Niko1221 ce6bbd7d54 DFlash2 on Intel Arc: the tuned bf16-selector pack for SYCL + IQ3_XXS at Arc's own confidence 0.6 ...

The SYCL 0.2.11 engine reads only bf16 selector codebooks, so Arc gets its own tuned file (qwen3.8-flash-next-dflash2.strata1.q4_0-bf16sel.pack, 522,458,112 bytes, own SHA-256 pin) through
SetDefault.pack. Only IQ3_XXS, the set measured on the A750 (tuned 0.6 x1.054, 3 of 3 rounds); other Arc sets keep the stock pack. SYCL never gets the selq4 file, CUDA / HIP / Metal never the
bf16sel one. NOTICE names the second file and no longer overclaims about the text.

(cherry picked from commit 6111c583b70e3c88a3d54039e13d76b19f15742d)

2026-11-06 15:47:42 +01:00

..

dflash2-tuned

DFlash2 on Intel Arc: the tuned bf16-selector pack for SYCL + IQ3_XXS at Arc's own confidence 0.6

2026-11-06 15:47:42 +01:00

dsv41

DS Coder points at the official repository ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF, revision 0b1d5da5, every file's size and LFS sha256 from the Hub: five shards in IQ2_XXS/, the Engram at the root is the shared Engram (same name, size and SHA-256 as the unpruned release's converted one, used where it is, on the SSD), the repository's mmproj is accepted beside the shared one; tests pin the Engram rule

2026-11-02 15:40:48 +01:00

experimental-speed-projection

Experimental speed projection: an optional control vector (EXPERIMENTAL, off by default)

2026-09-27 16:47:45 +02:00

quant_quality/v1

Formats cost nothing unless used: Q6_K gate/up experts in the normal build; the K-quant MMQ prompt kernels on by default, on CUDA as their own library (strata_kq) opened only for a model with these experts and STRATA_MMQ_KQ=1

2026-10-12 21:04:17 +02:00

draft_vocab_cyrillic.bin

setup: --draft-vocab cyrillic - the English/code subset plus the whole Cyrillic script

2026-10-01 08:13:17 +02:00

draft_vocab_en.bin

setup: --host/--api-key/--draft-vocab on a start are kept for the model (#179); --draft-vocab en keeps the English/code subset (~110 MiB less VRAM); CUDA 13.0 for RTX 50 builds (#220); llama.cpp's web UI is not unpacked (Windows path limit, #206); Swift Q2_0 hidden (#171); the earlier data folder and nested Strata-data move into --data-dir (#198); whole shards without a finish mark count (#173)

2026-09-30 10:54:45 +02:00

draft_vocab_fr.bin

setup: --draft-vocab fr - the English/code subset plus the tokens of French text (#597)

2026-10-03 18:19:49 +02:00

draft_vocab.bin

data/draft_vocab.bin: the draft subset with all Chinese, Japanese and Korean tokens (40,525 -> 106,299 ids; #137). Q2_0 on CJK answers +15-38% decode, English -1-2%, the same expert slots on 12 GB

2026-09-29 22:58:59 +02:00

expert-profile-coder.bin

The GSQ-RCO Coder (256 of 512 experts): setup, profile, speeds (PR #54)

2026-09-28 14:20:58 +02:00

expert-profile-sycl.bin

SYCL: data/expert-profile-sycl.bin, the experts' ranking for a card on xe, built with make_profile_mix.py from the routing of 45 varied prompts and eight document summaries; setup_intel writes it instead of the shipped profile there (i915 keeps the shipped one). B70: +5..64% decode (mean +21%) on 12 unseen prompts, the same tokens. The copy-engine queue is Battlemage-only by default (it hung the A750's prompt)

2026-10-11 08:25:30 +02:00

expert-profile.bin

The PCIe share by the copy kernel for every pack (#31, #29); setup and profile fixes (#46, #49, #45)

2026-09-28 11:02:44 +02:00

gguf_fingerprints.json

catalog: Unsloth v3.1 Qwen3.8-Flash-Next sets (UD-IQ1_M, UD-Q2_0, UD-IQ2_XS, UD-Q2_K_XL, UD-IQ3_XXS, UD-Q3_K_XL, UD-IQ4_XS-v31) at d9620119 with a per-model revision; UD-Q4_K_XL's four-shard layout accepted from local files; manual-load refusal lifted (UD-IQ1_S stays refused); mac_fit switch for the Q5_0 down sets; header-only fixtures and tests (about lines pending wording)

2026-11-05 07:23:41 +01:00