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
WSL on AMD: setup and launch say plainly that the AMD (HIP) engines do not run under WSL yet (no /dev/kfd, the GPU bridge never finishes a command) and name the Windows app or Terminal as the next step; a warning, nothing is refused; tests
7025a99 Niko1221 committed 25 days ago
Windows uninstaller: stop the app, host, node and engine that run from the install folder before files are removed, so the program folder no longer stays behind after use; only processes whose image path is inside the install folder; test with a real process
36c6dd9 Niko1221 committed 25 days ago
Merge f-026-copyall (DeepSeek copy-all CPU-lane experts) into f-026-rel
4a6a5fb Niko1221 committed 25 days ago
metal: no command-buffer split after the fp8-rounding finishes (decode +22%: code 11.9 -> 14.6, agent 9.7 -> 12.0 tok/s); DS4_FINISH_SPLIT=1 restores it
1ece7ef Niko1221 committed 25 days ago
Mac: the engine bundle inside a signed .app is not removed after the first start (deleting it broke the app's signature); outside an .app and on other systems the cleanup is unchanged; test
721c6f7 Niko1221 committed 25 days ago
Monitor/serve: a one-token reply has no decode span, so no tok/s is shown for it (a 1-token probe read 1,250 tok/s); test
5d5c26d Niko1221 committed 25 days ago
serve: a chat request with no message is a 400 (messages must be a non-empty list), as OpenAI answers it; test
8a3ca07 Niko1221 committed 25 days ago
About/Updates: never show the host's placeholder engine version (strata-engine 0.1.0) while the real one is still being read; a forced status read that meets one already running waits and reads again, so the version after an engine update is not the old one; test
dd9b9f7 Niko1221 committed 25 days ago
Storage and Manage models: a model whose files are all gone (old config, 0.0 GB, not ready) is no longer listed as a download
4d67218 Niko1221 committed 25 days ago
Models screen: DeepSeek is installed and shown at its one 8K context (as the guided setup), not the 32K default
ad66a75 Niko1221 committed 25 days ago
Monitor: All GPUs leaves out a small integrated Radeon/Intel chip beside a real card (totals and count); the hardware report gives each card a stable id since each make numbers from 0; tests with the RTX 5070 + Radeon iGPU pair
59599f6 Niko1221 committed 25 days ago
metal: sparse attention on simdgroup matrices is the default for 8-head groups (32K prompt 269 -> 293 tok/s, KL equal to the scalar kernel); DS4_ATTN_MX=0 for the old one
2a207db Niko1221 committed 25 days ago
WIP metal: ds4_sparse_attn_mx (sparse attention on simdgroup matrices, DS4_ATTN_MX=1; =2 compares with the scalar kernel: rms rel 1.8e-6); timing and KL gate pending
b6dcb23 Niko1221 committed 26 days ago
metal: Q5_K dequant four values a step (bit-identical); DS4_SKIP_K and the timeline's overlap count as debug tools
100a237 Niko1221 committed 26 days ago
UPDATES.md: the Mac dmg keeps its Metal engine; only the Windows and Linux installers are engine-less
36f580a Niko1221 committed 26 days ago
metal: prompt hc_mix reads each token's row once for four mixes (bit-identical, +4.5% on the 3K prompt)
8810e00 Niko1221 committed 26 days ago
DeepSeek Metal: head matvec back to 4 columns a pass (8 moved the KL by +0.002)
050752f Niko1221 committed 26 days ago
DeepSeek Metal: prompt flow reads experts straight into the ring (unified memory: no staging, no DMA), head matvec 8 columns a pass
97c091c Niko1221 committed 26 days ago
DeepSeek Metal: cached compute pipelines (the prompt path created one per expert copy), F_NOCACHE for the unbuffered expert reads, dispatch-seq log
6b2a530 Niko1221 committed 26 days ago
Merge f-0243-dsimport (the app's model-folder preview recognises official DeepSeek sets) into f-0243-rel
ad638b9 Niko1221 committed 26 days ago
Model preview: recognise official DeepSeek-V4.1-Flash IST-DAS sets (1.75/2.00/2.25 bpw, official and first-release names, with or without the Engram) as family deepseek41
32ad109 Niko1221 committed 26 days ago
DeepSeek Metal: GPU stale-router lookahead prefetch (opt-in DS4_LOOKAHEAD; measured slower), recall counters
e527bfe Niko1221 committed 26 days ago
DeepSeek Metal: CPU-lane page prefetch in parallel across the call's matrices (the cold reads were 78% of the lane), CPU-lane phase log
9e3fe04 Niko1221 committed 26 days ago
collect_libs: a library an earlier binary's pass already put in --out is not copied onto itself (SameFileError stopped the AMD packages' second pass)
6270e2d Niko1221 committed 26 days ago
Merge f-0250-tui-uma (Faster is the default Speed on shared-RAM GPUs) into f-0243-rel
ce27d13 Niko1221 committed 26 days ago