Files
enginex-ascend-910-vllm/tests/e2e/weekly/single_node/configs/DeepSeek-V3.2-W8A8_A3_weekly.yaml
Sun Ruoxi 7f8a1b1f7a init v0.23.0
Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>
2026-08-27 15:11:51 +08:00

150 lines
4.4 KiB
YAML

# ==========================================
# ACTUAL TEST CASES
# ==========================================
_envs: &envs
OMP_PROC_BIND: "false"
OMP_NUM_THREADS: "1"
HCCL_OP_EXPANSION_MODE: "AIV"
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
VLLM_USE_V1: "1"
HCCL_BUFFSIZE: "256"
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_FLASHCOMM1: "1"
SERVER_PORT: "DEFAULT_PORT"
_server_cmd: &server_cmd
- "--enable-expert-parallel"
- "--tensor-parallel-size"
- "8"
- "--data-parallel-size"
- "2"
- "--port"
- "$SERVER_PORT"
- "--seed"
- "1024"
- "--max-model-len"
- "67000"
- "--max-num-batched-tokens"
- "4096"
- "--max-num-seqs"
- "8"
- "--trust-remote-code"
- "--quantization"
- "ascend"
- "--gpu-memory-utilization"
- "0.95"
- "--compilation-config"
- '{"cudagraph_capture_sizes":[1,2,4,8,16,24,32,40,48], "cudagraph_mode":"FULL_DECODE_ONLY"}'
- "--no-enable-prefix-caching"
- "--speculative-config"
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp"}'
- "--reasoning-parser"
- "deepseek_v3"
- "--tokenizer-mode"
- "deepseek_v32"
_benchmarks: &benchmarks
perf_16384_bs4:
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: 4
max_out_len: 1024
batch_size: 1
request_rate: 0
baseline: 26.10
threshold: 0.97
perf_16384_bs16:
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: 16
max_out_len: 1024
batch_size: 4
request_rate: 0.5
baseline: 60.53
threshold: 0.97
perf_32768_bs4:
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: 4
max_out_len: 512
batch_size: 1
request_rate: 0
baseline: 22.68
threshold: 0.97
perf_32768_bs8:
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: 8
max_out_len: 512
batch_size: 2
request_rate: 0.2
baseline: 32.73
threshold: 0.97
perf_65536_bs4:
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
baseline: 22.77
threshold: 0.97
perf_65536_bs8:
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: 8
max_out_len: 1024
batch_size: 2
request_rate: 1
baseline: 33.80
threshold: 0.97
_benchmarks_2048: &benchmarks_2048
perf_16384_bs4:
case_type: performance
dataset_path: vllm-ascend/GSM8K-in2048-bs4000-ds3.2
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: 220
threshold: 0.97
test_cases:
- name: "DeepSeek-V3.2-W8A8-weekly"
model: "vllm-ascend/DeepSeek-V3.2-W8A8"
envs:
<<: *envs
server_cmd: *server_cmd
benchmarks:
<<: *benchmarks
- name: "DeepSeek-V3.2-W8A8-mix-perf"
model: "vllm-ascend/DeepSeek-V3.2-W8A8"
envs:
<<: *envs
server_cmd: *server_cmd
server_cmd_extra:
- "--async-scheduling"
benchmarks:
<<: *benchmarks_2048