Markus Fritsche marfrit
marfrit pushed to master at marfrit/rkopnu 2026-07-19 08:43:04 +00:00
00dff56288 Rename platform_driver "rocket" -> "rkopnu"
marfrit pushed to master at marfrit/rkopnu 2026-07-19 08:38:04 +00:00
2b68134925 Package rkopnu as a reproducible repo for third parties
27472de2f8 rkopnu: Fable-review stability fixes for persistent 128K-context server
0a6774ae61 rkopnu: detach IOMMU domain on file close (fix UAF/panic from attach-once)
6a8b20ab3f rkopnu Phase 4c: honor core_mask -> vendor parity (51.5 vs 51.6 tok/s)
ff5316b2be rkopnu Phase 4b: cache the task-BO kernel vmap across submits (perf)
Compare 10 commits »
marfrit created branch master in marfrit/rkopnu 2026-07-19 08:38:04 +00:00
marfrit created repository marfrit/rkopnu 2026-07-19 08:37:51 +00:00
marfrit pushed to rocket-npu-backend at marfrit/rk-llama.cpp 2026-07-18 13:22:17 +00:00
67316774a0 ggml-rocket: make ROCKET_MIN_BATCH runtime-tunable via GGML_ROCKET_MIN_BATCH
marfrit pushed to master at marfrit/lmcp 2026-07-18 11:23:30 +00:00
2bb7b94a66 tools.d: add boltzmann stash tool (fleet-memory CLI wrapper)
marfrit pushed to master at marfrit/lmcp 2026-07-18 06:53:18 +00:00
80fb60c60f hub: serve apropos + wake_fleet (fleet-central tools) alongside remote_*
marfrit pushed to main at marfrit/rosenblatt 2026-07-15 07:59:02 +00:00
c77be48151 docs: genericize host references to machine type (Rock 5 ITX+)
marfrit pushed to main at marfrit/rosenblatt 2026-07-14 21:41:47 +00:00
27df00a201 Add REPLICATION.md: reproduce rocket-NPU 1GHz LLM accel on vanilla kernel+userspace
a2244675b1 npu-probe: backport mmap-leak fix + K>8192 backstop from ggml-rocket
8c430d48b3 npu-probe: ggml-linked bridge test — NPU reproduces ggml Q8_0 mul_mat
Compare 3 commits »
marfrit pushed to rocket-ksplit-wip at marfrit/rk-llama.cpp 2026-07-14 20:58:24 +00:00
6d5a075a6c ggml-rocket: K-segmentation for K>8192 (int8-partial float-accumulate)
marfrit created branch rocket-ksplit-wip in marfrit/rk-llama.cpp 2026-07-14 20:58:24 +00:00
marfrit pushed to rocket-npu-backend at marfrit/rk-llama.cpp 2026-07-14 20:58:23 +00:00
288d70413a ggml-rocket: multi-fd (3-core) column-split + task_num param
412cba2601 ggml-rocket: CBUF weight-reuse via multi-task job chaining
367143edf3 ggml-rocket: implement offload_op; add GGML_NO_REPACK measurement guard
3d395c1612 ggml-rocket: per-call BO pool + column-major weight caching (prefill 39->49 t/s)
698c6b731c ggml-rocket: min_batch floor — keep M<32 (decode) on CPU
Compare 7 commits »
marfrit created branch rocket-npu-backend in marfrit/rk-llama.cpp 2026-07-14 20:58:22 +00:00
marfrit pushed to main at marfrit/rosenblatt 2026-07-14 07:15:52 +00:00
9d4981891e Rosenblatt U6: quantization bridge validated on NPU (float->quant->NPU->dequant)
marfrit pushed to main at marfrit/rosenblatt 2026-07-14 07:13:51 +00:00
4a0bfcde15 Rosenblatt U6: per-output-channel requant in rkt_npu_matmul
marfrit pushed to main at marfrit/rosenblatt 2026-07-14 07:05:01 +00:00
eb2397ccf5 Rosenblatt: rkt_npu_matmul() — reusable arbitrary-size NPU matmul (U6 primitive)
marfrit pushed to main at marfrit/rosenblatt 2026-07-14 07:01:12 +00:00
562373ab83 Rosenblatt: full 2D-tiled arbitrary matmul on NPU (M and N), gemma-scale verified
marfrit pushed to main at marfrit/rosenblatt 2026-07-14 06:58:55 +00:00
2270f749bf Rosenblatt: arbitrary-size matmul on the NPU via tiling (U6 core primitive)
marfrit pushed to main at marfrit/rosenblatt 2026-07-14 06:44:31 +00:00
dcf7061893 Rosenblatt: expose real requant + bias in the matmul builder API
marfrit pushed to main at marfrit/rosenblatt 2026-07-14 06:36:16 +00:00
1fb810d93c Rosenblatt: FIRST CORRECT INT8 MATMUL ON THE MAINLINE ROCKET NPU