Markus Fritsche marfrit
0 Followers · 0 Following
Repositories
27
 Projects Packages Public Activity Starred Repositories
marfrit pushed to master at marfrit/rocky-vulkan-llama 2026-07-12 04:17:49 +00:00
d32c653c9a Kernel bring-up: A1 hangs panvk compute, plain 7.0.0-rc3 runs it; rebuild-vs-A1 disproven (uABI byte-identical); both panthor 1.7.0
marfrit opened issue marfrit/rocky-vulkan-llama#2 2026-07-11 19:17:30 +00:00
Vulkan backend: garbage output for Qwen3 & Gemma architectures (Qwen2.5 works) — QK-norm suspect
marfrit opened issue marfrit/rocky-vulkan-llama#1 2026-07-11 18:46:15 +00:00
Cataract: NPU (RKNPU2) and open GPU (panvk/panthor) can't share a kernel
marfrit pushed to master at marfrit/rocky-vulkan-llama 2026-07-11 18:36:14 +00:00
107ccddbb1 Clean up for public release: README rewrite, LICENSE (MIT), patches reorganized
5b541d596e Shader CHAMPION converged: 28 t/s (4.8x), correctness-gated; f16acc refused; ~15% left is quantize-dispatch (profile, not tile)
6ee32814b2 Shader iter6: BK_STEP=1 regressed (barrier crossover); BK_STEP=2 is sweet spot, reverted to iter5 champion
220a3b1667 Shader iter5 CHAMPION: scoped BK_STEP (spec const 11) -> L=28 t/s, M restored to 20.6 (clean)
1e23ce0cd2 Shader iter4: BK_STEP=2 -> L path 28 t/s (+9%), but M regressed (global #define); needs L-scoping
Compare 10 commits »
marfrit created branch master in marfrit/rocky-vulkan-llama 2026-07-11 18:36:14 +00:00
marfrit created repository marfrit/rocky-vulkan-llama 2026-07-11 18:36:01 +00:00
marfrit closed issue marfrit/lmcp#23 2026-07-10 18:59:27 +00:00
ampere: apply lmcp v1.2.1 cutover (tools.d/ migration)
marfrit pushed to mtp-b9549 at marfrit/rk-llama.cpp 2026-07-10 14:59:59 +00:00
0ebb855635 docs: int4 closed — no weight-only int4 dtype on RK3588 (hardware limit)
marfrit pushed to mtp-b9549 at marfrit/rk-llama.cpp 2026-07-10 14:24:41 +00:00
2635371317 docs: E8 prefill pipeline result — flat, DDR-bandwidth-bound
marfrit pushed to mtp-b9549 at marfrit/rk-llama.cpp 2026-07-10 10:17:34 +00:00
7fbbd7acc3 Merge rknpu-int8-clamp: saturate int8 activation quantization
447672d647 rknpu2: saturate int8 activation quantization
Compare 2 commits »
marfrit pushed to mtp-b9549 at marfrit/rk-llama.cpp 2026-07-10 10:16:25 +00:00
f7449d2698 docs: RKNPU2 perf campaign results + matmul micro-benchmarks
marfrit pushed to main at marfrit/marfrit-packages 2026-07-10 07:51:04 +00:00
e2991ff55b linux-pinetab2-danctnix-besser: re-enable CONFIG_SHADOW_CALL_STACK
marfrit closed issue marfrit/besser#20 2026-07-10 07:51:01 +00:00
Re-enable CONFIG_SHADOW_CALL_STACK once GCC 15 arm_neon.h pragma issue is fixed
marfrit closed issue marfrit/marfrit-packages#81 2026-07-10 07:40:21 +00:00
bohr runner: sid sources leak transitively into trixie-target builds — root cause of #78
marfrit pushed to rknpu-int8-clamp at marfrit/rk-llama.cpp 2026-07-10 07:01:59 +00:00
447672d647 rknpu2: saturate int8 activation quantization
marfrit created branch rknpu-int8-clamp in marfrit/rk-llama.cpp 2026-07-10 07:01:59 +00:00
marfrit pushed to mtp-b9549 at marfrit/rk-llama.cpp 2026-07-10 05:26:04 +00:00
3cd6e67b55 docs: remove local infrastructure references from gemma4-optimized.md
marfrit pushed to mtp-b9549 at marfrit/rk-llama.cpp 2026-07-10 05:13:41 +00:00
3258224a57 docs: re-measure CPU/NPU baselines on an idle host
marfrit closed issue marfrit/marfrit-packages#110 2026-07-10 05:04:19 +00:00
upstream bump: lmcp 0.3.0 → 1.2.3
marfrit pushed to mtp-b9549 at marfrit/rk-llama.cpp 2026-07-10 04:53:23 +00:00
5c66767159 docs: gemma4-optimized.md — NPU serving config, models, measurements