Commit Graph

Select branches

Hide Pull Requests

main

Mono Color

• c1a36daada tests: no-alloc ctx for CUDA backend path (fix GGML_ASSERT on alloc) main Marvin 2026-09-05 21:53:48 -0300
• a34feeb0f9 cuda: chunked WY delta-net kernel v1 (opt-in DELTA_WY_CUDA=1) + harness --cuda mode Marvin 2026-09-05 21:13:04 -0300
• 3a7d1016b3 tests: WY candidate validation, exp-diff ratios, --cuda mode, chunk sweep Marvin 2026-09-05 21:03:27 -0300
• f5b494b0e2 tests: WY chunked delta-net candidate with fallback guard (test-delta-chunk 42/42) Marvin 2026-09-05 20:56:24 -0300
• a47a8a65e8 tests: delta-net chunked-recurrence correctness harness (test-delta-chunk) Marvin 2026-09-05 20:29:11 -0300
• e2728c8547 qwen4exp: PP/TG opts for Ampere+Zen3 (merged experts, fused PLE taps, Zen3 tiling) Marvin 2026-09-05 20:14:29 -0300
• fe215a8ccd
  qwen4exp: gather selected cells for depth-constant TG attention (#2404) Joel Farthing 2026-09-03 12:35:31 -0500
• 6d6fe936db
  Support SWA compression with DFlash and DSpark (#2384) Samuel Oliveira Alves 2026-09-03 13:23:23 -0300
• caf7eae528
  map dense Qwen DFlash packages correctly (#2370) Samuel Oliveira Alves 2026-09-02 13:54:01 -0300
• 563b798ac9
  qwen4exp: MTP (NextN) self-speculative decoding support (#2369) James 2026-09-02 12:49:42 -0400
• 68bf92bfc3
  gemma4: compacted sliding-window KV cache (--swa-compress) (#2378) Alexey Zilber 2026-09-03 00:47:53 +0800
• e560283754
  Remove usless check in llama-quantize (#2394) Kawrakow 2026-09-02 18:45:49 +0200
• c2206b80da
  docs: --cache-ram-similarity is a minimum, not a maximum (#2395) mattafaak 2026-09-02 12:44:32 -0400
• 3c58ae373a
  loader: add `--defer-ple` to keep per-layer token embedding out of resident memory (#2389) Joel Farthing 2026-08-31 11:49:44 -0500
• ab6d816818
  CUDA DSA: fix v_offset for quantized K/V caches (#2387) Alex 2026-08-31 19:35:17 +0300
• 15dddc60b3
  Qwen3.8-Flash-Next: faster TG on CUDA (#2373) Kawrakow 2026-08-28 18:11:17 +0200
• 555330fbba
  CUDA: handle GQA = 12 for head size = 256 via new MMA (#2372) Kawrakow 2026-08-28 18:10:50 +0200
• 7cff686d37
  Quantization fudge factors (#2361) Kawrakow 2026-08-27 17:35:04 +0200
• 96ac84c587
  Update README.md Kawrakow 2026-08-27 17:33:10 +0200
• 0b4d09a227
  model: Add Qwen3.8-Flash-Next (qwen4exp) runtime support (#2365) Joel Farthing 2026-08-27 10:29:33 -0500
• 78ce50c163
  Fix grid overflow in get_rows on CUDA (#2367) Kawrakow 2026-08-27 15:40:41 +0200
• 6b261d771e
  cuda: fix iq3_k value table selection in MMVQ (#2364) Yap Sok Ann 2026-08-27 13:40:14 +0700
• ef40550042
  server: fix finish_reason for token-limit stops (#2362) Joel Farthing 2026-08-26 12:54:11 -0500
• 1d76336eeb
  fix(server): capture all server log sinks in --log-file, gated on explicit flag (#2313) Skelectric 2026-08-26 13:34:25 -0400
• 2f068b5d87
  dflash: use draft context as capacity contract (#2341) Samuel Oliveira Alves 2026-08-26 13:26:50 -0300
• 850320be8b
  metal: initialize encode_async in ggml_backend_metal_init (#2334) Riccardo Chiumiento 2026-08-26 17:15:46 +0200
• 28fbe34ce9
  Dflash 2 speculative decoding (#2345) Samuel Oliveira Alves 2026-08-26 12:09:34 -0300
• 73ad16269b
  rpc: fix crash running GLM-5.2 (glm-dsa) split over RPC (#2360) Coenie Beyers 2026-08-26 17:02:43 +0200
• b166e2696e
  server: charge the generation budget per accepted token (#2358) Joel Farthing 2026-08-26 01:00:47 -0500
• 08b500b958
  ggml: fix HC_POST single-token CPU chunk count (#2357) Joel Farthing 2026-08-25 09:29:06 -0500
• c49f7db34a
  Fix MMQ check when quant does not support MMQ (#2356) Kawrakow 2026-08-25 13:44:32 +0200
• d206417cb3
  server: fix prompt re-use with `--reasoning-tokens none` (#2353) Joel Farthing 2026-08-25 02:40:30 -0500
• 97370e3f27
  chat: fix multi-argument tool calls for tagged templates (#2351) Joel Farthing 2026-08-25 02:38:39 -0500
• d180050f89
  cuda : repair the HIP build, and validate IQ4_KS and IQ4_KT on RDNA3 (#2339) Guy Barel 2026-08-25 07:50:44 +0100
• 0ed847d314
  Adaptive P Sampler: Quality Control (#2337) dungquixote42 2026-08-24 12:49:49 -0400
• 26113d1dd3
  ggml-cuda: bind cublas handle to the backend stream in DSA attention (#2347) Yap Sok Ann 2026-08-24 23:48:12 +0700
• c574620b12
  Fix KQ mask padding for the Vulkan back-end (#2350) Kawrakow 2026-08-24 18:31:17 +0200
• 64109a4d60
  vulkan : add IQ4_KS and IQ4_KT support (#2332) Guy Barel 2026-08-24 17:20:35 +0100
• ad26e68bee
  Apply callback to extract features in spec (#2348) Samuel Oliveira Alves 2026-08-24 07:17:23 -0300
• 477852c1c9
  Load standalone Qwen3.5 MTP GGUFs passed with -md (#2328) Thireus ☠ 2026-08-24 09:03:27 +0100
• 66b2f50ce3
  Allow dspark to draft more that the amount of block size (#2323) Samuel Oliveira Alves 2026-08-24 04:55:26 -0300
• 6831fa6d8e
  CUDA graphs improvements (#2316) Kawrakow 2026-08-24 09:51:59 +0200
• 8337e4cd38
  Fix Qwen35+ MTP (#2322) Kawrakow 2026-08-15 19:35:03 +0200
• 1794846f73
  Fix Gemma4 MTP (#2324) Kawrakow 2026-08-15 19:34:47 +0200
• 7cd62a3eb2
  More principled CUDA DSA (#2315) Kawrakow 2026-08-15 09:41:21 +0200
• 8e703ddd8a
  server: accept max_completion_tokens as alias for max_tokens (#2321) Jun Yamog 2026-08-15 19:40:07 +1200
• 85a784505d
  cast embeds for F32 if necessary (#2319) Samuel Oliveira Alves 2026-08-15 01:21:56 -0300
• 43afea46c2
  Adapt Muse-Glimmer loading (#2314) Kawrakow 2026-08-14 07:55:11 +0200
• 981e5ea0d7
  DSpark: gather BF16 Markov rows (#2304) Samuel Oliveira Alves 2026-08-13 13:02:46 -0300
• cf711918e2
  Synch DFlash Tokens ID (#2303) Samuel Oliveira Alves 2026-08-13 13:02:08 -0300
• 37d82c2313
  Change the default amb value from 0 to 256 (#2312) Kawrakow 2026-08-13 17:47:37 +0200
• 6e7378f616
  Another minor optimization on CUDA for split mode graph (#2298) Kawrakow 2026-08-13 15:25:28 +0200
• 8b276c08ef
  CUDA: fuse rms -> add -> rms (#2297) Kawrakow 2026-08-13 15:24:45 +0200
• 3c949f3399
  Add work buffer size calculation for the ds4_comp op (#2307) Kawrakow 2026-08-13 15:23:38 +0200
• ff141691a1
  Use f32 accumulation in CUDA DSA implementation (#2311) Kawrakow 2026-08-13 15:22:01 +0200
• a10ef3eb00
  laguna: compacted sliding-window KV cache (--swa-compress) (#2310) Joel Farthing 2026-08-13 07:54:36 -0500
• 2cda8d2daf
  speculative: Allow --swa-compress with DeepSeek4 MTP (#2309) Joel Farthing 2026-08-13 04:45:39 -0500
• 4b0320381a
  Fix #2183 (#2308) Kawrakow 2026-08-13 09:58:03 +0200
• ee77f7ffb8
  Fix MXFP4 non-interlevaed type (#2306) Kawrakow 2026-08-13 08:41:39 +0200
• 79c1e16a41
  Update docs (#2299) Marian M. 2026-08-13 09:05:27 +0300
• 87644e36bc
  model: Ling-3.0 (bailingmoe3) runtime support (#2295) Joel Farthing 2026-08-13 01:02:50 -0500
• c46ffaa566
  fix dspark: seed draft block at id_last's true position (+1 off-by-one) (#2296) rumas77 2026-08-12 12:21:30 -0400
• 1dede1d79e
  Adding Muse-Glimmer support (#2293) Kawrakow 2026-08-12 15:53:54 +0200
• 26ceed9d40
  CUDA: clear MMQ row padding on partially offloaded quantized weights (#2292) Joel Farthing 2026-08-11 02:07:01 -0500
• b382ebd848
  llama: pass rope freq factors to build_std_attention (#2291) mb8565 2026-08-11 01:17:25 -0500
• c8772a8429
  Fix rope type for 39 architectures (#2290) mb8565 2026-08-11 01:09:24 -0500
• 5763a901de
  DSA: do not copy V rows when V == K (#2287) Kawrakow 2026-08-10 18:49:05 +0200
• b37189aae4
  Actually fix quantized indexer cache on CUDA (#2286) Kawrakow 2026-08-10 18:46:02 +0200
• b8b3034b0f
  Indexer topk: on the CPU repack Q8_0 indexer cache (#2285) Kawrakow 2026-08-10 18:45:43 +0200
• 7ebbb906d2
  Initial implementation of DSpark (#2280) Samuel Oliveira Alves 2026-08-10 03:46:03 -0300
• a7c81affa4
  GLM-5.2 vision hack (#2283) abc-nix 2026-08-09 13:54:53 +0000
• 7c57e445b7
  state: include compacted sliding-window rows in partial sequence state (#2281) Joel Farthing 2026-08-09 04:45:05 -0500
• f2328aa0c1 Fix -ctk / -ctv / -ictk that I broke earlier Kawrakow 2026-08-08 14:41:48 +0000
• da5884a2db Fix not commented out fprintf Kawrakow 2026-08-08 14:37:35 +0000
• 7642ac3eca
  Fix massive inefficiency in CUDA Q->f32/f16 and f32/f16->Q copies (#2279) Kawrakow 2026-08-08 17:26:59 +0300
• daa54abd0b
  DS4: do not cast caches to f32 (#2278) Kawrakow 2026-08-08 17:19:23 +0300
• 76c1942716
  Allow Q8_0 cache in the CUDA DSA implementation (#2276) Kawrakow 2026-08-08 17:18:36 +0300
• c5f37f0fba
  Indexer topk optimization (#2274) Kawrakow 2026-08-08 17:14:18 +0300
• 1ce4bb9736
  sweep-bench: fixes and new options (#2273) Joel Farthing 2026-08-08 08:53:35 -0500
• 86ad770f2a
  DS4: enable graph reuse (#2270) Kawrakow 2026-08-08 11:18:21 +0300
• 40dffce685
  CUDA: Fix non-F16 indexer top-k (#2272) Joel Farthing 2026-08-07 11:35:17 -0500
• 61b37f5f44
  Reduce the indexer temporary buffer size (#2271) Kawrakow 2026-08-07 17:09:18 +0300
• 36925b86c1 Compiler warning Kawrakow 2026-08-07 06:10:32 +0000
• e21eed5f58
  deepseek4: compacted sliding-window KV cache (--swa-compress) (#2266) Joel Farthing 2026-08-07 00:36:08 -0500
• 466d9bb15f
  Merged up/gate for DS4 (#2257) Kawrakow 2026-08-07 07:51:58 +0300
• da2293ded3
  Better placement of MoE tensors with -ncmoe and 1 GPU (#2262) Kawrakow 2026-08-07 07:51:35 +0300
• 113f426a3c
  Do not include ggml-impl.h in ggml-cuda.cu (#2264) Kawrakow 2026-08-07 07:23:43 +0300
• ea791ac569
  speculative : add Step 3.7 MTP support (#2250) Samuel Oliveira Alves 2026-08-07 01:23:21 -0300
• 21819724b3
  Compute the sequence fingerprint once per decode (#2265) ShubhamPriyadarshi 2026-08-06 20:04:42 +0530
• 4a4a6d3c14
  Fix wrong output for hybrid/recurrent models at -np > 1 (graph reuse key) (#2260) ShubhamPriyadarshi 2026-08-06 19:15:20 +0530
• d44e2cbe57
  openpangu: per-sequence state save/restore with --swa-compress (#2261) Joel Farthing 2026-08-06 03:00:14 -0500
• cf1aa57e1a
  openpangu: opt-in compacted sliding-window KV cache (--swa-compress) (#2253) Joel Farthing 2026-08-05 10:00:32 -0500
• b4be4b17a0
  Another minor indexer optimization on the CPU (#2231) Kawrakow 2026-08-05 08:15:41 +0300
• 1474935e31
  Bucket top_k (CPU): ~3% better TG at 128k context (#2225) Kawrakow 2026-08-05 07:50:41 +0300
• 7257cabc49
  Fix #2201 (#2255) Kawrakow 2026-08-05 07:48:20 +0300
• 6633d19df8
  Fix Antirez DS4 GGUFs (#2256) Kawrakow 2026-08-05 07:47:42 +0300
• 6b55d2c750
  Fix DSV4 tool calls and reasoning (#2242) Yap Sok Ann 2026-08-04 23:28:06 +0700
• af4e0cbdb6
  ggml : fuse the delta-net recurrent state copy into the op (#2251) Petr Vilím 2026-08-04 18:27:27 +0200
• 0a93e73a72
  Make antirez GGUFs run at least (#2254) Forkoz 2026-08-04 13:00:19 +0000
• f8b5664c18
  Chores : Typos fixing round 3 (project wide, ggml dir included, comments and user facing msg only) (#2249) Nexesenex 2026-08-04 06:15:28 +0200