apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 00:33:03 +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

12
fastdeploy/model_executor/layers/quantization/w8a8.py
View File

@@ -69,7 +69,7 @@ class W8A8LinearMethod(QuantMethodBase):
self.smooth_quant_method = SmoothQuantLinearMethod(quant_config)
def create_weights(self, layer):
layer.linear_weight_shape.reverse()
layer.weight_shape.reverse()
layer.weight_dtype = "int8"
if self.quant_config.use_smooth_quant:
self.smooth_quant_method.create_weights(layer)
@@ -101,21 +101,21 @@ class W8A8LinearMethod(QuantMethodBase):
if self.skip_quant:
logger.debug(f"{layer.prefix} skip quant")
weight_tensor = weights.cast(layer._dtype)
layer.linear_weight.set_value(weight_tensor)
layer.weight.set_value(weight_tensor)
else:
weight_tensor = weights.transpose([1, 0])
weight_tensor = paddle.cast(weight_tensor, "int8")
layer.linear_weight.set_value(weight_tensor)
layer.weight.set_value(weight_tensor)
def apply(self, layer, x):
if self.skip_quant:
linear_out = paddle.matmul(x, layer.linear_weight, False, True)
linear_out = paddle.matmul(x, layer.weight, False, True)
return linear_out
if self.quant_config.use_gemm_dequant:
linear_out = fastdeploy.model_executor.ops.gpu.gemm_dequant(
x, layer.linear_weight, layer.linear_out_scale, layer._dtype)
x, layer.weight, layer.linear_out_scale, layer._dtype)
else:
linear_out = paddle.matmul(x, layer.linear_weight, False, True)
linear_out = paddle.matmul(x, layer.weight, False, True)
linear_out = fastdeploy.model_executor.ops.gpu.dequant_int8(
linear_out, layer.linear_out_scale, layer._dtype)
return linear_out