332 lines
9.4 KiB
YAML
332 lines
9.4 KiB
YAML
# ==========================================
|
|
# Shared Configurations
|
|
# ==========================================
|
|
|
|
_envs: &envs
|
|
HCCL_BUFFSIZE: "512"
|
|
SERVER_PORT: "DEFAULT_PORT"
|
|
HCCL_OP_EXPANSION_MODE: "AIV"
|
|
OMP_PROC_BIND: "false"
|
|
OMP_NUM_THREADS: "1"
|
|
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
|
TASK_QUEUE_ENABLE: "1"
|
|
VLLM_ASCEND_ENABLE_MLAPO: "1"
|
|
VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
|
|
VLLM_ASCEND_ENABLE_NZ: "1"
|
|
VLLM_ENGINE_READY_TIMEOUT_S: "3000"
|
|
|
|
_server_cmd: &server_cmd
|
|
- "--enable-expert-parallel"
|
|
- "--allowed-local-media-path"
|
|
- "/"
|
|
- "--tensor-parallel-size"
|
|
- "4"
|
|
- "--data-parallel-size"
|
|
- "4"
|
|
- "--port"
|
|
- "$SERVER_PORT"
|
|
- "--quantization"
|
|
- "ascend"
|
|
- "--trust-remote-code"
|
|
- "--gpu-memory-utilization"
|
|
- "0.9"
|
|
- "--seed"
|
|
- "42"
|
|
- "--mm-processor-cache-gb"
|
|
- "0"
|
|
- "--mm-encoder-tp-mode"
|
|
- "data"
|
|
|
|
_benchmark_acc: &_benchmark_acc
|
|
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
|
|
|
|
_benchmark_TPOT50_3_5k_1_5k: &_benchmark_TPOT50_3_5k_1_5k
|
|
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: 392
|
|
max_out_len: 1500
|
|
batch_size: 98
|
|
trust_remote_code: true
|
|
request_rate: 0
|
|
baseline: 1
|
|
threshold: 0.97
|
|
|
|
_benchmark_TPOT50_16K_1K: &_benchmark_TPOT50_16K_1K
|
|
case_type: performance
|
|
dataset_path: vllm-ascend/GSM8K_prefix0_in16384_bs200_kimi
|
|
request_conf: vllm_api_stream_chat
|
|
dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
|
|
num_prompts: 56
|
|
max_out_len: 1024
|
|
batch_size: 14
|
|
trust_remote_code: true
|
|
request_rate: 0
|
|
baseline: 1
|
|
threshold: 0.97
|
|
|
|
_benchmark_TPOT20_16K_1K: &_benchmark_TPOT20_16K_1K
|
|
case_type: performance
|
|
dataset_path: vllm-ascend/GSM8K_prefix0_in16384_bs200_kimi
|
|
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
|
|
trust_remote_code: true
|
|
request_rate: 0
|
|
baseline: 1
|
|
threshold: 0.97
|
|
|
|
_benchmark_TPOT50_64K_1K_pc90: &_benchmark_TPOT50_64K_1K_pc90
|
|
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: 56
|
|
max_out_len: 1024
|
|
batch_size: 14
|
|
trust_remote_code: true
|
|
request_rate: 0
|
|
baseline: 1
|
|
threshold: 0.97
|
|
|
|
_benchmark_TPOT20_64K_1K_pc90: &_benchmark_TPOT20_64K_1K_pc90
|
|
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: 4
|
|
max_out_len: 1024
|
|
batch_size: 1
|
|
trust_remote_code: true
|
|
request_rate: 0
|
|
baseline: 1
|
|
threshold: 0.97
|
|
|
|
_benchmark_TPOT50_128K_1K_pc90: &_benchmark_TPOT50_128K_1K_pc90
|
|
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: 8
|
|
max_out_len: 1024
|
|
batch_size: 2
|
|
trust_remote_code: true
|
|
request_rate: 0
|
|
baseline: 1
|
|
threshold: 0.97
|
|
|
|
_benchmark_TPOT20_128K_1K_pc90: &_benchmark_TPOT20_128K_1K_pc90
|
|
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: 4
|
|
max_out_len: 1024
|
|
batch_size: 1
|
|
trust_remote_code: true
|
|
request_rate: 0
|
|
baseline: 1
|
|
threshold: 0.97
|
|
|
|
# ==========================================
|
|
# ACTUAL TEST CASES
|
|
# ==========================================
|
|
|
|
test_cases:
|
|
- name: "Kimi-K2.6-W4A8-Case"
|
|
model: "Eco-Tech/Kimi-K2.6-w4a8"
|
|
envs:
|
|
<<: *envs
|
|
server_cmd: *server_cmd
|
|
server_cmd_extra:
|
|
- "--max-model-len"
|
|
- "32768"
|
|
- "--no-enable-prefix-caching"
|
|
- "--enable-chunked-prefill"
|
|
- "--max-num-batched-tokens"
|
|
- "16384"
|
|
- "--max-num-seqs"
|
|
- "500"
|
|
- "--compilation-config"
|
|
- '{"cudagraph_capture_sizes":[48,64,80,96,112,128], "cudagraph_mode":"FULL_DECODE_ONLY"}'
|
|
- "--speculative-config"
|
|
- '{"method":"eagle3", "model":"lightseekorg/kimi-k2.6-eagle3", "num_speculative_tokens":3}'
|
|
- "--additional-config"
|
|
- '{"enable_shared_expert_dp":true}'
|
|
benchmarks:
|
|
perf: *_benchmark_TPOT50_3_5k_1_5k
|
|
- name: "Kimi-K2.6-W4A8-TOPT50-16k-1k"
|
|
model: "Eco-Tech/Kimi-K2.6-w4a8"
|
|
envs:
|
|
<<: *envs
|
|
server_cmd: *server_cmd
|
|
server_cmd_extra:
|
|
- "--max-model-len"
|
|
- "32768"
|
|
- "--no-enable-prefix-caching"
|
|
- "--enable-chunked-prefill"
|
|
- "--max-num-batched-tokens"
|
|
- "8192"
|
|
- "--max-num-seqs"
|
|
- "16"
|
|
- "--compilation-config"
|
|
- '{"cudagraph_capture_sizes":[4,8,12,16,20], "cudagraph_mode":"FULL_DECODE_ONLY"}'
|
|
- "--speculative-config"
|
|
- '{"method":"eagle3", "model":"lightseekorg/kimi-k2.6-eagle3", "num_speculative_tokens":3}'
|
|
- "--additional-config"
|
|
- '{"enable_shared_expert_dp":true}'
|
|
benchmarks:
|
|
perf: *_benchmark_TPOT50_16K_1K
|
|
- name: "Kimi-K2.6-W4A8-TOPT20-16k-1k"
|
|
model: "Eco-Tech/Kimi-K2.6-w4a8"
|
|
envs:
|
|
<<: *envs
|
|
server_cmd: *server_cmd
|
|
server_cmd_extra:
|
|
- "--max-model-len"
|
|
- "18000"
|
|
- "--max-num-batched-tokens"
|
|
- "8192"
|
|
- "--max-num-seqs"
|
|
- "1"
|
|
- "--no-enable-prefix-caching"
|
|
- "--compilation-config"
|
|
- '{"cudagraph_capture_sizes":[8], "cudagraph_mode":"FULL_DECODE_ONLY"}'
|
|
- "--speculative-config"
|
|
- '{"method":"eagle3", "model":"lightseekorg/kimi-k2.6-eagle3", "num_speculative_tokens":7}'
|
|
- "--tool-call-parser"
|
|
- "kimi_k2"
|
|
- "--reasoning-parser"
|
|
- "kimi_k2"
|
|
- "--enable-auto-tool-choice"
|
|
benchmarks:
|
|
perf: *_benchmark_TPOT20_16K_1K
|
|
- name: "Kimi-K2.6-W4A8-TOPT50-64k-1k-pc90"
|
|
model: "Eco-Tech/Kimi-K2.6-w4a8"
|
|
envs:
|
|
<<: *envs
|
|
HCCL_BUFFSIZE: "400"
|
|
VLLM_ASCEND_BALANCE_SCHEDULING: "1"
|
|
server_cmd: *server_cmd
|
|
server_cmd_extra:
|
|
- "--tool-call-parser"
|
|
- "kimi_k2"
|
|
- "--reasoning-parser"
|
|
- "kimi_k2"
|
|
- "--enable-auto-tool-choice"
|
|
- "--api-server-count"
|
|
- "1"
|
|
- "--aggregate-engine-logging"
|
|
- "--max-num-seqs"
|
|
- "3"
|
|
- "--max-model-len"
|
|
- "67000"
|
|
- "--max-num-batched-tokens"
|
|
- "8192"
|
|
- "--enable-prefix-caching"
|
|
- "--compilation-config"
|
|
- '{"cudagraph_capture_sizes":[8,12], "cudagraph_mode":"FULL_DECODE_ONLY"}'
|
|
- "--speculative-config"
|
|
- '{"method":"eagle3", "model":"lightseekorg/kimi-k2.6-eagle3", "num_speculative_tokens":3}'
|
|
benchmarks:
|
|
perf: *_benchmark_TPOT50_64K_1K_pc90
|
|
- name: "Kimi-K2.6-W4A8-TOPT20-64k-1k-pc90"
|
|
model: "Eco-Tech/Kimi-K2.6-w4a8"
|
|
envs:
|
|
<<: *envs
|
|
HCCL_BUFFSIZE: "400"
|
|
VLLM_ASCEND_BALANCE_SCHEDULING: "1"
|
|
server_cmd:
|
|
- "--port"
|
|
- "$SERVER_PORT"
|
|
- "--tool-call-parser"
|
|
- "kimi_k2"
|
|
- "--reasoning-parser"
|
|
- "kimi_k2"
|
|
- "--enable-auto-tool-choice"
|
|
- "--quantization"
|
|
- "ascend"
|
|
- "--allowed-local-media-path"
|
|
- "/"
|
|
- "--trust-remote-code"
|
|
- "--data-parallel-size"
|
|
- "1"
|
|
- "--tensor-parallel-size"
|
|
- "16"
|
|
- "--enable-expert-parallel"
|
|
- "--max-num-seqs"
|
|
- "8"
|
|
- "--max-model-len"
|
|
- "67000"
|
|
- "--max-num-batched-tokens"
|
|
- "8192"
|
|
- "--gpu-memory-utilization"
|
|
- "0.9"
|
|
- "--seed"
|
|
- "1024"
|
|
- "--enable-prefix-caching"
|
|
- "--compilation-config"
|
|
- '{"cudagraph_capture_sizes":[4,8,16], "cudagraph_mode":"FULL_DECODE_ONLY"}'
|
|
- "--speculative-config"
|
|
- '{"method":"eagle3", "model":"lightseekorg/kimi-k2.6-eagle3", "num_speculative_tokens":3}'
|
|
- "--mm-processor-cache-gb"
|
|
- "0"
|
|
- "--mm-encoder-tp-mode"
|
|
- "data"
|
|
benchmarks:
|
|
perf: *_benchmark_TPOT20_64K_1K_pc90
|
|
- name: "Kimi-K2.6-W4A8-TOPT50-128k-1k-pc90"
|
|
model: "Eco-Tech/Kimi-K2.6-w4a8"
|
|
envs:
|
|
<<: *envs
|
|
server_cmd: *server_cmd
|
|
server_cmd_extra:
|
|
- "--enable-chunked-prefill"
|
|
- "--max-num-seqs"
|
|
- "16"
|
|
- "--max-model-len"
|
|
- "133120"
|
|
- "--max-num-batched-tokens"
|
|
- "8192"
|
|
- "--compilation-config"
|
|
- '{"cudagraph_capture_sizes":[4,8,12,16,32], "cudagraph_mode":"FULL_DECODE_ONLY"}'
|
|
- "--speculative-config"
|
|
- '{"method":"eagle3", "model":"lightseekorg/kimi-k2.6-eagle3", "num_speculative_tokens":3}'
|
|
- "--additional-config"
|
|
- '{"enable_shared_expert_dp":true}'
|
|
benchmarks:
|
|
perf: *_benchmark_TPOT50_128K_1K_pc90
|
|
- name: "Kimi-K2.6-W4A8-TOPT20-128k-1k-pc90"
|
|
model: "Eco-Tech/Kimi-K2.6-w4a8"
|
|
envs:
|
|
<<: *envs
|
|
server_cmd: *server_cmd
|
|
server_cmd_extra:
|
|
- "--enable-chunked-prefill"
|
|
- "--max-num-seqs"
|
|
- "16"
|
|
- "--max-model-len"
|
|
- "133120"
|
|
- "--max-num-batched-tokens"
|
|
- "8192"
|
|
- "--compilation-config"
|
|
- '{"cudagraph_capture_sizes":[4,8,12,16,32], "cudagraph_mode":"FULL_DECODE_ONLY"}'
|
|
- "--speculative-config"
|
|
- '{"method":"eagle3", "model":"lightseekorg/kimi-k2.6-eagle3", "num_speculative_tokens":3}'
|
|
- "--additional-config"
|
|
- '{"enable_shared_expert_dp":true}'
|
|
benchmarks:
|
|
perf: *_benchmark_TPOT20_128K_1K_pc90
|
|
<<: *_benchmark_acc
|