apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 08:21:53 +08:00
Code Issues Actions 19 Packages Projects Releases Wiki Activity

[RL] Support SM100 FP8 quantization in RL (#6601)

Browse Source 
* RL SM100 Fix

* update
This commit is contained in:
bukejiyu
2026-03-04 20:55:04 +08:00
committed by GitHub
parent 1256fd3806
commit 598cce8545
3 changed files with 34 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/model_executor/layers/quantization/block_wise_fp8.py
+6 -2
View File
@@ -281,8 +281,12 @@ class BlockWiseFP8LinearMethod(QuantMethodBase):
if self.model_format != "torch":
process_weight_transpose(layer, "weight")
process_weight_transpose(layer, "weight_scale_inv")
else:
return
if self.quant_config.deepgemm_scale_ue8m0:
new_weight_scale_inv = paddle.empty(
layer.weight_scale_inv.shape[::-1], dtype=layer.weight_scale_inv.dtype
)
new_weight_scale_inv = new_weight_scale_inv.transpose([1, 0])
layer.weight_scale_inv.data = new_weight_scale_inv
def process_loaded_weights(self, layer, weights) -> None:
weight_tensor = weights.transpose([1, 0])