Files
enginex-ascend-910-vllm/tests/e2e/nightly/single_node/models/configs/DeepSeek-V3.2-W8A8.yaml
Sun Ruoxi 7f8a1b1f7a init v0.23.0
Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>
2026-08-27 15:11:51 +08:00

80 lines
2.3 KiB
YAML

# ==========================================
# ACTUAL TEST CASES
# ==========================================
test_cases:
- name: "DeepSeek-V3.2-W8A8-TP8-DP2"
model: "vllm-ascend/DeepSeek-V3.2-W8A8"
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"
VLLM_ENGINE_READY_TIMEOUT_S: "3000"
server_cmd:
- "--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"
- "--async-scheduling"
- "--no-enable-prefix-caching"
- "--enable-expert-parallel"
- "--gpu-memory-utilization"
- "0.95"
- "--compilation-config"
- '{"cudagraph_capture_sizes":[1,2,4,8,16,24,32,40,48], "cudagraph_mode":"FULL_DECODE_ONLY"}'
- "--speculative-config"
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp"}'
- "--reasoning-parser"
- "deepseek_v3"
- "--tokenizer_mode"
- "deepseek_v32"
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: 86.67
temperature: 1.0
top_p: 0.95
thinking: true
threshold: 10
perf_2:
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: 16
max_out_len: 1500
batch_size: 4
request_rate: 0
baseline: 1
threshold: 0.97