Skip to content

removed redundancies from QEFFHybridCache #582

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

removed redundancies from QEFFHybridCache #582

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
30 changes: 12 additions & 18 deletions QEfficient/transformers/cache_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -346,15 +346,15 @@ def update(
sliding_window_pattern = cache_kwargs.get("sliding_window_pattern")
is_sliding_layer = torch.tensor(bool((layer_idx + 1) % sliding_window_pattern))
layer_ctx_len = self.key_cache[layer_idx].shape[2]
kv_position_ids = torch.where(
(~is_sliding_layer | (position_ids == -1)), position_ids, position_ids % (layer_ctx_len - 1)
)

kv_position_ids = torch.where(
is_sliding_layer & (position_ids.max() >= (layer_ctx_len - 1) * 2),
(position_ids + 1) % layer_ctx_len,
kv_position_ids,
)
if is_sliding_layer:
kv_position_ids = torch.where(position_ids == -1, position_ids, position_ids % (layer_ctx_len - 1))

kv_position_ids = torch.where(
position_ids.max() >= (layer_ctx_len - 1) * 2, (position_ids + 1) % layer_ctx_len, kv_position_ids
)
else:
kv_position_ids = position_ids

valid_mask = (kv_position_ids != -1).unsqueeze(1).unsqueeze(-1)
key_states = torch.where(valid_mask == 1, key_states, torch.zeros_like(key_states))
Expand All @@ -368,23 +368,17 @@ def update(
# Original Gather
ctx_len = self.key_cache[layer_idx].shape[2]
ctx_indices = torch.arange(ctx_len)[None, None, ...]
gather_limit = kv_position_ids.max(1, keepdim=True).values.unsqueeze(1)
Copy link
Contributor

@quic-rishinr quic-rishinr Nov 4, 2025

Choose a reason for hiding this comment

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

if we are using position_ids it would go overboard for sliding window right?

gather_limit = position_ids.max(1, keepdim=True).values.unsqueeze(1)
invalid_mask = ctx_indices > gather_limit
if torch.onnx.is_in_onnx_export():
invalid_idx_value = torch.iinfo(torch.int32).max
else:
invalid_idx_value = 0
ctx_indices = torch.where(invalid_mask, invalid_idx_value, ctx_indices)

all_indices = torch.arange(layer_ctx_len) + kv_position_ids.max() + 1
rolling_indices = torch.where(all_indices > layer_ctx_len - 1, all_indices % layer_ctx_len, all_indices)
final_indices = torch.where(
(is_sliding_layer & (position_ids.max() >= (layer_ctx_len - 1))), rolling_indices, ctx_indices
)
k_out = CtxGatherFunc.apply(k_out, final_indices)
v_out = CtxGatherFunc.apply(v_out, final_indices)
ctx_v_out = torch.where(invalid_mask.unsqueeze(-1), torch.tensor(0.0, dtype=torch.float32), v_out)
v_out = torch.where((is_sliding_layer & (position_ids.max() >= (layer_ctx_len - 1))), v_out, ctx_v_out)
k_out = CtxGatherFunc.apply(k_out, ctx_indices)
v_out = CtxGatherFunc.apply(v_out, ctx_indices)
v_out = torch.where(invalid_mask.unsqueeze(-1), torch.tensor(0.0, dtype=torch.float32), v_out)
return k_out, v_out


Expand Down
Loading