Skip to content

[BugFix] dynamic cache kv block_wise_fp8 not need create layer.cache_k_scale #5362

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

Merged
yuanlehome merged 1 commit into from
Dec 3, 2025
Merged

[BugFix] dynamic cache kv block_wise_fp8 not need create layer.cache_k_scale #5362

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
87 changes: 44 additions & 43 deletions fastdeploy/model_executor/layers/quantization/kv_cache.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -180,69 +180,70 @@ def create_weights(self, layer: nn.Layer, **extra_weight_attrs):
else:
raise NotImplementedError(f"{self.cache_quant_config.quant_type} is not implemented")

scale_shape = [layer.fd_config.model_config.num_key_value_heads]
if self.cache_quant_config.is_channel_wise:
scale_shape = [layer.kv_num_heads * layer.head_dim]

layer.cache_k_scale = layer.create_parameter(
shape=scale_shape,
dtype=paddle.get_default_dtype(),
default_initializer=paddle.nn.initializer.Constant(0),
)
layer.cache_v_scale = layer.create_parameter(
shape=scale_shape,
dtype=paddle.get_default_dtype(),
default_initializer=paddle.nn.initializer.Constant(0),
)

set_weight_attrs(
layer.cache_k_scale,
{
**extra_weight_attrs,
},
)
set_weight_attrs(
layer.cache_v_scale,
{
**extra_weight_attrs,
},
)
if "block_wise" not in layer.cache_quant_type_str: # dynamic cache kv block_wise_fp8 not need
scale_shape = [layer.fd_config.model_config.num_key_value_heads]
if self.cache_quant_config.is_channel_wise:
scale_shape = [layer.kv_num_heads * layer.head_dim]

layer.cache_k_out_scale = layer.create_parameter(
shape=scale_shape,
dtype=paddle.get_default_dtype(),
default_initializer=paddle.nn.initializer.Constant(0),
)
layer.cache_v_out_scale = layer.create_parameter(
shape=scale_shape,
dtype=paddle.get_default_dtype(),
default_initializer=paddle.nn.initializer.Constant(0),
)

if self.cache_quant_config.has_zero_point:
layer.cache_k_zp = layer.create_parameter(
layer.cache_k_scale = layer.create_parameter(
shape=scale_shape,
dtype=paddle.get_default_dtype(),
default_initializer=paddle.nn.initializer.Constant(0),
)
layer.cache_v_zp = layer.create_parameter(
layer.cache_v_scale = layer.create_parameter(
shape=scale_shape,
dtype=paddle.get_default_dtype(),
default_initializer=paddle.nn.initializer.Constant(0),
)

set_weight_attrs(
layer.cache_k_zp,
layer.cache_k_scale,
{
**extra_weight_attrs,
},
)
set_weight_attrs(
layer.cache_v_zp,
layer.cache_v_scale,
{
**extra_weight_attrs,
},
)

layer.cache_k_out_scale = layer.create_parameter(
shape=scale_shape,
dtype=paddle.get_default_dtype(),
default_initializer=paddle.nn.initializer.Constant(0),
)
layer.cache_v_out_scale = layer.create_parameter(
shape=scale_shape,
dtype=paddle.get_default_dtype(),
default_initializer=paddle.nn.initializer.Constant(0),
)

if self.cache_quant_config.has_zero_point:
layer.cache_k_zp = layer.create_parameter(
shape=scale_shape,
dtype=paddle.get_default_dtype(),
default_initializer=paddle.nn.initializer.Constant(0),
)
layer.cache_v_zp = layer.create_parameter(
shape=scale_shape,
dtype=paddle.get_default_dtype(),
default_initializer=paddle.nn.initializer.Constant(0),
)
set_weight_attrs(
layer.cache_k_zp,
{
**extra_weight_attrs,
},
)
set_weight_attrs(
layer.cache_v_zp,
{
**extra_weight_attrs,
},
)

def process_loaded_weights(self, layer: nn.Layer, state_dict):
"""
use for loader v0
Expand Down
Loading