Skip to content

feat: QuantizedRotatingKVCache + KVSplit (K/V different bits) #1074

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
deceptech-packet-ninja wants to merge 3 commits into
base: main
Choose a base branch
from
+428 −20
Open

feat: QuantizedRotatingKVCache + KVSplit (K/V different bits) #1074

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
6377f15
feat: add --kv-bits, --kv-group-size, --quantized-kv-start to server
deceptech-packet-ninja Mar 30, 2026
6cebf08
feat: QuantizedRotatingKVCache + KVSplit (K/V different bits)
deceptech-packet-ninja Mar 30, 2026
54be43f
merge: resolve conflict with upstream batch generation refactor
deceptech-packet-ninja Apr 3, 2026
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
1 change: 1 addition & 0 deletions mlx_lm/generate.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,7 @@
CacheList,
KVCache,
QuantizedKVCache,
QuantizedRotatingKVCache,
RotatingKVCache,
load_prompt_cache,
)
Expand Down
20 changes: 17 additions & 3 deletions mlx_lm/models/base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,7 +69,15 @@ def quantized_scaled_dot_product_attention(
mask: Optional[mx.array],
group_size: int = 64,
bits: int = 8,
value_group_size: Optional[int] = None,
value_bits: Optional[int] = None,
) -> mx.array:
# Default value params to key params for backward compat
if value_group_size is None:
value_group_size = group_size
if value_bits is None:
value_bits = bits

B, n_q_heads, L, D = queries.shape
n_kv_heads = q_keys[0].shape[-3]
n_repeats = n_q_heads // n_kv_heads
Expand All @@ -96,7 +104,11 @@ def quantized_scaled_dot_product_attention(
scores += mask
scores = mx.softmax(scores, axis=-1, precise=True)
out = mx.quantized_matmul(
scores, *q_values, transpose=False, group_size=group_size, bits=bits
scores,
*q_values,
transpose=False,
group_size=value_group_size,
bits=value_bits,
)

if n_repeats > 1:
Expand All @@ -123,8 +135,10 @@ def scaled_dot_product_attention(
values,
scale=scale,
mask=mask,
group_size=cache.group_size,
bits=cache.bits,
group_size=getattr(cache, "key_group_size", cache.group_size),
bits=getattr(cache, "key_bits", cache.bits),
value_group_size=getattr(cache, "value_group_size", cache.group_size),
value_bits=getattr(cache, "value_bits", cache.bits),
)
else:
return mx.fast.scaled_dot_product_attention(
Expand Down
Loading