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

Generate: Mistral/Mixtral FA2 cache fix when going beyond the context window #28037

Merged
merged 8 commits into from
Dec 14, 2023
+28 −10
Merged

Generate: Mistral/Mixtral FA2 cache fix when going beyond the context window #28037

Show file tree
Hide file tree
Changes from 7 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
aacffa8
add breakpoint
gante Dec 14, 2023
70cc502
only slice if the cache has contents
gante Dec 14, 2023
564750c
get the right slice
gante Dec 14, 2023
f038066
.
gante Dec 14, 2023
3f24fb0
.
gante Dec 14, 2023
d4867f7
tmp commit
gante Dec 14, 2023
cf7cc6d
fix FA2 + mis(x)tral
gante Dec 14, 2023
fbfb543
add self.layer_idx exception (as suggested by @tomaarsen)
gante Dec 14, 2023
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
13 changes: 8 additions & 5 deletions src/transformers/models/mistral/modeling_mistral.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -385,11 +385,16 @@ def forward(

if past_key_value is not None:
Copy link
Member Author

Choose a reason for hiding this comment

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

context: when use_cache is True, past_key_value is now a Cache object even if it is an empty cache (previously it was None).

As such, with a slicing window, we need to check whether the cache has contents before attempting to slice, as we can't slice None.

# Activate slicing cache only if the config has a value `sliding_windows` attribute
if getattr(self.config, "sliding_window", None) is not None and kv_seq_len > self.config.sliding_window:
cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
if (
getattr(self.config, "sliding_window", None) is not None
and kv_seq_len > self.config.sliding_window
and cache_has_contents
):
slicing_tokens = 1 - self.config.sliding_window

past_key = past_key_value[0]
past_value = past_key_value[1]
past_key = past_key_value[self.layer_idx][0]
Copy link
Member

Choose a reason for hiding this comment

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

self.layer_idx is guaranteed to be defined here, right?

Copy link
Member Author

Choose a reason for hiding this comment

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

Nope, good catch! Going to add an appropriate exception.

past_value = past_key_value[self.layer_idx][1]

past_key = past_key[:, :, slicing_tokens:, :].contiguous()
past_value = past_value[:, :, slicing_tokens:, :].contiguous()
Expand All @@ -400,8 +405,6 @@ def forward(
f" {past_key.shape}"
)

past_key_value = (past_key, past_value)
Copy link
Member Author

@gante gante Dec 14, 2023 •
edited
Loading

Choose a reason for hiding this comment

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

past_key_value is now a Cache instance that is updated in place with the .update() function (L413 in the updated file). We don't need to set it.


if attention_mask is not None:
attention_mask = attention_mask[:, slicing_tokens:]
attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
Expand Down
13 changes: 8 additions & 5 deletions src/transformers/models/mixtral/modeling_mixtral.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -436,11 +436,16 @@ def forward(

if past_key_value is not None:
# Activate slicing cache only if the config has a value `sliding_windows` attribute
if getattr(self.config, "sliding_window", None) is not None and kv_seq_len > self.config.sliding_window:
cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
if (
getattr(self.config, "sliding_window", None) is not None
and kv_seq_len > self.config.sliding_window
and cache_has_contents
):
slicing_tokens = 1 - self.config.sliding_window

past_key = past_key_value[0]
past_value = past_key_value[1]
past_key = past_key_value[self.layer_idx][0]
past_value = past_key_value[self.layer_idx][1]

past_key = past_key[:, :, slicing_tokens:, :].contiguous()
past_value = past_value[:, :, slicing_tokens:, :].contiguous()
Expand All @@ -451,8 +456,6 @@ def forward(
f" {past_key.shape}"
)

past_key_value = (past_key, past_value)

if attention_mask is not None:
attention_mask = attention_mask[:, slicing_tokens:]
attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
Expand Down
Loading