Commit Graph

  • b3f50180ec Read half-aligned Qwen Q2 weight rows correctly on Metal main antirez 2026-11-03 16:55:02 +01:00
  • 5e3a3449e2 Check both staging widths in the Qwen compact expert dispatch test antirez 2026-11-03 16:52:44 +01:00
  • cd1c5507b6 Preserve resident Qwen prefill speed with speculative decoding antirez 2026-11-03 16:50:55 +01:00
  • 2cd0ae7064 Keep M6 Qwen tuning separate from DeepSeek and GLM kernel selection antirez 2026-11-03 16:41:21 +01:00
  • 5c9ed63c16 Keep generated Qwen research dumps out of the release tree antirez 2026-11-03 16:38:47 +01:00
  • 7bba0f7d2c Load Qwen version 2 checkpoints and rebuild missing predictor history on demand antirez 2026-11-03 16:38:47 +01:00
  • a2798605b7 Use resident Qwen arithmetic when experts are streamed from SSD antirez 2026-11-03 16:37:20 +01:00
  • 0dbfd05e16 Keep Metal-only Qwen predictor kernel tests out of the CUDA binary antirez 2026-11-03 16:32:40 +01:00
  • afd33efab2 Keep current M5 expert tuning alongside the streamed Metal path antirez 2026-11-03 16:29:39 +01:00
  • c942a90fdd Restore Qwen predictor scratch views after the fused resident expert path antirez 2026-11-03 16:28:41 +01:00
  • 4ef9e4d0d0 Merge #1056: Giorgio Oppo’s Qwen3.8 Metal SSD expert streaming antirez 2026-11-03 16:28:10 +01:00
  • 9eb4a12f25 Merge upstream/main into qwen-kernel-opt Giorgio Oppo 2026-11-02 21:35:23 +01:00
  • e4709b44be Share DeepSeek V4 Metal index queries across four keys antirez 2026-11-02 11:36:48 +01:00
  • 9fe88cd312 Integrate the per-model Metal weight policy fix antirez 2026-11-02 11:26:14 +01:00
  • 890898cfb4 Apply Metal weight tracking to the model being opened antirez 2026-11-02 11:26:14 +01:00
  • a4bbb38ce4 Integrate the V4.1 memory-aware Metal TP fix antirez 2026-11-02 11:17:03 +01:00
  • c54841002d Keep V4.1 TP sharded when either Mac needs the smaller working set antirez 2026-11-02 11:15:23 +01:00
  • 8c54d7fb01 Merge #1073 with local Metal and Spark optimizations antirez 2026-11-02 11:11:33 +01:00
  • 0accf8b395 Merge #1073: speed up DeepSeek V4.1 Flash on Metal antirez 2026-11-02 11:04:52 +01:00
  • 6075227f1e Preserve the Spark small-batch shared-expert projection antirez 2026-11-02 11:04:22 +01:00
  • 224e77e791 Keep Metal-only V4.1 decode helpers out of the Spark build antirez 2026-11-02 11:04:22 +01:00
  • a88711246c Merge reviewed server PRs #1165 and #1245 from upstream antirez 2026-11-01 18:07:54 +01:00
  • b4e3faf526 Merge #1245: allow up to 256 images per server request Salvatore Sanfilippo 2026-11-01 18:07:47 +01:00
  • da2a102dd0 Merge #1165: advertise the loaded model modalities Salvatore Sanfilippo 2026-11-01 18:06:35 +01:00
  • b7615c141f Keep the default CUDA unroll policy for multirow Qwen Q8 antirez 2026-11-01 09:46:06 +01:00
  • 36e2680676 Link ROCm fallback symbols into standalone GPU test tools antirez 2026-11-01 09:43:04 +01:00
  • ae40c5c4f0 Carry the complete V4.1 raw window through medium prefill pruning antirez 2026-10-31 20:43:40 +01:00
  • c1bce729fc Cross the legacy GLM sparse boundary in one SSD layer sweep antirez 2026-10-31 20:42:03 +01:00
  • ef8e536501 Reuse Q2_K bit planes in Qwen Metal expert prefills antirez 2026-10-31 20:35:44 +01:00
  • 524c874506 Skip wired GLM weights during full-layer SSD read-ahead antirez 2026-10-31 20:32:49 +01:00
  • 46ade1bb3e Reuse rotated compact keys across legacy GLM prefill queries antirez 2026-10-31 20:31:53 +01:00
  • 1c6d058d9f Gather Qwen BF16 prompt embeddings directly on the GPU antirez 2026-10-31 20:30:32 +01:00
  • d9fc9cf30c Merge #1147: validate sparse selection prefixes in Metal GLM prefill antirez 2026-10-31 20:30:04 +01:00
  • b799c90e3a Resize independent image rows in parallel before vision prefill antirezandMarcello Costagliola 2026-10-31 20:28:10 +01:00
  • 773247a881 Skip whole decoder chunks in medium V4.1 prefills antirez 2026-10-31 20:21:25 +01:00
  • 8c9b1986a4 Merge #1127: overlap short callback-driven Metal Flash prefills antirez 2026-10-31 20:15:45 +01:00
  • c2779bb4e8 Keep the four-head Flash indexer schedule on Metal TP antirez 2026-10-31 17:49:26 +01:00
  • 32604bc6de Issue Qwen Q8 weights ahead during single-token CUDA decode antirez 2026-10-31 17:46:11 +01:00
  • dc9580daf4 Merge #1213: stop discarded GLM MTP drafts after their KV write antirez 2026-10-31 17:38:36 +01:00
  • babe24db9f Fill single-token GLM indexer tensor tiles with heads on Spark antirez 2026-10-31 17:33:15 +01:00
  • db67455bed Reduce Metal Flash and Pro index scores with one barrier antirez 2026-10-31 17:33:15 +01:00
  • efd018cc43 Tile single-host GLM 5.3 Flash indexed attention on Metal antirezandtrueimage 2026-10-31 17:32:53 +01:00
  • b00bbf42da Remove per-head barriers in the CUDA Flash and Pro decode indexer antirez 2026-10-31 17:29:39 +01:00
  • 79e59d7ac2 Merge #1270: keep legacy GLM MTP live across the sparse boundary antirez 2026-10-31 17:26:35 +01:00
  • 95976d737f Merge #1238: speed up Qwen3.8 Metal decode antirez 2026-10-31 17:16:16 +01:00
  • d61aa0c139 Reuse V4.1 index queries across eight keys per Spark warp antirez 2026-10-31 16:27:39 +01:00
  • e3137ce6bc Batch disk Engram reads for Metal verifier rows antirezandbadelon 2026-10-31 15:54:28 +01:00
  • 23c333bd1c Skip rejected V4.1 index keys in Metal SSD decode antirez 2026-10-31 15:52:26 +01:00
  • c7bb2d7656 Skip masked V4.1 index scores on DGX Spark antirez 2026-10-31 15:49:39 +01:00
  • 4c29e82b3e Merge #1034: queue single-host V4.1 Metal SSD decode by default antirez 2026-10-31 15:47:40 +01:00
  • 47400ed008 Merge #1262: skip rejected V4.1 keys without changing Metal top-k ties antirez 2026-10-31 15:42:39 +01:00
  • 32f4243a09 Merge #1261: pipeline resident V4.1 Metal decode submissions antirez 2026-10-31 15:37:19 +01:00
  • f2e60304a3 Merge upstream/main into qwen-kernel-opt Giorgio Oppo 2026-10-31 09:39:36 +01:00
  • e14d7f8b1e Release cached engines between regression tests antirez 2026-10-30 19:32:07 +01:00
  • c97964b8aa Validate Qwen tokens before prefetched input staging antirez 2026-10-30 18:10:33 +01:00
  • 21178462db Test indexer row-group tails and output bounds antirez 2026-10-30 17:57:19 +01:00
  • 345d7f44cb Score four indexer rows per SIMD group HDYS-TTBYS 2026-10-24 09:43:03 +09:00
  • e73e516e75 metal: reuse Q4_K raw loads in Qwen NAX prefill Matthijs van den Bos 2026-10-09 15:30:41 +02:00
  • 64baaa0678 Test Qwen PLE prefetch parity and recovery Matthijs van den Bos 2026-09-29 23:11:25 +02:00
  • 347c2e1fbe Overlap Qwen PLE reads with prefill GPU work Matthijs van den Bos 2026-09-29 22:49:30 +02:00
  • 82aa7e8e60 metal: use shuffle network for non-causal top-k block sort Emilian Bold 2026-10-07 15:02:24 +03:00
  • 5382eac5e0 metal: process medium V4.1 SSD appends in one layer sweep nomish 2026-10-19 20:22:48 -07:00
  • a8ce74afd5 server: render blank reasoning as absent Tim Lawrenz 2026-10-28 14:55:16 +00:00
  • 97e32768e6 metal: use exact divisions in the V4.1 FP4 activation quantizer nomish 2026-10-19 11:55:07 -07:00
  • 381808d839 Test CPU-touched Metal buffer reclamation antirez 2026-10-30 17:49:58 +01:00
  • a4c922c560 Fix a Metal session memory leak on /ctx Mark Shank 2026-10-21 11:11:03 -04:00
  • 2768d0ed3c fix(jpeg): validate SOS spectral parameters in progressive scans Giulio Davide Carparelli 2026-10-30 00:45:26 +01:00
  • 3459d5546b docs: revalidate streaming decode queue on Flash 1017 Dango233 2026-10-30 18:57:29 +08:00
  • fc69b626ae Merge remote-tracking branch 'origin/main' into dango-codex/queue-1017-validation Dango233 2026-10-30 15:33:45 +08:00
  • 1e6a0e9602 Decode: give batched Engram tables separate row buffers HDYS-TTBYS 2026-10-26 23:12:42 +09:00
  • 694620956b Qwen decode: DS4_QWEN4_NO_PLE_OVERLAP restores up-front PLE staging; name the moe down_reduce n_hc slot paradise-runner 2026-10-26 06:36:53 -06:00
  • 1be7b4db2a Decode: give each Engram table its own input buffer HDYS-TTBYS 2026-10-25 23:16:51 +09:00
  • d9f2d15b40 glm: bridge MTP verification across the indexed attention boundary Luo Hanjun 2026-10-25 15:22:31 +08:00
  • 5fd33a4e53 Set the candidate-decode flag after the Engram reads HDYS-TTBYS 2026-10-24 00:00:24 +09:00
  • 9d7a00b4fb Restrict candidate-row scoring to single-token decode HDYS-TTBYS 2026-10-23 23:04:04 +09:00
  • 01772c087a Optimize V4.1 long-context decode HDYS-TTBYS 2026-10-23 22:31:35 +09:00
  • 57df7a5bf4 Pipeline V4.1 single-host batched decode command buffers HDYS-TTBYS 2026-10-23 18:24:30 +09:00
  • be2773bc6f Pipeline V4.1 single-host decode command buffers HDYS-TTBYS 2026-10-23 18:24:30 +09:00
  • 72ccf36317 Revert the V4.1 decode pre-encode: held command buffer, per-token GPU spin Tarjei Mandt 2026-10-22 15:04:10 +11:00
  • 51fef8784a Merge pull request #9 from IngeniousIdiocy/v41-decode-on-kp7 Tarjei Mandt 2026-10-22 15:04:09 +11:00
  • 3650da359e Turn off the V4.1 TP concurrent shared expert with the fixed-route switch Tarjei Mandt 2026-10-22 11:54:16 +11:00
  • ce5b718be8 Start V4.1 short-sweep suffix pruning at 512 rows Tarjei Mandt 2026-10-22 11:54:06 +11:00
  • 5b83cdfe8c Merge pull request #7 from ivanfioravanti/v41-tp-decode-engram-warm Tarjei Mandt 2026-10-22 11:53:51 +11:00
  • e1d8114167 Encode the next V4.1 decode token while this one finishes Mark Shank 2026-10-21 11:55:48 -04:00
  • 4ff3af604f Speed up the V4.1 decode top-512 Mark Shank 2026-10-21 11:55:16 -04:00
  • 64ceeea4f5 Score long V4.1 decode index rows from bf16 keys Mark Shank 2026-10-21 11:54:36 -04:00
  • e99dc21751 Spread two single-core V4.1 decode reductions over the GPU Mark Shank 2026-10-21 11:54:00 -04:00
  • f0b4f55fcc Compute one RoPE table per V4.1 decode token Mark Shank 2026-10-21 11:53:25 -04:00
  • 55b1e75e64 Narrow more V4.1 decode weights and read them with wider loads Mark Shank 2026-10-21 11:52:48 -04:00
  • b66c926046 Compile V4.1's fixed decode shapes into its decode matvecs Mark Shank 2026-10-21 11:52:09 -04:00
  • 95606b5045 Fix a data race in the V4.1 FP8 weight-copy builder Mark Shank 2026-10-21 11:51:33 -04:00
  • 26335c7044 Overlap V4.1 decode hyper-connection weights with the projections Mark Shank 2026-09-18 19:27:54 -04:00
  • f5127dfc82 Allow up to 256 images per server request ivanfioravanti 2026-10-21 19:43:58 +02:00
  • 5a588dbf12 Share weight loads across two rows in the exact TP K-slice matvec ivanfioravanti 2026-10-21 17:54:49 +02:00
  • 565db43559 Prune the V4.1 decoder suffix of short sweeps without changing the output ivanfioravanti 2026-10-21 17:20:24 +02:00
  • 187f46b4fa ds4-bench: per-step decode logits hash and dump ivanfioravanti 2026-10-20 07:45:39 +02:00
  • bdae7640ef Warm the V4.1 Engram page cache when the tables fit in memory ivanfioravanti 2026-10-20 07:45:21 +02:00
  • e91c453ec2 Overlap more of V4.1 two-Mac TP decode ivanfioravanti 2026-10-21 08:13:51 +02:00
  • ee7e840523 cuda: improve SSD-streamed prefill with ordered expert reads antirez 2026-10-21 09:30:33 +02:00
  • 0f9b03c9df Speed up exact V4.1 prefill matmuls and indexer scores on M3 ivanfioravanti 2026-10-20 07:43:50 +02:00