ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

Add more control arguments for benchmark_vllm_throughput (#11291)

Browse source
This commit is contained in:
Guancheng Fu 2024-06-12 17:43:06 +08:00 committed by GitHub
parent 592f7aa61e
commit 2e75bbccf9
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 12 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

15
docker/llm/serving/xpu/docker/benchmark_vllm_throughput.py
View file

@ -77,6 +77,7 @@ def run_vllm(
gpu_memory_utilization: float = 0.9, gpu_memory_utilization: float = 0.9,
load_in_low_bit: str = "sym_int4", load_in_low_bit: str = "sym_int4",
max_num_batched_tokens: int = 5000, max_num_batched_tokens: int = 5000,
max_num_seqs: int = 256,
) -> float: ) -> float:
from vllm import SamplingParams from vllm import SamplingParams
from ipex_llm.vllm.xpu.engine import IPEXLLMClass as LLM from ipex_llm.vllm.xpu.engine import IPEXLLMClass as LLM
@ -94,7 +95,8 @@ def run_vllm(
device=device, device=device,
enable_prefix_caching=enable_prefix_caching, enable_prefix_caching=enable_prefix_caching,
load_in_low_bit=load_in_low_bit, load_in_low_bit=load_in_low_bit,
max_num_batched_tokens=max_num_batched_tokens,) max_num_batched_tokens=max_num_batched_tokens,
max_num_seqs=max_num_seqs,)
# Add the requests to the engine. # Add the requests to the engine.
@ -238,7 +240,8 @@ def main(args: argparse.Namespace):
args.tensor_parallel_size, args.seed, args.n, args.use_beam_search, args.tensor_parallel_size, args.seed, args.n, args.use_beam_search,
args.trust_remote_code, args.dtype, args.max_model_len, args.trust_remote_code, args.dtype, args.max_model_len,
args.enforce_eager, args.kv_cache_dtype, args.device, args.enforce_eager, args.kv_cache_dtype, args.device,
args.enable_prefix_caching, args.gpu_memory_utilization, args.load_in_low_bit, args.max_num_batched_tokens) args.enable_prefix_caching, args.gpu_memory_utilization, args.load_in_low_bit,
args.max_num_batched_tokens,args.max_num_seqs)
elif args.backend == "hf": elif args.backend == "hf":
assert args.tensor_parallel_size == 1 assert args.tensor_parallel_size == 1
elapsed_time = run_hf(requests, args.model, tokenizer, args.n, elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
@ -348,9 +351,15 @@ if __name__ == "__main__":
parser.add_argument('--max-num-batched-tokens', parser.add_argument('--max-num-batched-tokens',
type=int, type=int,
default=5000, default=4096,
help='maximum number of batched tokens per iteration') help='maximum number of batched tokens per iteration')
parser.add_argument('--max-num-seqs',
type=int,
default=256,
help='Maximum number of sequences per iteration.')
args = parser.parse_args() args = parser.parse_args()
if args.tokenizer is None: if args.tokenizer is None:
args.tokenizer = args.model args.tokenizer = args.model