6ed1b99aaa
refactor(attention): move into a package
...
modules/attention.py becomes modules/attention/: hijacks.py keeps the six
sdpa monkeypatch setters, dispatcher.py the diffusers-side processor and
dispatcher setup with the kernels hub hijack, and the package facade
re-exports every public name so call sites are unchanged. The devices
import moves inside set_diffusers_attention, which removes the
devices <-> attention import cycle.
2026-08-23 08:20:45 +01:00
0218a0aadc
Add FP16 Accumulation option to SDNQ Atten
2026-08-18 20:44:05 +03:00
76941efc7e
strict typechecking
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2026-08-16 12:45:50 +02:00
712a13f1a0
Move SDNQ to upstream SDNQ repo
2026-08-10 12:13:10 +03:00
a1df374773
add support for nunchaku-lite models and engine
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2026-08-04 15:31:05 +02:00
177f4c417c
Add SDNQ_USE_TRITON_SCALED_MM
2026-07-22 23:11:57 +03:00
9b7a8a3408
skip reapply attention
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2026-07-22 20:12:07 +02:00
a2f611df6c
correct attention reapply
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2026-07-22 19:47:51 +02:00
75926e6fc4
Remove sdnq_use_quantized_matmul and use sdnq_quantize_matmul_mode instead
2026-07-15 18:45:30 +03:00
d8a29a68ef
Update IPEX and ROCm to Torch 2.13
2026-07-15 04:05:26 +03:00
1cf2c9012d
add min block size check to sdnq atten
2026-07-15 03:20:43 +03:00
dc7e3929cd
SDNQ Atten skip TE
2026-07-13 04:49:55 +03:00
e1ac9650b7
Make SDNQ Triton kernels trigger way less tuning and fix accuracy issues on SDNQ Atten
2026-07-01 17:31:57 +03:00
d3e3948787
update lens and bria
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2026-06-29 09:19:29 +02:00
e89f1a1fe8
SDNQ add causal atten support
2026-06-28 20:08:40 +03:00
6ee13520a1
SDNQ atten remove query.device == key.device checks
2026-06-27 16:23:49 +03:00
2622291de7
SDNQ Atten Fix FP16 PV
2026-06-27 15:39:58 +03:00
3fed183564
update sdnq atten min sizes
2026-06-27 14:51:56 +03:00
718e24c50d
increase the min lenghts for SDNQ atten
2026-06-27 02:54:39 +03:00
8029156fa2
SDNQ Atten use simple row-wise quant instead
2026-06-24 17:59:38 +03:00
0af57a0af1
And no quant atten
2026-06-23 20:41:19 +03:00
72a4d2c2bd
Add Hadamard support to SDNQ Atten
2026-06-23 14:17:34 +03:00
dfa713e6fc
Add SDNQ attention
2026-06-23 08:10:02 +03:00
5e99dee3c2
full codespell coverage
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2026-06-04 12:36:10 +02:00
a0531c70fd
add new attention dispatcher
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2026-05-29 21:10:56 +02:00
d2d9f7caea
captioning improvements and cleanup
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2026-05-29 21:10:56 +02:00
a56d1d299e
add wildcards to image metadata and note separate template from prompt field
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2026-05-19 09:45:20 +02:00
68f00d85ca
attention execution guard against cpu tensors
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2026-05-17 08:44:29 +02:00
375591ecdd
Revert non-critical changes
...
The following errors are re-introduced and will need follow-up:
- models.py, line 92
- sd_unet.py, line 41
- sd_models.py, line 1440
2026-03-25 13:47:27 -07:00
de86927c1b
Remove unneeded type and default
2026-03-23 22:54:40 -07:00
fb2f9ea650
new server info panel
...
Signed-off-by: vladmandic <mandic00@live.com >
2026-03-18 15:37:57 +01:00
d65a2d1ebc
ruff lint
2026-02-19 11:13:44 +01:00
e5c494f999
cleanup logger
2026-02-19 11:09:13 +01:00
a3074baf8b
unified logger
2026-02-19 09:46:42 +01:00
bfe014f5da
modernize typing
2026-02-19 09:15:37 +01:00
49965dfda8
get_hip_arch_name -> get_hip_agent, use amdhip64_7.dll served within rocm package
2026-01-03 21:00:36 +09:00
c536637908
add kanvas integration
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-11-09 15:41:20 -05:00
113c582171
IPEX fix early init failure
2025-11-09 00:30:03 +03:00
f4ee9c7052
Add Flex attention
2025-11-09 00:14:38 +03:00
2bbbb684cc
Rename CK Flash attention to just Flash attention
2025-11-08 23:24:40 +03:00
a93715e0da
Don't expose AMD Triton Flash Atten for non AMD
2025-11-08 23:20:55 +03:00
56026c4e61
refactor attention handling
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-11-08 10:55:41 -05:00
155ee7f84c
fix sage-attention checks on sm86
...
Signed-off-by: Vladimir Mandic <mandic00@live.com >
2025-11-08 08:47:21 -05:00
CalamitousFelicitousness
Dity0
Vladimir Mandic
awsr
Seunghoon Lee