Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

feat : add npu to ppl #12071

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

feat : add npu to ppl #12071

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c66716b
fix: add npu to ppl
cranechu0131 Sep 12, 2024
70ee2e7
fix: remove lm_eval
cranechu0131 Sep 12, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 16 additions & 4 deletions python/llm/dev/benchmark/perplexity/run_wikitext.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,7 +36,18 @@
parser.add_argument("--mixed_precision", action="store_true")
args = parser.parse_args()

if args.precision == "fp16": # ipex fp16
if args.device == "npu":
from ipex_llm.transformers.npu_model import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(
args.model_path,
trust_remote_code=True,
torch_dtype=torch.float16,
max_output_len=4096,
max_prompt_len=4096,
load_in_low_bit=args.precision,
attn_implementation="eager"
)
elif args.precision == "fp16": # ipex fp16
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(args.model_path,
use_cache=args.use_cache,
Expand All @@ -57,7 +68,7 @@
trust_remote_code=True,
mixed_precision=args.mixed_precision)
model = model.half()
model = model.to(args.device)

model = model.eval()

from transformers import AutoTokenizer
Expand Down Expand Up @@ -98,7 +109,7 @@ def parse_kwargs(kwstr):
else:
end_loc = begin_loc + stride
trg_len = -stride//2
input_ids = encodings.input_ids[:, begin_loc:end_loc].to(args.device)
input_ids = encodings.input_ids[:, begin_loc:end_loc]
if args.stride == 0: input_ids[:, 0] = tokenizer.bos_token_id
target_ids = input_ids.clone()
target_ids[:, :-trg_len] = -100
Expand All @@ -110,6 +121,7 @@ def parse_kwargs(kwstr):
# N.B. the model only calculates loss over trg_len - 1 labels, because it internally shifts the labels
# to the left by 1.
neg_log_likelihood = outputs.loss
print(neg_log_likelihood)

nlls.append(neg_log_likelihood)
if "xpu" in args.device:
Expand All @@ -118,6 +130,6 @@ def parse_kwargs(kwstr):
prev_end_loc = end_loc
if end_loc == seq_len:
break

print(neg_log_likelihood)
ppl = torch.exp(torch.stack(nlls).mean())
print("Final ppl estimate: {}".format(ppl.item()))
Loading