apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 08:37:06 +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

16
fastdeploy/model_executor/layers/quantization/block_wise_fp8.py
View File

@@ -78,8 +78,8 @@ class BlockWiseFP8LinearMethod(QuantMethodBase):
self.quant_config = quant_config
def create_weights(self, layer):
layer.linear_weight_shape.reverse()
layer.linear_weight_scale = layer.create_parameter(
layer.weight_shape.reverse()
layer.weight_scale = layer.create_parameter(
shape=[
(layer.output_size + self.quant_config.weight_block_size[0] -
1) // self.quant_config.weight_block_size[0],
@@ -95,8 +95,8 @@ class BlockWiseFP8LinearMethod(QuantMethodBase):
weight_tensor = weights.transpose([1, 0])
quanted_weight_tensor, weight_block_scale_tensor = (
per_block_cast_to_fp8(weight_tensor))
layer.linear_weight.copy_(quanted_weight_tensor, False)
layer.linear_weight_scale.set_value(weight_block_scale_tensor)
layer.weight.copy_(quanted_weight_tensor, False)
layer.weight_scale.set_value(weight_block_scale_tensor)
def process_prequanted_weights(self, layer, state_dict):
"""
@@ -106,10 +106,10 @@ class BlockWiseFP8LinearMethod(QuantMethodBase):
weight_scale = get_tensor(state_dict.pop(layer.weight_scale_key))
quant_weight = quant_weight.transpose([1, 0]).contiguous()
layer.linear_weight.copy_(quant_weight.view("float8_e4m3fn"), False)
layer.weight.copy_(quant_weight.view("float8_e4m3fn"), False)
weight_scale = weight_scale.transpose([1, 0])
layer.linear_weight_scale.set_value(weight_scale)
layer.weight_scale.set_value(weight_scale)
def apply(self, layer, x):
x, x_scale_tensor = fastdeploy.model_executor.ops.gpu.per_token_quant_padding(
@@ -119,9 +119,9 @@ class BlockWiseFP8LinearMethod(QuantMethodBase):
import fastdeploy.model_executor.ops.gpu.deep_gemm as deep_gemm
deep_gemm.gemm_fp8_fp8_bf16_nt(
(x, x_scale_tensor),
(layer.linear_weight, layer.linear_weight_scale),
(layer.weight, layer.weight_scale),
linear_out,
)
if layer.with_bias:
linear_out = paddle.add(linear_out, layer.linear_bias)
linear_out = paddle.add(linear_out, layer.bias)
return linear_out