53 Commits

Author	SHA1	Message	Date
Vladimir Mandic	e972141917	linting ... Signed-off-by: Vladimir Mandic <mandic00@live.com>	2025-10-18 08:33:58 -04:00
Vladimir Mandic	c8ca5cd75c	load rocm.py only when needed ... Signed-off-by: Vladimir Mandic <mandic00@live.com>	2025-10-18 08:31:28 -04:00
Seunghoon Lee	a36916616f	windows catch rocm detection failure	2025-10-18 12:00:30 +09:00
Seunghoon Lee	552c223569	use driver library, more checks for windows rocm	2025-10-18 01:17:30 +09:00
Disty0	6b67a9d0c4	SDNQ add check_mats to matmul	2025-09-30 01:58:13 +03:00
Vladimir Mandic	c661497b87	linting ... Signed-off-by: Vladimir Mandic <mandic00@live.com>	2025-09-28 18:09:09 -04:00
Disty0	4a70e82b0c	ROCm always use numpy on cholesky	2025-09-28 14:05:20 +03:00
Disty0	a47959b114	move ROCm Windows hijacks outside of torch install	2025-09-28 13:33:15 +03:00
Seunghoon Lee	06b14b070b	hijack torch.linalg.cholesky()	2025-09-28 18:43:37 +09:00
Seunghoon Lee	9c19e8a9b8	return correct item	2025-09-28 18:34:10 +09:00
Seunghoon Lee	6e8abf8dc3	hijack cholesky decomposition for therock pytorch	2025-09-28 18:03:02 +09:00
Seunghoon Lee	0de3732884	prioritize python pacakage check over PATH	2025-09-28 11:31:43 +09:00
Seunghoon Lee	35b1c6b07d	windows install rocm if not installed	2025-09-28 05:39:41 +09:00
Seunghoon Lee	059392afc7	clean up	2025-09-27 17:51:59 +09:00
Disty0	503a178794	Don't load rocm hsa on non wsl envs and remove unused lib hijacks	2025-09-27 11:39:24 +03:00
Seunghoon Lee	579b1f3175	do not load global amdhip64, let pytorch load it	2025-09-27 17:01:14 +09:00
Seunghoon Lee	d52d84eda5	clean up	2025-09-27 16:49:29 +09:00
Luca Beltrame	f00486c8f3	Fix HIP library detection ... The code unconditionally checked `lib`, but on modern Linux distributions, 64-bit binaries are under `lib64` instead. Note that this might be slightly different on Debian-based distributions, but as I don't have one I can't test it.	2025-09-27 09:23:32 +02:00
Seunghoon Lee	e3e41298d9	remove unused import	2025-09-27 11:29:55 +09:00
Seunghoon Lee	8d2aedd924	hotfix	2025-09-27 11:17:15 +09:00
Seunghoon Lee	d63bc05b6c	update rocm.py to detect rocm-sdk packages	2025-09-27 11:12:13 +09:00
Disty0	4b76049aed	fix basicsr and gfpgan	2025-09-10 19:13:08 +03:00
Disty0	266c9c0d3d	Move Zluda Triton flash atten hijack to Triton Flash attention option	2025-09-03 16:16:41 +03:00
Disty0	d99777616b	ROCm fix Segfault and remove the conflicting hipblaslt override	2025-08-29 19:03:49 +03:00
Disty0	74b6edf2df	revert gfx1101	2025-06-11 19:25:05 +03:00
Disty0	6aa5c08fb0	Cleanup and update changelog	2025-06-11 16:03:33 +03:00
Disty0	71be3c7d45	ROCm don't override gfx with gfx1100 and gfx1101 + rocm 6.4	2025-06-11 15:47:25 +03:00
Disty0	df6b13ea47	Don't set gfx override with RX 9000 and above	2025-06-11 15:09:03 +03:00
Disty0	bd2d9d1677	Python 3.13 support	2025-06-09 22:58:08 +03:00
Disty0	748bdbf437	Update rocm flash attention repo with navi rotary fix	2025-04-08 00:15:57 +03:00
Seunghoon Lee	0eaa2c0378	rocm wsl better arch detection	2025-04-01 21:14:02 +09:00
Disty0	dc551bfd85	Set blaslt_tensile_libpath to empty string instead of None	2025-02-15 18:04:42 +03:00
Seunghoon Lee	13e9f4f3e1	flash attn triton is already merged into main	2025-02-14 23:21:56 +09:00
Disty0	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
Seunghoon Lee	389bb78b8e	Fix bug.	2025-01-11 23:41:17 +09:00
Seunghoon Lee	e43d8e9448	check hipblaslt availability in windows	2025-01-11 23:36:51 +09:00
Seunghoon Lee	d597d5912d	use bitmasking for agent detection	2024-10-26 13:40:54 +09:00
Vladimir Mandic	e0d702a3dd	add gfx autodetect options ... Signed-off-by: Vladimir Mandic <mandic00@live.com>	2024-10-07 11:38:10 -04:00
Seunghoon Lee	470ea10e95	fix hipinfo	2024-09-27 10:24:55 +09:00
Seunghoon Lee	b74af75622	add more known devices	2024-09-27 09:42:13 +09:00
Seunghoon Lee	96922d70a7	rocm&zluda handle apu	2024-09-26 13:34:11 +09:00
Seunghoon Lee	1395f5bf9e	add triton backend flash-attn (experimental)	2024-09-24 12:53:18 +09:00
Seunghoon Lee	e246e55734	rocm install flash-attn if needed	2024-09-24 12:27:43 +09:00
Seunghoon Lee	9c4213e4a8	tcmalloc experiment	2024-08-09 16:11:12 +09:00
Seunghoon Lee	864263f570	accurate wsl check	2024-08-09 10:38:28 +09:00
Seunghoon Lee	c490b9cb51	fix first launch	2024-08-09 00:45:12 +09:00
Seunghoon Lee	61961f5d6d	hipblaslt check torch version	2024-08-09 00:35:46 +09:00
Seunghoon Lee	4492dedf16	fix hip path detection	2024-07-28 00:55:40 +09:00
Seunghoon Lee	bdf6501f40	fix	2024-07-28 00:34:54 +09:00
Seunghoon Lee	5a9474180d	rocm.py	2024-07-28 00:00:19 +09:00