Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-18 16:54:33 +02:00
Code Issues Packages Projects Releases Wiki Activity

Override CPU to use FP32 by default

Browse Source
This commit is contained in:
Disty0
2025-06-06 15:33:51 +03:00
parent 2ccc76ab91
commit 7679028c1a
2 changed files with 13 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CHANGELOG.md
+11 -7
View File
@@ -2,15 +2,19 @@
## Update for 2025-06-06
- **Torch**:
- set default to `torch==2.7.1`
- **Changes**
- Increase the medvram mode threshold from 8GB to 12GB
- Set CPU backend to use FP32 by default
- **Torch**
- set default to `torch==2.7.1`
- **SDNQ Quantization**
- Add group size support for convolutional layers
- Add quantized matmul support for for convolutional layers
- Fix forced FP32 with tensorwise FP8 matmul
- Fix PyTorch <= 2.4 compatibility with FP8 matmul
- Fix VAE with conv quant
- Add group size support for convolutional layers
- Add quantized matmul support for for convolutional layers
- Fix forced FP32 with tensorwise FP8 matmul
- Fix PyTorch <= 2.4 compatibility with FP8 matmul
- Fix VAE with conv quant
## Update for 2025-06-02
modules/devices.py
+2 -2
View File
@@ -331,7 +331,7 @@ def test_fp16():
if fp16_ok is not None:
return fp16_ok
if opts.cuda_dtype != 'FP16': # don't override if the user sets it
if sys.platform == "darwin" or backend == 'openvino': # override
if sys.platform == "darwin" or backend in {'openvino', 'cpu'}: # override
fp16_ok = False
return fp16_ok
elif backend == 'rocm':
@@ -362,7 +362,7 @@ def test_bf16():
if bf16_ok is not None:
return bf16_ok
if opts.cuda_dtype != 'BF16': # don't override if the user sets it
if sys.platform == "darwin" or backend == 'openvino' or backend == 'directml': # override
if sys.platform == "darwin" or backend in {'openvino', 'directml', 'cpu'}: # override
bf16_ok = False
return bf16_ok
elif backend == 'rocm' or backend == 'zluda':