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

[CodeGen] Deepsparse Pipeline #1078

Closed
wants to merge 1 commit into from
Closed

[CodeGen] Deepsparse Pipeline #1078

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
14 changes: 10 additions & 4 deletions src/deepsparse/transformers/pipelines/text_generation.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -133,6 +133,13 @@ def __init__(

# override tokenizer to pad to left
self.tokenizer.padding_side = "left"
# read from the tokenizer whether is
# uses a prefix to determine the sos token
self.sos_token_offset = 0
if hasattr(self.tokenizer, "prefix"):
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is this a universal attribute or just for OPT?
also let's move this to a helper function

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Just for OPT. We do not have this attribute neither for CodeGen nor for BLOOM.

self.sos_token_offset = len(self.tokenizer.prefix)
if len(self.tokenizer.prefix) > 1:
raise NotImplementedError("Prefix length > 1 not supported yet")

@staticmethod
def route_input_to_bucket(
Expand Down Expand Up @@ -416,7 +423,6 @@ def assemble_kv_cache(
self,
cache_values: List[numpy.ndarray],
tokens: List[int],
consider_sos_token: bool = False,
) -> Dict[str, numpy.ndarray]:
"""
Restructure the kv cache values from the engine output, so
Expand All @@ -427,7 +433,7 @@ def assemble_kv_cache(

There are two modes:
1. Some values in the cache represent dummy (pad) tokens, padding is
to the left, so the left-most cache value is deleted
to the left, so the left-most cache value is deleted
2. The cache is saturated with non-dummy (meaningful) tokens:
- if there is a mandatory start-of-sequence (SOS) token,
we delete the left-most cache value that is not a cache
Expand All @@ -436,12 +442,12 @@ def assemble_kv_cache(

:param cache_values: the cache values from the engine output
:param tokens: the tokens from the previous inference run
:param consider_sos_token: whether to consider the SOS token in the cache

:return kv_cache: the restructured cache values
"""
for idx, cache_value in enumerate(cache_values):
if len(tokens) > self.sequence_length - 1:
idx_to_remove = int(not consider_sos_token)
idx_to_remove = int(self.sos_token_offset)
else:
idx_to_remove = 0

Expand Down