2cb9195ff0
### What this PR does / why we need it? Fix the OOM (Out-of-Memory) error in the single-node-deepseek-v3-2-w8a8 nightly test of vllm-ascend: - Reduced the value of HCCL_BUFFSIZE - Lowered the gpu-memory-utilization Optimize service-side performance: Updated service-oriented configuration parameters (e.g., max-num-seqs, cudagraph_capture_sizes, batch_size) to improve the inference performance,so that the performance is closer to the optimal performance of the current mainline. Align performance baseline with main branch: Updated the performance baseline according to the latest performance data ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? The test has passed. https://github.com/vllm-project/vllm-ascend/actions/runs/23734079080/job/69134387320?pr=7793 --------- Signed-off-by: wyh145 <1987244901@qq.com>
79 lines
2.3 KiB
YAML
79 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:
|
|
HCCL_OP_EXPANSION_MODE: "AIV"
|
|
OMP_PROC_BIND: "false"
|
|
OMP_NUM_THREADS: "1"
|
|
HCCL_BUFFSIZE: "256"
|
|
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
|
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
|
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
|
VLLM_ENGINE_READY_TIMEOUT_S: "1800"
|
|
SERVER_PORT: "DEFAULT_PORT"
|
|
server_cmd:
|
|
- "--enable-expert-parallel"
|
|
- "--tensor-parallel-size"
|
|
- "8"
|
|
- "--data-parallel-size"
|
|
- "2"
|
|
- "--port"
|
|
- "$SERVER_PORT"
|
|
- "--max-model-len"
|
|
- "8192"
|
|
- "--max-num-batched-tokens"
|
|
- "8192"
|
|
- "--max-num-seqs"
|
|
- "8"
|
|
- "--trust-remote-code"
|
|
- "--quantization"
|
|
- "ascend"
|
|
- "--gpu-memory-utilization"
|
|
- "0.93"
|
|
- "--compilation-config"
|
|
- '{"cudagraph_capture_sizes":[4, 8, 16, 20, 24, 28, 32], "cudagraph_mode":"FULL_DECODE_ONLY"}'
|
|
- "--speculative-config"
|
|
- '{"num_speculative_tokens": 3, "method":"deepseek_mtp"}'
|
|
- "--additional-config"
|
|
- '{"layer_sharding": ["q_b_proj", "o_proj"]}'
|
|
- "--reasoning-parser"
|
|
- "deepseek_v3"
|
|
- "--tokenizer_mode"
|
|
- "deepseek_v32"
|
|
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: 5
|
|
perf_1:
|
|
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: 1
|
|
max_out_len: 1500
|
|
batch_size: 1
|
|
request_rate: 11.2
|
|
baseline: 1
|
|
threshold: 0.97
|
|
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: 128
|
|
max_out_len: 1500
|
|
batch_size: 32
|
|
request_rate: 11.2
|
|
baseline: 210
|
|
threshold: 0.97
|