65eae6de7b
### What this PR does / why we need it?
Change recurrent_gated_delta_rule ops from triton to ascend C version
for better performance.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
- vLLM version: v0.15.0
- vLLM main:
9562912cea
---------
Signed-off-by: SunnyLee219 <3294305115@qq.com>
78 lines
1.8 KiB
YAML
78 lines
1.8 KiB
YAML
# ==========================================
|
|
# Shared Configurations
|
|
# ==========================================
|
|
|
|
_envs: &envs
|
|
OMP_NUM_THREADS: "10"
|
|
OMP_PROC_BIND: "false"
|
|
HCCL_BUFFSIZE: "1024"
|
|
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
|
|
SERVER_PORT: "DEFAULT_PORT"
|
|
|
|
_server_cmd: &server_cmd
|
|
- "--tensor-parallel-size"
|
|
- "4"
|
|
- "--port"
|
|
- "$SERVER_PORT"
|
|
- "--max-model-len"
|
|
- "40960"
|
|
- "--trust-remote-code"
|
|
- "--async-scheduling"
|
|
- "--no-enable-prefix-caching"
|
|
- "--enable-expert-parallel"
|
|
- "--gpu-memory-utilization"
|
|
- "0.8"
|
|
- "--max-num-seqs"
|
|
- "64"
|
|
- "--compilation-config"
|
|
- '{"cudagraph_capture_sizes": [64]}'
|
|
|
|
_benchmarks: &benchmarks
|
|
perf:
|
|
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: 256
|
|
max_out_len: 1500
|
|
batch_size: 64
|
|
baseline: 1
|
|
threshold: 0.97
|
|
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: 64
|
|
top_k: 20
|
|
baseline: 95
|
|
threshold: 5
|
|
|
|
# ==========================================
|
|
# ACTUAL TEST CASES
|
|
# ==========================================
|
|
|
|
test_cases:
|
|
- name: "Qwen3-Next-80B-A3B-Instruct-aclgraph-8192-a3"
|
|
model: "Qwen/Qwen3-Next-80B-A3B-Instruct"
|
|
envs:
|
|
<<: *envs
|
|
server_cmd: *server_cmd
|
|
server_cmd_extra:
|
|
- "--max-num-batched-tokens"
|
|
- "8192"
|
|
benchmarks:
|
|
<<: *benchmarks
|
|
|
|
- name: "Qwen3-Next-80B-A3B-Instruct-aclgraph-32768-a3"
|
|
model: "Qwen/Qwen3-Next-80B-A3B-Instruct"
|
|
envs:
|
|
<<: *envs
|
|
server_cmd: *server_cmd
|
|
server_cmd_extra:
|
|
- "--max-num-batched-tokens"
|
|
- "32768"
|
|
benchmarks:
|
|
<<: *benchmarks
|