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

LLM: Change benchmark bf16 load format. (#9035)

Browse source 
* LLM: Change benchmark bf16 load format.

* comment on bf16 chatglm.

* fix.
This commit is contained in:
Cengguang Zhang 2023-09-22 17:38:38 +08:00 committed by GitHub
parent 023555fb1f
commit 26213a5829
1 changed files with 3 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
python/llm/dev/benchmark/all-in-one/run.py
View file

@ -176,18 +176,16 @@ def run_pytorch_autocast_bf16(repo_id,
st = time.perf_counter() st = time.perf_counter()
if repo_id in ['THUDM/chatglm-6b', 'THUDM/chatglm2-6b']: if repo_id in ['THUDM/chatglm-6b', 'THUDM/chatglm2-6b']:
# TODO: need verify chatglm family run bf16. # TODO: need verify chatglm family run bf16.
model = AutoModel.from_pretrained(model_path, trust_remote_code=True, torch_dtype='auto').float() invalidInputError(False, "Currently pytorch do not support bfloat16 on cpu for chatglm models.")
#model = AutoModel.from_pretrained(model_path, trust_remote_code=True, torch_dtype='auto').bfloat()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
elif repo_id in ['meta-llama/Llama-2-7b-chat-hf','meta-llama/Llama-2-13b-chat-hf', elif repo_id in ['meta-llama/Llama-2-7b-chat-hf','meta-llama/Llama-2-13b-chat-hf',
'meta-llama/Llama-2-70b-chat-hf','decapoda-research/llama-7b-hf', 'meta-llama/Llama-2-70b-chat-hf','decapoda-research/llama-7b-hf',
'decapoda-research/llama-65b-hf','lmsys/vicuna-7b-v1.5', 'decapoda-research/llama-65b-hf','lmsys/vicuna-7b-v1.5',
'lmsys/vicuna-13b-v1.3','project-baize/merged-baize-30b']: 'lmsys/vicuna-13b-v1.3','project-baize/merged-baize-30b']:
model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, torch_dtype='auto') model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, torch_dtype=torch.bfloat16)
# Need to use LlamaTokenizer, reason please refer to issue: https://github.com/intel-analytics/BigDL/issues/8944 # Need to use LlamaTokenizer, reason please refer to issue: https://github.com/intel-analytics/BigDL/issues/8944
tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True) tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
else: else:
model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, torch_dtype='auto') model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
end = time.perf_counter() end = time.perf_counter()
print(">> loading of model costs {}s".format(end - st)) print(">> loading of model costs {}s".format(end - st))