33 Commits

Author	SHA1	Message	Date
vladmandic	a4671045b6	lint and crlf ... Signed-off-by: vladmandic <mandic00@live.com>	2026-01-24 10:28:46 +01:00
Disty0	8d6bfcd827	Update SDNQ	2026-01-23 14:39:07 +03:00
Disty0	784cda80aa	update sdnq	2026-01-14 16:23:26 +03:00
Disty0	db59d2b507	SDNQ handle packed floats in fp mm	2025-12-27 16:29:18 +03:00
Disty0	949ff04577	SDNQ fix fp16 mm with fp8 weights and improve stochastic rounding performance	2025-12-09 17:41:29 +03:00
Disty0	aaef4992c3	SDNQ fix svd + fp8 tw and fp16 mm	2025-11-28 22:31:09 +03:00
Disty0	b6e9332cfe	SDNQ de-couple matmul dtype and add fp16 matmul	2025-11-22 02:16:20 +03:00
Disty0	1745ed53f8	Refactor SDNQDequantizer	2025-11-18 01:42:58 +03:00
Disty0	6f33ec3357	SDNQ use the model quant params instead of user settings on Lora	2025-11-10 00:12:38 +03:00
Disty0	f12caf81f9	SDNQ skip bad layers on svd and fix svd with dequantize_fp32	2025-10-17 17:25:50 +03:00
Disty0	c7aba8589b	SDNQ fix Qwen loading	2025-10-11 00:05:09 +03:00
Disty0	be91bbff75	SDNQ add SVD support for Convs	2025-10-06 18:26:42 +03:00
Disty0	9e52d0c1fb	SDNQ add SVDQuant quantization method	2025-10-05 22:50:30 +03:00
Disty0	99113947bf	SDNQ add RDNA2 INT8 support via Triton	2025-10-04 18:31:25 +03:00
Disty0	54acf1760b	Make SDNQ scales compatible with balanced offload	2025-10-03 18:13:55 +03:00
Disty0	c5cab96223	SDNQ simplify check_mats	2025-10-03 02:58:17 +03:00
Disty0	03382bdd4c	SDNQ simplify check_mats	2025-10-01 01:35:51 +03:00
Disty0	0c1d34721c	SDNQ use contiguous for intel	2025-09-30 02:37:58 +03:00
Disty0	6b67a9d0c4	SDNQ add check_mats to matmul	2025-09-30 01:58:13 +03:00
Disty0	e6715ba8d3	Cleanup SDNQ compile	2025-09-19 19:29:36 +03:00
Disty0	a12edc1e90	SDNQ use nan_to_num_ with fp8 quantization in case of zeros	2025-09-15 20:22:39 +03:00
Vladimir Mandic	9743c8e4bf	keep previous processed state ... Signed-off-by: Vladimir Mandic <mandic00@live.com>	2025-08-31 15:20:15 -04:00
Disty0	bbb345cf44	Fix bias dtype mismatch	2025-08-30 02:31:41 +03:00
Disty0	6c36433a14	SDNQ fix row-wise FP8 matmul with fp32 and fp16 inputs	2025-08-30 02:27:15 +03:00
Disty0	a8de3f7282	SDNQ add quantized matmul support for all quantization types and group sizes	2025-08-29 22:26:47 +03:00
Disty0	f324b7c0e5	SDNQ remove unnecessary .contiguous()	2025-08-21 02:21:05 +03:00
Disty0	8460be662c	SDNQ use inplace transpose and use view instead of reshape	2025-08-17 05:07:55 +03:00
Disty0	9992338187	sdnq fix convs	2025-08-11 23:24:13 +03:00
Disty0	26461f1d8d	fix conv in8 matmul	2025-08-11 23:15:30 +03:00
Disty0	dc7b25d387	Cleanup SDNQ and add SDNQ_USE_TENSORWISE_FP8_MATMUL env var	2025-08-11 14:50:17 +03:00
Disty0	3f45c4e570	Cleanup SDNQ and skip transpose on packed int8 matmul	2025-08-10 19:31:34 +03:00
Disty0	22d86acda3	Make SDNQ MatMul listen to the dequantize fp32 setting	2025-08-09 01:10:07 +03:00
Disty0	c3d007b02c	SDNQ split forward.py into layers and cleanup	2025-08-02 17:36:55 +03:00