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

[LLM] Fix llm arc ut oom (#9300)

Browse source 
* Move model to cpu after testing so that gpu memory is deallocated

* Add code comment

---------

Co-authored-by: sgwhat <ge.song@intel.com>
This commit is contained in:
Cheen Hau, 俊豪 2023-10-30 14:38:34 +08:00 committed by GitHub
parent ee5becdd61
commit cee9eaf542
2 changed files with 4 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
python/llm/test/inference_gpu/test_optimize_model.py
View file

@ -41,6 +41,7 @@ def test_optimize_model(Model, Tokenizer, model_path):
trust_remote_code=True) trust_remote_code=True)
model = model.to(device) model = model.to(device)
logits_base_model = (model(input_ids)).logits logits_base_model = (model(input_ids)).logits
model.to('cpu') # deallocate gpu memory
model = Model.from_pretrained(model_path, model = Model.from_pretrained(model_path,
load_in_4bit=True, load_in_4bit=True,
@ -48,6 +49,7 @@ def test_optimize_model(Model, Tokenizer, model_path):
trust_remote_code=True) trust_remote_code=True)
model = model.to(device) model = model.to(device)
logits_optimized_model = (model(input_ids)).logits logits_optimized_model = (model(input_ids)).logits
model.to('cpu')
diff = abs(logits_base_model - logits_optimized_model).flatten() diff = abs(logits_base_model - logits_optimized_model).flatten()

3
python/llm/test/inference_gpu/test_transformers_api.py
View file

@ -41,10 +41,11 @@ def test_completion(Model, Tokenizer, model_path, prompt, answer):
load_in_4bit=True, load_in_4bit=True,
optimize_model=True, optimize_model=True,
trust_remote_code=True) trust_remote_code=True)
model = model.to(device) model = model.to(device) # deallocate gpu memory
input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device) input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device)
output = model.generate(input_ids, max_new_tokens=32) output = model.generate(input_ids, max_new_tokens=32)
model.to('cpu')
output_str = tokenizer.decode(output[0], skip_special_tokens=True) output_str = tokenizer.decode(output[0], skip_special_tokens=True)
assert answer in output_str assert answer in output_str