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

LLM: change fp16 benchmark to model.half (#10477)

Browse source 
* LLM: change fp16 benchmark to model.half

* fix
This commit is contained in:
Jin Qiao 2024-03-20 13:38:39 +08:00 committed by GitHub
parent 749bedaf1e
commit e41d556436
1 changed files with 8 additions and 4 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

12
python/llm/dev/benchmark/all-in-one/run.py
View file

@ -930,27 +930,31 @@ def run_transformer_int4_fp16_gpu_win(repo_id,
# which convert the relevant layers in the model into INT4 format # which convert the relevant layers in the model into INT4 format
st = time.perf_counter() st = time.perf_counter()
if repo_id in CHATGLM_IDS: if repo_id in CHATGLM_IDS:
model = AutoModel.from_pretrained(model_path, load_in_low_bit=low_bit, optimize_model=True, torch_dtype=torch.float16, model = AutoModel.from_pretrained(model_path, load_in_low_bit=low_bit, optimize_model=True,
trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval() trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = model.half()
model = model.to('xpu') model = model.to('xpu')
elif repo_id in LLAMA_IDS: elif repo_id in LLAMA_IDS:
model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit=low_bit, optimize_model=True, torch_dtype=torch.float16, model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit=low_bit, optimize_model=True,
trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval() trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval()
tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True) tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = model.half()
model = model.to('xpu') model = model.to('xpu')
elif repo_id in LLAVA_IDS: elif repo_id in LLAVA_IDS:
llava_repo_dir = os.environ.get('LLAVA_REPO_DIR') llava_repo_dir = os.environ.get('LLAVA_REPO_DIR')
sys.path.append(rf"{llava_repo_dir}") sys.path.append(rf"{llava_repo_dir}")
from llava.model.language_model.llava_llama import LlavaLlamaForCausalLM from llava.model.language_model.llava_llama import LlavaLlamaForCausalLM
model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit=low_bit, optimize_model=True, torch_dtype=torch.float16, model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit=low_bit, optimize_model=True,
trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval() trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = model.half()
model = model.to('xpu') model = model.to('xpu')
else: else:
model = AutoModelForCausalLM.from_pretrained(model_path, optimize_model=True, load_in_low_bit=low_bit, torch_dtype=torch.float16, model = AutoModelForCausalLM.from_pretrained(model_path, optimize_model=True, load_in_low_bit=low_bit,
trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval() trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = model.half()
model = model.to('xpu') model = model.to('xpu')
if isinstance(model, GPTJForCausalLM): if isinstance(model, GPTJForCausalLM):
# For gpt-j model family, this optimization can provide a better performance. # For gpt-j model family, this optimization can provide a better performance.