Niko1221 16e08b50cd Merge f-029-par-amd (4772f0cd9)

2026-10-30 18:39:27 +01:00

..

artifact

Gyro-L: read the pair-trellis types TQP5/TQP6/TQP7 (ggml 147-149) on the CPU, CUDA and the MMQ prompt path (#3806)

2026-10-27 01:57:57 +01:00

core

shared arena: memfd backing (not limited by the /dev/shm size), one engine reads the experts and the others skip the read, RAM check before the pages are taken

2026-10-30 04:46:51 +01:00

dsv41

DeepSeek engine opt-in: the busy experts of a prompt take FP16 GEMMs, not the tiles (DS4_FAT_ROWS=N, default 0: byte-identical path, the packed tiles run exactly as before). An expert of N rows or more has its weights decoded to FP16 once (iq::f16_rows, the values the tiles multiply), the rows gathered (gather_f16), cublasGemmEx, clamped SwiGLU in FP16 (swiglu_f16); SetArgs skip_ge leaves those experts out of the tiles. The tiles decode an expert's weights again for every 64 of its rows; a popular expert of a long chunk has hundreds. Measured on the RTX 5090 with the IST-DAS 2.25 bpw pack, v0.2.7 base, DS4_FAT_ROWS=64: a 13K-token prompt 1,400 -> 1,950 tok/s, a 5.4K prompt 1,065 -> 1,181, decode unchanged (28-33 short, 47-49 warm). 3x V100, Maya-XS: +6-8% on 8K and 32K prompts. The same products, the sums' order differs: greedy output can diverge after the first rounding (wording equivalent); KL vs the reference 0.276, the stock path 0.285 (the same band, measured on Maya-XS; not yet re-measured on this base).

2026-10-28 15:07:48 +01:00

hip_compat

fix(HIP compat): cudaMemcpyPeer and cudaDeviceDisablePeerAccess for the #2666 peer-copy probe (peer_experts.cpp did not compile on HIP in f-145-int)

2026-10-16 22:39:02 +02:00

kernels

Merge f-028-par

2026-10-29 15:11:01 +01:00

ngram

UD-Q6_K_XL: Q6_K gate/up kernels behind -DSTRATA_Q6K_EXPERTS=ON; the Q8_0 PLE hunk is left to the #651 series

2026-10-05 19:58:13 +02:00

plan

Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install

2026-09-24 18:42:20 +02:00

platform

Build the engine for ARM64 (DGX Spark / GB10) without changing the x86 build (#3622)

2026-10-27 01:11:53 +01:00

prefill

Merge f-029-par-harness

2026-10-30 18:37:54 +01:00

program

Opt-in research switches: STRATA_PAR_COUNTERS, STRATA_PIPED_CONCURRENT_MIN, multi-prompt layer-major read (Prefill::run_multi, BCO co-reads, server STRATA_COALESCE_PREFILL), STRATA_CRASH_BT

2026-10-30 11:08:53 +01:00

qwen35

qwen35: a penalty for a history token outside the logit row no longer indexes past the row (#3911)

2026-10-26 18:07:29 +01:00

spec

spec: a gated request's lookup policy learns from committed rounds only, DraftPolicy choose() unchanged otherwise (#3416)

2026-10-27 01:45:36 +01:00