apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-06 09:07:10 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

refactor rl get_name_mappings_to_training (#2847)
Some checks failed
Deploy GitHub Pages / deploy (push) Has been cancelled
Details

Browse Source 
* refactor rl get_name_mappings_to_training

* fix tp>1

* change variable name(ffn1->up_gate_proj/ffn2->down_proj)

* change variable name(linear_weight->weight/linear_bias->bias)

* add rl names mapping for vl

* fix ernie 0.3B error

* fix develop code

* fix
This commit is contained in:
Yuanle Liu
2025-07-15 22:31:42 +08:00
committed by GitHub
parent e7bcbbab52
commit 61b3997b85
47 changed files with 1591 additions and 1629 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
fastdeploy/model_executor/layers/backends/gcu/quantization/weight_only.py
View File

@@ -38,14 +38,14 @@ class GCUWeightOnlyLinearMethod(WeightOnlyLinearMethod):
def create_weights(self, layer):
# The scale shape should be equal to the output dim of weight using Per-Channel Quantization.
linear_weight_scale_shape = [layer.linear_weight_shape[1]]
weight_scale_shape = [layer.weight_shape[1]]
layer.linear_weight_shape.reverse()
layer.weight_shape.reverse()
if self.quant_config.name() == "wint4":
layer.linear_weight_shape[0] //= 2
layer.weight_shape[0] //= 2
layer.weight_dtype = "int8"
layer.linear_weight_scale = layer.create_parameter(
shape=linear_weight_scale_shape,
layer.weight_scale = layer.create_parameter(
shape=weight_scale_shape,
dtype=layer._dtype,
is_bias=False,
)
@@ -61,8 +61,8 @@ class GCUWeightOnlyLinearMethod(WeightOnlyLinearMethod):
"""
quant_weight = get_tensor(state_dict.pop(layer.weight_key))
weight_scale = get_tensor(state_dict.pop(layer.weight_scale_key))
layer.linear_weight.set_value(quant_weight)
layer.linear_weight_scale.set_value(
layer.weight.set_value(quant_weight)
layer.weight_scale.set_value(
weight_scale.astype(paddle.get_default_dtype()))
@@ -73,8 +73,8 @@ class GCUWeightOnlyLinearMethod(WeightOnlyLinearMethod):
self.group_size, # group_size
)
layer.linear_weight.set_value(quanted_weight_tensor)
layer.linear_weight_scale.set_value(
layer.weight.set_value(quanted_weight_tensor)
layer.weight_scale.set_value(
weight_scale_tensor.astype(paddle.get_default_dtype()))
@@ -82,8 +82,8 @@ class GCUWeightOnlyLinearMethod(WeightOnlyLinearMethod):
def apply(self, layer, x):
linear_out = linear_quant(
lhs=x,
rhs=layer.linear_weight,
scale=layer.linear_weight_scale,
rhs=layer.weight,
scale=layer.weight_scale,
bias=None,
group_size=self.group_size,
)