Commit Graph
99 Commits
Author SHA1 Message Date
cumal 250ea6e385 fix: layer split across cuda (#2262) 2026-11-03 00:03:43 +08:00
leejet 290148d2e3 feat: add custom modulate operator and ConvRot fusion (#2258) 2026-11-02 01:19:28 +08:00
leejet 133087deb8 fix: prefer discrete GPUs in auto-fit device selection (#2259) 2026-11-02 01:09:36 +08:00
leejet 62c54439c8 fix: include missing cassert headers in SageAttention CUDA code 2026-11-02 00:17:49 +08:00
leejet 45a63ac460 refactor: move model-specific operators and fusions into backend extensions (#2253) 2026-11-01 21:18:37 +08:00
leejet 0a067120b3 perf: optimize Vulkan kernels, graph fusion, and workspace reuse (#2251) 2026-10-31 14:55:58 +08:00
leejet e239e4b0fd refactor: replace manual gated activations with native ggml GLU (#2250) 2026-10-30 01:32:46 +08:00
leejet 37bc88b5e9 perf: optimize H3 inference and segmented weight prefetch (#2240) 2026-10-30 00:38:14 +08:00
leejet 77b72a2efe perf: accelerate INT8 diffusion inference with CUDA fusions (#2235) 2026-10-28 22:02:46 +08:00
project516 2032526e8c fix: include integrated GPUs in automatic memory planning (#2228) 2026-10-28 01:28:11 +08:00
Mirekandleejet e6370bf093 perf: avoid redundant MOSS QKV and value layout copies (#2209)
Co-authored-by: leejet <leejet714@gmail.com>
2026-10-25 16:54:51 +08:00
leejet ea209dbfcf perf: cache CPU position embeddings across DiT runners (#2211) 2026-10-25 05:18:30 +08:00
leejet e0ed724e28 chore: format code 2026-10-24 01:23:42 +08:00
losewayy 40b1788476 feat: allow grouped temporal decoding for Wan2.2 TI2V 5B (#2196) 2026-10-24 01:09:05 +08:00
boriselec ddd6b3aa35 fix: correct Linux sysfs path for physical core detection (#2186) 2026-10-22 22:53:18 +08:00
leejet b15c8aac39 fix: include RPC devices in --list-devices (#2182) 2026-10-21 23:47:50 +08:00
vmobilis 939895f00e feat: display remaining time (#2076) 2026-10-21 23:29:05 +08:00
leejet af0fc72c96 chore: sync ggml and remove obsolete conv1d workarounds (#2147) 2026-10-17 02:39:51 +08:00
therealshulgin 8c442fb0ad fix: open model files with UTF-8 paths on Windows (#2144) 2026-10-17 00:50:09 +08:00
leejet e47426c044 fix: benchmark Vulkan direct convolution in auto mode (#2141) 2026-10-15 23:08:30 +08:00
leejet 46362fc4c0 feat: add automatic direct convolution selection (#2134) 2026-10-13 23:25:19 +08:00
leejet d763dae3b8 feat: add per-module automatic flash attention selection (#2132) 2026-10-13 22:58:45 +08:00
losewayy ce95c09ce1 fix: halve VAE decode tile size on OOM retry instead of clamping to 256px (#2120) 2026-10-10 22:33:37 +08:00
leejet a1ded76da5 chore: resolve MSVC warnings in ggml extensions 2026-10-06 20:10:36 +08:00
leejet fd43b542a6 refactor: escape newlines in prompt logs and clarify source separators (#2106) 2026-10-06 19:51:12 +08:00
stduhpfandleejet 96f67fe631 feat: Server: Add support for previews via sdcpp API (#2093)
Co-authored-by: leejet <leejet714@gmail.com>
2026-10-06 19:20:02 +08:00
leejet 203a882a28 refactor: shorten log levels and move source locations to the end (#2104) 2026-10-06 18:26:49 +08:00
leejet 42ab1c12f1 fix: keep scaled INT8 convrot matmuls on Vulkan (#2070) 2026-09-27 20:43:06 +08:00
leejet 19bbbca1c7 refactor: define VAE tile dimensions in image pixels (#2059) 2026-09-25 18:20:32 +08:00
leejet 0a9340c599 fix: scale Qwen Image 2.1 VAE convolutions (#2054) 2026-09-25 15:52:35 +08:00
leejet 1a2330de68 fix: reserve 128 MiB headroom when selecting monolithic execution (#2046) 2026-09-25 01:53:11 +08:00
leejet 740c7ae193 feat: add configurable Qwen cache types and early cache scheduling (#2045) 2026-09-25 01:51:01 +08:00
Dong Wang 500ef5fa7c fix: map mmapped weights through Metal buffers instead of CPU buffers (#2037) 2026-09-24 00:53:49 +08:00
leejet 2dc7f5408a feat: add Qwen Image 2.1 prefix KV cache (#2035) 2026-09-23 23:08:21 +08:00
leejet 97d932b8f8 fix: restrict VAE tiling retries to allocation failures (#2019) 2026-09-21 23:40:05 +08:00
leejet 187b2561ea feat: add native CUDA SageAttention support (#2005) 2026-09-21 01:16:59 +08:00
Wagner Bruna b8248a869c fix: avoid passing ggml logs as format strings (#2002) 2026-09-21 00:50:52 +08:00
leejet 1330cebae8 feat: support building with upstream ggml (#1999) 2026-09-19 22:25:01 +08:00
leejet 17860c0e45 perf: parallelize host tensor elementwise and broadcast ops (#1998) 2026-09-19 21:46:16 +08:00
leejet 275ab58e01 perf: reduce CPU overhead in graph execution and sampling (#1997) 2026-09-19 18:32:58 +08:00
Fabrice Aneche d32b4e893b fix: prevent clip_preprocess center crop from exceeding the resized image (#1995) 2026-09-19 18:08:34 +08:00
leejet 3e037a81e4 perf: accelerate VAE direct 3D convolutions (#1996) 2026-09-19 17:50:01 +08:00
leejet adcac69650 perf: pad small attention heads to 64 for MMA Flash Attention (#1992) 2026-09-18 23:59:58 +08:00
leejet 656a1354c3 refactor: remove obsolete unused tensor filtering (#1984) 2026-09-18 23:41:37 +08:00
leejet cc515a01f9 perf: eliminate temporary allocations in Philox rounds (#1982) 2026-09-17 02:09:12 +08:00
Wagner Bruna 07a85c74cb feat: support Brownian tree noise in all noise injection samplers (#1899) 2026-09-15 02:37:31 +08:00
leejet 4964abdfc5 feat: support external Hugging Face tokenizer JSON files (#1973) 2026-09-15 00:27:39 +08:00
Georgeandleejet 0bd72f075a feat: add Wan2.2 S2V (audio+img-to-video) support (#1925)
Co-authored-by: leejet <leejet714@gmail.com>
2026-09-13 23:33:50 +08:00
fszontagh 4a7da26b73 fix: bound plain-text runs in parse_prompt_attention regex (#1919) 2026-09-13 23:29:19 +08:00
leejet 9a977388a8 fix: guard GPU memory capacity and propagate encoding failures (#1958) 2026-09-13 21:35:46 +08:00