Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
Commits on main
GPU roles in strata launch (--gpu-roles, M/H/O/R/D Cards screen), setup (--gpu-roles, one roles question) and the host's TypeScript/JS rules; golden and setup tests follow the new defaults
8cfe44b Niko1221 committed 1 month ago
Metal: Gyro read-ahead skips the n-gram table, Q6_K/IQ4_NL hyper-connection matrices go through BF16 before their Q8_0 copy (as the pack does for CUDA)
1b47c1d Niko1221 committed 1 month ago
dsv41 cudalab: gfx12 WMMA port recipe of the trellis GEMV: host emulation of the data flow (matches the reference), HIP kernel source, layout probe
d251b2c Niko1221 committed 1 month ago
serve: DeepSeek-V4.1 through the server (deepseek-v3 pre-tokenizer, DSML tool calls, effort words, end token, card sampling)
5a52007 Niko1221 committed 1 month ago
Progress words: a file being repaired after a wrong SHA-256 says so in plain words
4dd6273 Niko1221 committed 1 month ago
dsv41 cudalab: computed (LUT-free) codebooks for the 15-bit trellis (1mad, 3inst, 3sum, 3single), encoder hook TQE_CODE, LUT-free GEMV kernel, code SQNR tool, pilot scripts from f-dsv41-enc
6f703a6 Niko1221 committed 1 month ago
metal/tests/gyro_parity.sh: the Gyro kernel checks in one script
f6b4101 Niko1221 committed 1 month ago
Hardware report: speed per NVIDIA card (SMs x clock) for the Main-card pick
d581286 Niko1221 committed 1 month ago
Gyro on the Mac: model list tests, resident wrap span without the n-gram table, trellis gate/up tile
0b736f7 Niko1221 committed 1 month ago
Server: gpu_roles (Main last, Off excluded, hand-edited gpu wins), roles log line, every-card-Off handling
b6825a4 Niko1221 committed 1 month ago
Downloads: every ranged answer is checked, a file with a wrong SHA-256 is repaired piece by piece instead of deleted
c0b0869 Niko1221 committed 1 month ago
GPU roles: one shared rule for Main / Helper / Off (recommend, set, order, mirror, warnings) with shared test vectors
2219cc9 Niko1221 committed 1 month ago
setup: FP4-Local conversion installs torch 2.14.1 on Python 3.14+ (2.9.1 has no cp315 CPU wheel), older Pythons keep 2.9.1 (#3474)
4681762 beplee committed 1 month ago
dsv41: pack_gguf takes unsloth's DSpark drafter GGUF (dflash) and the mmproj vision GGUF into the pack; the vision loader reads F32 rows and gate/up halves
2d718de Niko1221 committed 1 month ago
tests: Gyro-M per-tensor parity (--experts-first on a sparse copy), moe_mm_parity takes the trellis types
6b93b42 Niko1221 committed 1 month ago
Merge f-022-gpu0: strata launch keeps the installed card choice; engine/nvidia-smi number map; hip_ordinal and image gpu list fixes
d3cb2f5 Niko1221 committed 1 month ago
Keep the card a person chose: strata launch no longer re-picks GPU 0, the engine's GPU 0 is explained, old hip_ordinal and list GPUs for images
9bb1dd2 Niko1221 committed 1 month ago
dsv41: split-KV attention opt-in (DS4_ATTN_SPLIT=1): A/B on 4x R9700 shows no decode gain, KL unchanged
0138920 Niko1221 committed 1 month ago
Metal: Gyro in the Mac setup and app model lists, the n-gram table stays out of the dense copy, decode-check test
c3472ca Niko1221 committed 1 month ago
SYCL Gyro: APR pack log line in generate, second-GPU helper declines Hadamard-folded experts
f1bfcc2 Niko1221 committed 1 month ago
Metal: Gyro engine message fix, expert_parity prints the pipelines' limits (PIPE_INFO)
a64f1dd Niko1221 committed 1 month ago
dsv41: split-KV sparse attention from the CUDA lab for decode / short verify windows (packed rows, T<=5), DS4_ATTN_SPLIT=0 off
ff5577a Niko1221 committed 1 month ago
dsv41: --score writes end_line in the segment file (quant_quality.py compare reads it)
d17c32d Niko1221 committed 1 month ago