Skip to content

Maybe DSv4 #5

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

Open
IzzyPutterman wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Maybe DSv4 #5

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
47 changes: 35 additions & 12 deletions benchmark_serving.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,6 +57,7 @@
from argparse import ArgumentParser as FlexibleArgumentParser

from benchmark_utils import convert_to_pytorch_benchmark_format
from encoding_dsv4 import encode_messages as dsv4_encode_messages

MILLISECONDS_TO_SECONDS_CONVERSION = 1000

Expand Down Expand Up @@ -371,18 +372,25 @@ def sample_random_requests(
range_ratio: float,
tokenizer: PreTrainedTokenizerBase,
use_chat_template: bool = False,
dsv4: bool = False,
) -> List[Tuple[str, int, int]]:
prefix_token_ids = np.random.randint(0,
tokenizer.vocab_size,
size=prefix_len).tolist()
if use_chat_template:
chat_template_dummy = tokenizer.apply_chat_template(
[{"role": "user", "content": "a"}],
add_generation_prompt=True,
tokenize=False,
)
tokenized_chat_template_dummy = tokenizer.encode(chat_template_dummy, add_special_tokens=False)
chat_template_len = len(tokenized_chat_template_dummy) - 1
if dsv4:
template_dummy = dsv4_encode_messages(
[{"role": "user", "content": "a"}],
thinking_mode="thinking",
)
else:
template_dummy = tokenizer.apply_chat_template(
[{"role": "user", "content": "a"}],
add_generation_prompt=True,
tokenize=False,
)
tokenized_template_dummy = tokenizer.encode(template_dummy, add_special_tokens=False)
chat_template_len = len(tokenized_template_dummy) - 1
input_len = input_len - chat_template_len

input_lens = np.random.randint(
Expand All @@ -406,11 +414,17 @@ def sample_random_requests(
]
prompt = tokenizer.decode(re_encoded_sequence)
if use_chat_template:
prompt = tokenizer.apply_chat_template(
[{"role": "user", "content": prompt}],
add_generation_prompt=True,
tokenize=False,
)
if dsv4:
prompt = dsv4_encode_messages(
[{"role": "user", "content": prompt}],
thinking_mode="thinking",
)
else:
prompt = tokenizer.apply_chat_template(
[{"role": "user", "content": prompt}],
add_generation_prompt=True,
tokenize=False,
)
input_lens[i] += chat_template_len

input_requests.append((prompt, int(prefix_len + input_lens[i]),
Expand Down Expand Up @@ -967,6 +981,7 @@ def main(args: argparse.Namespace):
range_ratio=args.random_range_ratio,
tokenizer=tokenizer,
use_chat_template=args.use_chat_template,
dsv4=args.dsv4,
)

else:
Expand Down Expand Up @@ -1353,5 +1368,13 @@ def main(args: argparse.Namespace):
"launching the server. For each request, the "
"script chooses a LoRA module at random.")

dsv4_group = parser.add_argument_group("DeepSeek-V4 chat template options")
dsv4_group.add_argument(
"--dsv4",
action="store_true",
help="Use the DeepSeek-V4 chat template (encoding_dsv4.py) instead of "
"the tokenizer's built-in chat template. Applies to random and sonnet "
"datasets.",
)
args = parser.parse_args()
main(args)
Loading