Commit Graph

329 Commits

Author SHA1 Message Date
Disty0 e484850e9e Cleanup SDNQ 2026-05-22 17:32:41 +03:00
Disty0 f91893d0df Fix bias dtype mismatch with pre-quants 2026-05-22 00:08:42 +03:00
Disty0 3e879ccb69 cleanup 2026-05-21 23:44:44 +03:00
Disty0 cfafbed868 don't duplicate modules_to_not_convert 2026-05-21 20:44:28 +03:00
Disty0 9adf20a47c fix asym quant with hadamard 2026-05-21 17:22:45 +03:00
Disty0 40bd5ca00d Make Triton MM block sizes configurable 2026-05-20 19:16:21 +03:00
Disty0 4341193b35 SDNQ trigger triton_mm autotune for every multiple of 64 instead of 1 2026-05-20 15:36:12 +03:00
Disty0 cad5ab19c8 SDNQ use torch.compile friendly hadamard cache 2026-05-19 22:18:58 +03:00
Disty0 46f37232b3 Fix hadamard not registering 2026-05-18 23:54:23 +03:00
Disty0 a896719713 Fix dynamic quant with svd + hadamard combination 2026-05-18 23:31:18 +03:00
Disty0 069a1c927d Fix dynamic quant 2026-05-18 22:44:50 +03:00
Disty0 927994d47d gate recompile_limit behind torch version 2026-05-18 19:28:05 +03:00
Disty0 65b9c8803f SDNQ use cache for hadamard matrix 2026-05-18 18:20:32 +03:00
Disty0 ac31b897b9 PyTorch ignores the global re_compile limit now 2026-05-18 17:53:27 +03:00
Disty0 e3b1eb71a1 cleanup 2026-05-18 17:10:31 +03:00
Disty0 c227d5fa10 SDNQ add Hadamard Rotation support 2026-05-18 17:06:11 +03:00
Disty0 58891d3a4e cleanup comments 2026-05-15 14:36:03 +03:00
Vladimir Mandic 8ca0568bd2 fix apply_sdnq_to_module
Signed-off-by: Vladimir Mandic <mandic00@live.com>
2026-05-15 12:17:01 +02:00
Disty0 6eebf2566d Refactor SDNQ quantizer handling and add modules_to_not_use_matmul 2026-05-15 05:47:32 +03:00
Vladimir Mandic 7194b4b553 add anima and hidream-o1 prequants
Signed-off-by: Vladimir Mandic <mandic00@live.com>
2026-05-13 10:15:37 +02:00
Vladimir Mandic c8d6fd5cf8 Revert "much stricter ruff linting"
This reverts commit 310dbf1574.
2026-05-11 08:13:57 +02:00
Vladimir Mandic 310dbf1574 much stricter ruff linting
Co-authored-by: Copilot <copilot@github.com>
Signed-off-by: Vladimir Mandic <mandic00@live.com>
2026-05-11 07:15:37 +02:00
CalamitousFelicitousness 543fd35484 feat(hidream): enable sdnq quantization for o1
- Add module_skip_keys entry for HiDreamO1Qwen3VLTransformer so SDNQ
  leaves embedding/output layers (lm_head, embed_tokens, x_embedder,
  t_embedder1, final_layer2, patch_embed, pos_embed) at full precision.
- Flip allow_quant to True in load_hidream_o1 and surface the resolved
  quant type in the load log.
2026-05-11 03:57:58 +01:00
Vladimir Mandic 6261c68396 add Step1X-Edit
Co-authored-by: Copilot <copilot@github.com>
Signed-off-by: Vladimir Mandic <mandic00@live.com>
2026-05-05 09:31:05 +02:00
Disty0 cc0f06d64f SNDQ unify Triton MM configs 2026-05-04 15:14:30 +03:00
awsr e9d74cea72 Fix typing 2026-04-23 16:49:50 -07:00
resonantsky cefad871b9 fix rdna2 variable name: is_rdna2 -> is_rdna2_and_older 2026-04-17 18:44:24 +02:00
resonantsky 576d566429 established rdna2 best configs 2026-04-17 16:46:41 +02:00
Disty0 33726b5019 Update triton mm config 2026-04-16 20:51:10 +03:00
Disty0 eba2c3a8a3 SDNQ update RDNA2 detection 2026-04-16 08:34:17 +03:00
Disty0 85ec9bf137 SDNQ fix post load quant 2026-04-16 00:38:54 +03:00
Disty0 ed5019116b cleanup 2026-04-16 00:25:21 +03:00
Disty0 9960b371d1 cleanup 2026-04-16 00:22:52 +03:00
Disty0 68a5fe6342 SDNQ add fp8 mm info to modules_quant_config 2026-04-16 00:18:42 +03:00
Disty0 8a1743712a SDNQ disable FP8 MM on dynamic quant with unsupported GPUs 2026-04-15 23:28:20 +03:00
Disty0 202f12ea6c SDNQ disable fp8 mm with pre-quants on unsupported gpus 2026-04-15 23:12:46 +03:00
vladmandic 84c2ddc5ca linting
Signed-off-by: vladmandic <mandic00@live.com>
2026-04-13 12:18:05 +02:00
Disty0 a16ff95f18 SDNQ add nn.Embedding quantization support and add Gemma4 keys 2026-04-10 18:46:58 +03:00
Disty0 b2e071dc52 cleanup 2026-04-04 01:39:26 +03:00
Disty0 470a0d816e SDNQ add tensor descriptor kernel to triton mm for Intel Arc 2026-04-04 01:32:34 +03:00
Disty0 06fe4c7f20 SDNQ cleanup triton_mm 2026-03-29 00:44:58 +03:00
Disty0 005eb789a3 SDNQ use combination config for Triton MM 2026-03-28 18:02:17 +03:00
Disty0 6eadf40f33 SDNQ add ufp aliases 2026-03-27 21:31:36 +03:00
Disty0 10942032a3 Add SDNQ_COMPILE_KWARGS env var 2026-03-25 02:10:58 +03:00
Disty0 11f3921832 SDNQ make dyn loss threshold auto selected by default 2026-03-25 01:47:59 +03:00
awsr c4ebef29a9 RUF013 updates 2026-03-24 05:48:19 -07:00
Disty0 f35e644d44 cast svd to torch_dtype 2026-03-11 14:50:14 +03:00
Disty0 e202d43801 SDNQ set sdnq_dequantize_fp32 to True by default 2026-03-11 14:35:11 +03:00
Disty0 30d21f420f Allow all types in dyn quant 2026-03-11 13:46:24 +03:00
Disty0 b02a63c60c SDNQ update minimum dtype check 2026-03-11 04:12:37 +03:00