Skip to content

[Executorch][kv cache] Make quantized cache return only the updated cache portion #9351

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

Closed
wants to merge 1 commit into from
Closed

[Executorch][kv cache] Make quantized cache return only the updated cache portion #9351

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: 24 additions & 6 deletions examples/models/llama/source_transformation/quantized_kv_cache.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,6 +48,9 @@ def __init__(
f"Only affine symmetric and asymmetric cache types are supported: got {cache_type}"
)

# This reduces dequantize overhead by dequantizing only upto the
# current input position.
self.return_sliced_cache = True
# For now supporting int8 only
self.use_custom_update_cache_op = use_custom_update_cache_op
self.quantized_cache_dtype = torch.int8
Expand Down Expand Up @@ -125,19 +128,34 @@ def update(self, input_pos, k_val, v_val):
self.v_cache_scales[:, input_pos] = v_scales
self.v_cache_zero_points[:, input_pos] = v_zero_points

if self.return_sliced_cache:
k_cache_out = self.k_cache[:, :input_pos+1]
k_cache_scales = self.k_cache_scales[:, :input_pos+1]
k_cache_zero_points = self.k_cache_zero_points[:, :input_pos+1]
v_cache_out = self.v_cache[:, :input_pos+1]
v_cache_scales = self.v_cache_scales[:, :input_pos+1]
v_cache_zero_points = self.v_cache_zero_points[:, :input_pos+1]
else:
k_cache_out = self.k_cache
k_cache_scales = self.k_cache_scales
k_cache_zero_points = self.k_cache_zero_points
v_cache_out = self.v_cache
v_cache_scales = self.v_cache_scales
v_cache_zero_points = self.v_cache_zero_points

k_out = torch.ops.quantized_decomposed.dequantize_per_token(
self.k_cache,
self.k_cache_scales,
self.k_cache_zero_points,
k_cache_out,
k_cache_scales,
k_cache_zero_points,
torch.iinfo(self.quantized_cache_dtype).min,
torch.iinfo(self.quantized_cache_dtype).max,
self.quantized_cache_dtype,
self.cache_fp_type,
)
v_out = torch.ops.quantized_decomposed.dequantize_per_token(
self.v_cache,
self.v_cache_scales,
self.v_cache_zero_points,
v_cache_out,
v_cache_scales,
v_cache_zero_points,
torch.iinfo(self.quantized_cache_dtype).min,
torch.iinfo(self.quantized_cache_dtype).max,
self.quantized_cache_dtype,
Expand Down
Loading