bdc477d252
Refactor GPU backend selection for sage attention
...
Removed hot path, now everything is defined at setup
Also, passing device to get_device_capability so that it works properly with multi-gpu setups.
2025-11-05 16:31:13 +00:00
4c791fb795
Remove model check logic for SA2 workaround
2025-11-05 10:53:07 +00:00
18676996d0
Sage Attention 2 + Triton workaround Qwen-Image
...
Workaround to prevent black images generated with Qwen-Image models when Sage Attention 2 is enabled with Triton as backend on devices with compute capability 8.0 and 8.6.
Simply switches back to Cuda backend for these models only.
Proof of concept, feel free to close if this is not appropriate.
2025-11-04 23:31:14 +00:00
780cd26587
triton test hide errors behind debug flag
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-11-04 07:26:21 -05:00
495cfd8632
fix cn
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-11-01 12:21:19 -04:00
58f218a560
add cudnn enable/disable override
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-11-01 11:33:39 -04:00
408b82ef08
cleanup
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-10-29 09:44:18 -04:00
46876060ab
kandinsky 10s force flex attn
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-10-28 10:53:19 -04:00
4282762c8c
fix late init
...
Signed-off-by: vladmandic <mandic00@live.com >
2025-10-26 18:57:51 -04:00
60ac82b191
add basic xpu gpu monitor
...
Signed-off-by: vladmandic <mandic00@live.com >
2025-10-26 18:55:54 -04:00
0271e0830c
triton split check into early and full
...
Signed-off-by: vladmandic <mandic00@live.com >
2025-10-26 11:48:10 -04:00
818b0c0821
Add basic triton test
2025-10-26 10:44:04 +03:00
4b95d72d45
video tab layout
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-10-18 14:07:52 -04:00
1ebd96fdc6
add kandinsky5-lite t2v
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-10-18 12:15:36 -04:00
4a70e82b0c
ROCm always use numpy on cholesky
2025-09-28 14:05:20 +03:00
a47959b114
move ROCm Windows hijacks outside of torch install
2025-09-28 13:33:15 +03:00
6766563510
Add info log for Building CK Flash attention
2025-09-10 19:25:07 +03:00
bc7c89c070
add typing to sdpa hijacks
2025-09-10 04:43:57 +03:00
c51552af90
Enable triton flash atten option for rocm linux too
2025-09-03 16:26:00 +03:00
c42e0e0b37
Cleanup
2025-09-03 16:18:02 +03:00
266c9c0d3d
Move Zluda Triton flash atten hijack to Triton Flash attention option
2025-09-03 16:16:41 +03:00
fa44521ea3
offload-never and offload-always per-module and new highvram profile
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-07-31 11:40:24 -04:00
04af23a3bc
refactore pipeline apply/unapply optional components & features
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-07-26 20:04:07 -04:00
ad716b118b
fix enable_gqa with dyn atten
2025-07-26 01:49:17 +03:00
a5b77b8ee2
remove dead code
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-07-05 16:47:25 -04:00
c9f49720c5
Cleanup
2025-06-25 23:37:41 +03:00
dd84fb541f
Always set sdpa params
2025-06-11 21:43:48 +03:00
7679028c1a
Override CPU to use FP32 by default
2025-06-06 15:33:51 +03:00
299d189276
When using the openvino backend, do not look for an nvidia gpu.
2025-05-12 19:14:26 -04:00
b0e5a6c4df
Add devices.has_triton() and enable NNCF compile if triton is available
2025-05-09 22:24:36 +03:00
dfebc909eb
Disable cuDNN benchmark on ROCm and add cudnn_benchmark_limit option
2025-05-08 13:27:06 +03:00
90f887ac4a
Add dim checks to ck flash atten and fix dim check on dyn atten
2025-03-25 03:50:21 +03:00
0c890b50e0
proper zluda detection
2025-03-20 23:03:23 +09:00
1e0f512ccb
ROCm disable FP16 for gfx1102
2025-03-19 15:42:36 +03:00
878cab085f
Reverse the sdpa hijcak order
2025-02-14 19:56:39 +03:00
f94196bcd1
Rename ROCm Flash atten hijack to CK Flash atten and enable AOTriton memory and flash atten by default
2025-02-13 22:01:06 +03:00
49712ab9e7
update requirements
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-02-13 12:03:08 -05:00
e28b8cd920
add torch gc debug
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-01-31 13:12:59 -05:00
d2fee97e24
Update changelog
2025-01-31 20:27:22 +03:00
039746914f
Add check for missing cuda and ipex params
2025-01-31 19:12:56 +03:00
3dcb70e8a2
device init logging
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-01-31 09:31:12 -05:00
1697fb1508
add tunable ops path
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-01-30 13:35:09 -05:00
0ea7840608
add tunable ops
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-01-30 13:08:49 -05:00
a770b1c888
More correct Dynamic Atten SDPA implementation and deprecate IPEX Diffusers attention
2025-01-25 21:33:42 +03:00
af35296a68
IPEX 4GB alloc detection and log driver version
2025-01-22 18:15:25 +03:00
bb97e695da
log cleanup
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2024-12-31 09:15:36 -05:00
910f5d0a73
lora direct on-demand apply/unapply
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2024-12-29 12:38:19 -05:00
20f2554cec
add sd35-ipadapter and more balanced offload optimizations
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2024-12-20 10:22:42 -05:00
e9f951b2c5
offload logging
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2024-12-11 14:20:01 -05:00
9a588d9c91
update balanced offload
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2024-12-11 12:06:03 -05:00
CalamitousFelicitousness
Vladimir Mandic
Disty0
chrismuzyn
Seunghoon Lee