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
launch: the Access screen's help names --show-key and --api-key, so a person can see or set a key of their own (#3713 #3951); test
aef053e Niko1221 committed 26 days ago
missing files: name the engine program in words once, not as component:strata-engine (#3900)
c22e5e7 Niko1221 committed 26 days ago
chat model picker lists downloaded optimized models that are not loaded yet (#3814)
97c3116 Niko1221 committed 26 days ago
Chat: Copy response/message/code say Copied or Couldn't copy and fall back to the older copy route (#3731); a pasted screenshot is also read from clipboard items (#3748); the user bubble keeps its line breaks (#3720)
5faf357 Niko1221 committed 26 days ago
update: the test start runs the new engine with a short context and no picture reader (#3842), and an update keeps the CUDA build that runs here (#3835); launch: the summary names the engine log (#3656) and a tuned model says how to measure again (#3874); tests
77284b5 Niko1221 committed 26 days ago
monitor: read Windows disk counters from one long-lived PowerShell instead of a cold one every 3 s (#3939)
4b49729 Niko1221 committed 26 days ago
App update: log lines for every update step, an alert when the app came back as the old version, and Update engine works before the native snapshot has connected
014fd6b Niko1221 committed 26 days ago
setup: the build tools come from the virtualenv first and an older system cmake (below the 3.24 CMakeLists.txt needs) counts as missing, so pip installs the right one (#3822); tests
5bfce4b Niko1221 committed 26 days ago
In-app update: a finished app update is no longer shown as an interrupted operation; a test update feed used once is not remembered as the feed for later checks
15fdb69 Niko1221 committed 26 days ago
own model files: a qwen4exp file set that is no tested build but has the layout of one is used as that model with a plain notice, and setup keeps a complete GGUF of another size or hash with a warning instead of stopping (#3651 #3777); tests
ad9bb57 Niko1221 committed 26 days ago
Adaptive tier: the engine chooses the asynchronous tier by measurement (arena, one CUDA GPU)
b18700c Niko1221 committed 26 days ago
UPDATE.bat / update.sh: an engine that stayed older than the minimum is said as 'the engine was NOT updated' with one next step, not '[ok] Strata is updated' (#3653); test
c02a0b8 Niko1221 committed 26 days ago
strata connect: scale the app configs' compaction point with the server's context
7a03fc9 demetree committed 28 days ago
setup: a Mac with an Intel (Rosetta) Python is still an Apple Silicon Mac; the chip name and hw.optional.arm64 decide, one plain line says the Python is an Intel build (#3894); test
d924bef Niko1221 committed 26 days ago
launch: a setup option error ('setup.py: error: argument --kv: invalid choice') is the cause shown, not only 'could not be installed'; setup accepts --kv fp16 as DOCKER.md says (#3785); tests
bdaa93e Niko1221 committed 26 days ago
setup: a NVIDIA driver older than CUDA 13 needs (528 to 579) runs the CUDA 12 engine by itself and says so in one line instead of stopping and naming a flag strata launch does not have (#3843); tests
1736b55 Niko1221 committed 26 days ago
Windows: a card taken out of the PC no longer decides the engine; only the display adapters Windows reports as present count when two makes are in the registry (#3892); test
0cd4432 Niko1221 committed 26 days ago
metal: finish dst stride, BF16 ring pack, residency on the context queue, split the command buffer after fp8 finishes; ds4_launch.h makes the CUDA build work again; more DS4_DUMP_S points
1935292 Niko1221 committed 26 days ago
DeepSeek engine: the RAM store's background fill is opt-in (DS4_STORE_BG=1): ready 9 s sooner but the first prompt waits for the same SSD and decode measured 3.7% lower
e3eeb93 Niko1221 committed 26 days ago
DeepSeek engine: the RAM store's fill beyond the first 48 GB is read in the background after the model is ready (hottest first, behind the model's own reads); DS4_STORE_BG=0 reads it all first
58f2acc Niko1221 committed 26 days ago
DeepSeek CPU lane: compute the store hits while the misses are read, each missed expert as its read lands (DS4_MISS_SPLIT=0 restores the wait-for-all order); ExpertStore::ensure split into begin / wait_slots
6fcde7e Niko1221 committed 26 days ago
DeepSeek engine: the automatic RAM store may take every routed expert, not 48 GB at most, where the free RAM holds them (1 R9700 + 251 GB, 2.00 set: 12.6 -> 21.4 tokens/s); free - 12 GB and the explicit sizes unchanged
b3c86c7 Niko1221 committed 26 days ago
metal: bind the allocation's own MTLBuffer instead of a no-copy alias per call (the alias read stale data), deferred frees, matvec jobs cut at 8 activations, Q8/quant debug checks
9f45341 Niko1221 committed 26 days ago
DeepSeek engine: the automatic RAM store is sized again after the weights are on the GPU, and keeps room for the context's buffers, so a unified-memory PC (Strix Halo) keeps 12 GB free; explicit sizes unchanged
f1e41cd Niko1221 committed 26 days ago
ds4: DS4_DUMP_S sub-step dump inside the layers (same records on CUDA and Metal), DS4_FLUSH_EACH, per-layer debug prints
da9cf63 Niko1221 committed 26 days ago
metal: both queues drain at every sync, engram dequant wraps the strided extent, the ggml CPU lane is in strata_ds4
0c34a7d Niko1221 committed 26 days ago