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
|
||||
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
|
||||
benchmarks:
|
||||
perf:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in3500-bs2800
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 8
|
||||
max_out_len: 1500
|
||||
batch_size: 2
|
||||
request_rate: 0
|
||||
baseline: 1
|
||||
threshold: 0.97
|
||||
acc_aime2025:
|
||||
case_type: accuracy
|
||||
dataset_path: vllm-ascend/aime2025
|
||||
request_conf: vllm_api_general_chat
|
||||
dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
|
||||
max_out_len: 129024
|
||||
batch_size: 32
|
||||
baseline: 93.33
|
||||
threshold: 10
|
||||
temperature: 1.0
|
||||
top_p: 0.95
|
||||
@@ -0,0 +1,105 @@
|
||||
test_name: "multi-node-GLM-5.2-W4A8C8-MTP-A3_3.5k_1.5k_0_50"
|
||||
model: "Eco-Tech/GLM-5.2-w4a8c8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
env_common: &env_common
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
OMP_NUM_THREADS: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ENGINE_READY_TIMEOUT_S: "3000"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
SERVER_PORT: 8077
|
||||
|
||||
# TODO: need to identify why TP and mtp+1 divisibility rules break on dual-node case
|
||||
|
||||
deployment:
|
||||
- envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve Eco-Tech/GLM-5.2-w4a8c8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 8
|
||||
--data-parallel-size-local 4
|
||||
--data-parallel-start-rank 0
|
||||
--data-parallel-address $LOCAL_IP
|
||||
--enable-expert-parallel
|
||||
--data-parallel-rpc-port 12981
|
||||
--tensor-parallel-size 4
|
||||
--hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
--seed 1024
|
||||
--tool-call-parser glm47
|
||||
--reasoning-parser glm45
|
||||
--enable-auto-tool-choice
|
||||
--max-num-seqs 32
|
||||
--max-model-len 133120
|
||||
--max-num-batched-tokens 4096
|
||||
--trust-remote-code
|
||||
--gpu-memory-utilization 0.92
|
||||
--quantization ascend
|
||||
--enable-chunked-prefill
|
||||
--no-enable-prefix-caching
|
||||
--async-scheduling
|
||||
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
--additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
|
||||
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
|
||||
-
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve Eco-Tech/GLM-5.2-w4a8c8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 8
|
||||
--data-parallel-size-local 4
|
||||
--data-parallel-start-rank 4
|
||||
--headless
|
||||
--data-parallel-address $MASTER_IP
|
||||
--enable-expert-parallel
|
||||
--data-parallel-rpc-port 12981
|
||||
--tensor-parallel-size 4
|
||||
--hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
--seed 1024
|
||||
--tool-call-parser glm47
|
||||
--reasoning-parser glm45
|
||||
--enable-auto-tool-choice
|
||||
--max-num-seqs 32
|
||||
--max-model-len 133120
|
||||
--max-num-batched-tokens 4096
|
||||
--trust-remote-code
|
||||
--gpu-memory-utilization 0.95
|
||||
--quantization ascend
|
||||
--enable-chunked-prefill
|
||||
--no-enable-prefix-caching
|
||||
--async-scheduling
|
||||
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
--additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
|
||||
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
|
||||
benchmarks:
|
||||
perf_TPOT50:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_in3500_bs400
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 256
|
||||
max_out_len: 1500
|
||||
batch_size: 64
|
||||
request_rate: 0
|
||||
baseline: 1267.2573
|
||||
threshold: 0.97
|
||||
perf_TPOT20:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K_in3500_bs400
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 32
|
||||
max_out_len: 1500
|
||||
batch_size: 8
|
||||
request_rate: 0
|
||||
baseline: 345.13
|
||||
threshold: 0.97
|
||||
@@ -0,0 +1,105 @@
|
||||
test_name: "multi-node-GLM-5.2-W4A8C8-MTP-A3_64k/128k"
|
||||
model: "Eco-Tech/GLM-5.2-w4a8c8"
|
||||
num_nodes: 2
|
||||
npu_per_node: 16
|
||||
env_common: &env_common
|
||||
HCCL_OP_EXPANSION_MODE: "AIV"
|
||||
OMP_PROC_BIND: "false"
|
||||
VLLM_USE_MODELSCOPE: "true"
|
||||
OMP_NUM_THREADS: "1"
|
||||
HCCL_BUFFSIZE: "400"
|
||||
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
||||
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
||||
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
||||
VLLM_ENGINE_READY_TIMEOUT_S: "3000"
|
||||
VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
|
||||
SERVER_PORT: 8077
|
||||
|
||||
# TODO: need to identify why TP and mtp+1 divisibility rules break on dual-node case
|
||||
|
||||
deployment:
|
||||
- envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve Eco-Tech/GLM-5.2-w4a8c8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 8
|
||||
--data-parallel-size-local 4
|
||||
--data-parallel-start-rank 0
|
||||
--data-parallel-address $LOCAL_IP
|
||||
--enable-expert-parallel
|
||||
--data-parallel-rpc-port 12981
|
||||
--tensor-parallel-size 4
|
||||
--hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
--seed 1024
|
||||
--tool-call-parser glm47
|
||||
--reasoning-parser glm45
|
||||
--enable-auto-tool-choice
|
||||
--max-num-seqs 6
|
||||
--max-model-len 133120
|
||||
--max-num-batched-tokens 4096
|
||||
--trust-remote-code
|
||||
--gpu-memory-utilization 0.92
|
||||
--quantization ascend
|
||||
--enable-chunked-prefill
|
||||
--enable-prefix-caching
|
||||
--async-scheduling
|
||||
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
--additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
|
||||
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
|
||||
-
|
||||
envs:
|
||||
<<: *env_common
|
||||
server_cmd: >
|
||||
vllm serve Eco-Tech/GLM-5.2-w4a8c8
|
||||
--host 0.0.0.0
|
||||
--port $SERVER_PORT
|
||||
--data-parallel-size 8
|
||||
--data-parallel-size-local 4
|
||||
--data-parallel-start-rank 4
|
||||
--headless
|
||||
--data-parallel-address $MASTER_IP
|
||||
--enable-expert-parallel
|
||||
--data-parallel-rpc-port 12981
|
||||
--tensor-parallel-size 4
|
||||
--hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
|
||||
--seed 1024
|
||||
--tool-call-parser glm47
|
||||
--reasoning-parser glm45
|
||||
--enable-auto-tool-choice
|
||||
--max-num-seqs 6
|
||||
--max-model-len 133120
|
||||
--max-num-batched-tokens 4096
|
||||
--trust-remote-code
|
||||
--gpu-memory-utilization 0.9
|
||||
--quantization ascend
|
||||
--enable-chunked-prefill
|
||||
--enable-prefix-caching
|
||||
--async-scheduling
|
||||
--compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
|
||||
--additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
|
||||
--speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
|
||||
benchmarks:
|
||||
perf_64k:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in65536-bs1500-prefix90-glm51
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 128
|
||||
max_out_len: 1024
|
||||
batch_size: 32
|
||||
request_rate: 0
|
||||
baseline: 539.11
|
||||
threshold: 0.97
|
||||
perf_128k:
|
||||
case_type: performance
|
||||
dataset_path: vllm-ascend/GSM8K-in131072-bs500-prefix90-glm51
|
||||
request_conf: vllm_api_stream_chat
|
||||
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
||||
num_prompts: 256
|
||||
max_out_len: 1024
|
||||
batch_size: 64
|
||||
request_rate: 0
|
||||
baseline: 796.35
|
||||
threshold: 0.97
|
||||
Reference in New Issue
Block a user