92 lines
2.3 KiB
YAML
92 lines
2.3 KiB
YAML
# ==========================================
|
|
# Shared Configurations
|
|
# ==========================================
|
|
|
|
_envs: &envs
|
|
OMP_NUM_THREADS: "100"
|
|
OMP_PROC_BIND: "false"
|
|
HCCL_BUFFSIZE: "1024"
|
|
VLLM_RPC_TIMEOUT: "3600000"
|
|
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "3600000"
|
|
SERVER_PORT: "DEFAULT_PORT"
|
|
VLLM_ENGINE_READY_TIMEOUT_S: "7200"
|
|
|
|
_server_cmd: &server_cmd
|
|
- "--quantization"
|
|
- "ascend"
|
|
- "--seed"
|
|
- "1024"
|
|
- "--no-enable-prefix-caching"
|
|
- "--data-parallel-size"
|
|
- "2"
|
|
- "--tensor-parallel-size"
|
|
- "8"
|
|
- "--enable-expert-parallel"
|
|
- "--port"
|
|
- "$SERVER_PORT"
|
|
- "--max-model-len"
|
|
- "40960"
|
|
- "--max-num-seqs"
|
|
- "14"
|
|
- "--trust-remote-code"
|
|
|
|
_benchmarks_gsm8k: &benchmarks_gsm8k
|
|
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: 32768
|
|
batch_size: 32
|
|
baseline: 95
|
|
threshold: 10
|
|
|
|
_benchmarks_aime: &benchmarks_aime
|
|
acc:
|
|
case_type: accuracy
|
|
dataset_path: vllm-ascend/aime2024
|
|
request_conf: vllm_api_general_chat
|
|
dataset_conf: aime2024/aime2024_gen_0_shot_chat_prompt
|
|
max_out_len: 32768
|
|
batch_size: 32
|
|
baseline: 86.67
|
|
threshold: 10
|
|
|
|
# ==========================================
|
|
# ACTUAL TEST CASES
|
|
# ==========================================
|
|
|
|
test_cases:
|
|
- name: "MTPX-DeepSeek-R1-0528-W8A8-mtp2"
|
|
model: "vllm-ascend/DeepSeek-R1-0528-W8A8"
|
|
envs:
|
|
<<: *envs
|
|
server_cmd: *server_cmd
|
|
server_cmd_extra:
|
|
- "--max-num-batched-tokens"
|
|
- "4096"
|
|
- "--speculative-config"
|
|
- '{"num_speculative_tokens": 2, "method": "mtp"}'
|
|
- "--gpu-memory-utilization"
|
|
- "0.92"
|
|
benchmarks:
|
|
<<: *benchmarks_gsm8k
|
|
|
|
- name: "MTPX-DeepSeek-R1-0528-W8A8-mtp3"
|
|
model: "vllm-ascend/DeepSeek-R1-0528-W8A8"
|
|
envs:
|
|
<<: *envs
|
|
HCCL_OP_EXPANSION_MODE: "AIV"
|
|
server_cmd: *server_cmd
|
|
server_cmd_extra:
|
|
- "--max-num-batched-tokens"
|
|
- "2048"
|
|
- "--speculative-config"
|
|
- '{"num_speculative_tokens": 3, "method": "mtp"}'
|
|
- "--gpu-memory-utilization"
|
|
- "0.9"
|
|
- "--compilation-config"
|
|
- '{"cudagraph_capture_sizes": [56], "cudagraph_mode": "FULL_DECODE_ONLY"}'
|
|
benchmarks:
|
|
<<: *benchmarks_aime
|