Commit Graph

  • dbb2d168cf When the GPU reader did not start for any other reason, the summary says pictures are read on the processor main Niko1221 2026-11-10 16:10:16 +01:00
  • 7e8206b501 The picture reader's need in the sentences is its whole need at its image limit, not the one allocation that failed Niko1221 2026-11-10 15:04:36 +01:00
  • c53e45a35d The picture reader falls back to the processor in every mode instead of ending the launch Niko1221 2026-11-10 14:52:03 +01:00
  • 78eaaf1b01 A picture reader that cannot start for lack of GPU memory is reported as that, with the card, the size and the next step Niko1221 2026-11-10 14:41:16 +01:00
  • 4cce1ee4d0 The picture reader's READY pipe: one deadline, stray lines kept and logged, a printing reader is never killed Niko1221 2026-11-10 14:11:52 +01:00
  • 38ad0bcbd1 The product never starts a debugger: every child it starts gets GGML_NO_BACKTRACE=1 Niko1221 2026-11-10 16:14:44 +01:00
  • 42d502a066 Snapshot driver: a command run in the test process does not change the environment the engine's is compared with Niko1221 2026-11-10 16:10:06 +01:00
  • 8ac77c06c4 DIAGNOSTIC (ci/checks-green only, drop before main): macOS getfqdn and child server timing (stdlib, no-lookup, serve/server.py); Windows PowerShell 5.1 children of node and python parents ci/checks-green Niko1221 2026-11-10 10:41:48 +01:00
  • 9d11bc053e test_cfg_paths, test_component_start: the starts give the stub server 60 s, not 15-20 (a pass still returns as soon as it is ready; on a Windows PC with about 330 processes and a non-elevated account a start took 16 s against a 15 s wait) Niko1221 2026-11-10 13:17:49 +01:00
  • 7726f80cba test_cfg_paths, test_component_start: the starts give the stub server 60 s, not 15-20 (a pass still returns as soon as it is ready; on a Windows PC with about 330 processes and a non-elevated account a start took 16 s against a 15 s wait) Niko1221 2026-11-10 13:17:49 +01:00
  • 8bbd64e971 parallel-chats test: the scratch folder's removal retries (a chat save that landed while it was being removed failed the teardown with ENOTEMPTY on a macOS runner) Niko1221 2026-11-10 12:33:28 +01:00
  • 82ab8ac762 test_recipes: the interrupted run is found by excluding the two known runs, not as the newest by file time (three runs in one tick of a coarse Windows clock came back in any order) Niko1221 2026-11-10 12:33:05 +01:00
  • 79d529cc78 procs.orphans_of on Windows: names first, a command line only where the image name can be the engine's file (reading every command line took 4-9 s per server start on a Windows PC with about 280 processes and a non-elevated account; now 0.1 s warm, 2 s cold) Niko1221 2026-11-10 10:40:46 +01:00
  • 3e1fcaaf67 serve: the HTTP server binds without a host name lookup (http.server's server_bind asks getfqdn, which stalled 35 s per start on GitHub's macOS runner; nothing reads server_name) Niko1221 2026-11-10 10:41:47 +01:00
  • 42e97db851 test_runtime: the start, restart and verify waits give the stub server 60 s (operations 80 s), not 10-20 (a pass still returns as soon as it is ready; on a Windows PC with about 280 processes and a non-elevated account, the server's leftover-engine scan alone took 4-9 s and the budget ran out) Niko1221 2026-11-10 10:30:56 +01:00
  • 35ad5b4716 test stubs: mcp_fake_server and cutserver bind without the getfqdn lookup (they run as fresh child processes) Niko1221 2026-11-10 10:26:18 +01:00
  • c7b446fa85 test_setup_feed_models TerminalAndApp: the table's Strata reads a fixed NVIDIA PC, not the card of the PC the tests run on (on a PC with an Intel Arc card, install_plan went down the Intel path) Niko1221 2026-11-10 09:39:02 +01:00
  • 50c753f5d5 host tests: assert-private reads the DACL through .NET File.GetAccessControl and compares it by SID (Get-Acl's module did not load in Windows PowerShell under a PowerShell 7 parent; SDDL writes the runner's built-in Administrator as the alias LA) Niko1221 2026-11-10 10:41:47 +01:00
  • b250bc78a9 test_drafting_tuned: a same-size rewrite moves the file clock on before the marker check, as test_drafting_selector_q4 does (same tick on Windows kept mtime_ns) Niko1221 2026-11-10 09:33:51 +01:00
  • ea6826cc83 test_setup_golden: the config a run wrote is the new or changed file, not the newest time (same clock tick on GitHub's Windows runner picked the config placed before; setup wrote the named size) Niko1221 2026-11-10 09:33:51 +01:00
  • e906e7e979 test_strata_mcp: the fake model server binds without http.server's getfqdn lookup (a fresh child on GitHub's macOS runner did not answer within 30 s) Niko1221 2026-11-10 09:33:51 +01:00
  • 13e7188cc8 images: the answer's background says what the picture is: "transparent" also when a picture kept its alpha because it could not be put on white (the logged fallback), "opaque" for a white or an RGB sd-cli picture (_flatten reports whether the returned PNG has an alpha channel) Niko1221 2026-11-10 13:09:20 +01:00
  • c62a656b9e parallel-chats test: the scratch folder's removal retries (a chat save that landed while it was being removed failed the teardown with ENOTEMPTY on a macOS runner) Niko1221 2026-11-10 12:33:28 +01:00
  • 74577cb0f1 test_recipes: the interrupted run is found by excluding the two known runs, not as the newest by file time (three runs in one tick of a coarse Windows clock came back in any order) Niko1221 2026-11-10 12:33:05 +01:00
  • 3a0916c5e7 docs/IMAGES.md: the background request field, the answer's background and strata.flatten_s, and the picture on white (the API's PNG is no longer byte for byte sd-cli's unless background is transparent) Niko1221 2026-11-10 12:21:20 +01:00
  • 631b280308 images: pictures are put on white by default. Qwen-Image-2.1 decodes RGBA and leaves some backgrounds transparent over a flat key colour; the server now composites each picture on white and returns an 8-bit RGB PNG (flatten_png: zlib 6, every chunk other than IHDR / IDAT, sd-cli's "parameters" text above all, copied byte for byte; 0.12 s median, 0.22 s max per 1024-1536 picture with Pillow, 0.62 / 1.09 s on the zlib + numpy fallback). background: "transparent" keeps sd-cli's PNG unchanged; opaque / auto / none give the white one; other values are refused. A PNG that cannot be rewritten goes out unchanged with one log line. The response carries background and strata.flatten_s Niko1221 2026-11-10 12:21:15 +01:00
  • ea16bfd306 images: the text-encoder time reaches /v1/status and the response again: stable-diffusion.cpp 2988060 (the pinned commit) prints "get_learned_condition completed, taking 0.70s", the pattern only read the "N ms" form some older revisions printed (from b636886, 2024-02); both forms are read now (log_figures), the fake sd-cli prints the pinned form Niko1221 2026-11-10 12:01:16 +01:00
  • 102c9e722e procs.orphans_of on Windows: names first, a command line only where the image name can be the engine's file (every command line took 4-9 s per server start over ~280 processes; now 0.1 s warm, 2 s cold) Niko1221 2026-11-10 10:40:46 +01:00
  • f6a067cfa9 serve: the HTTP server binds without a host name lookup (http.server's server_bind asks getfqdn, which stalled 35 s per start on GitHub's macOS runner; nothing reads server_name) Niko1221 2026-11-10 10:41:47 +01:00
  • 51e73c2064 test_runtime: the start, restart and verify waits give the stub server 60 s (operations 80 s), not 10-20 (a pass still returns as soon as it is ready; on the owner PC the server's leftover-engine scan alone took 4-9 s and the budget ran out) Niko1221 2026-11-10 10:30:56 +01:00
  • 84477224a7 test stubs: mcp_fake_server and cutserver bind without the getfqdn lookup (they run as fresh child processes) Niko1221 2026-11-10 10:26:18 +01:00
  • e3e2cbfd23 test_setup_feed_models TerminalAndApp: the table's Strata reads a fixed NVIDIA PC, not the box's own card (p1006's Arc sent install_plan down the Intel path) Niko1221 2026-11-10 09:39:02 +01:00
  • f8ec0dc28f host tests: assert-private reads the DACL through .NET File.GetAccessControl and compares it by SID (Get-Acl's module did not load in Windows PowerShell under a PowerShell 7 parent; SDDL writes the runner's built-in Administrator as the alias LA) Niko1221 2026-11-10 10:41:47 +01:00
  • ea1ef8b58b test_drafting_tuned: a same-size rewrite moves the file clock on before the marker check, as test_drafting_selector_q4 does (same tick on Windows kept mtime_ns) Niko1221 2026-11-10 09:33:51 +01:00
  • 0aae1ee2e6 test_setup_golden: the config a run wrote is the new or changed file, not the newest time (same clock tick on GitHub's Windows runner picked the config placed before; setup wrote the named size) Niko1221 2026-11-10 09:33:51 +01:00
  • ac0992e533 test_strata_mcp: the fake model server binds without http.server's getfqdn lookup (a fresh child on GitHub's macOS runner did not answer within 30 s) Niko1221 2026-11-10 09:33:51 +01:00
  • 49bf891e72 make_model_release --checklist <file>: the internal release checklist by path, filled beside the repo's default notes and model card (no second copy of the public templates in a private folder); a missing file is refused before anything is written Niko1221 2026-11-10 08:48:04 +01:00
  • 5ce203f59a make_model_release: the release templates (the notes and the model card) live in the repo, packaging/templates/model-release, and are the default of --templates (another folder can still be named); the templates test reads them on every box (the internal release checklist stays out of the repo) Niko1221 2026-11-10 08:45:57 +01:00
  • 87dfca0e7c test_strata_mcp MacCatalog: the second temp data folder (f-0214-app's disk-short check) under the real home on macOS too Niko1221 2026-11-10 05:43:54 +01:00
  • 0abb4f869e Merge f-clean-fix: CI fixes and tests that no longer depend on the box they run on Niko1221 2026-11-10 06:47:31 +01:00
  • 663f5a5b83 Merge branch 'main' into f-0214-app Niko1221 2026-11-10 06:27:13 +01:00
  • c47df69e01 test: the phaseA-tests templates check gives --min-engine as a real case-A caller must, and first checks that without it the script refuses (078b52563 made it required; the check only runs on the owner PC, so it was not run after that) Niko1221 2026-11-10 06:16:39 +01:00
  • 1b30def0ed Merge commit '562fa312e' into f-0214-app Niko1221 2026-11-10 06:02:22 +01:00
  • 562fa312ec test: hermetic drafter-fetch test and no pytest tmp_path in the 0.2.14 tests (Windows gate of f-0214-app) Niko1221 2026-11-10 05:57:01 +01:00
  • e3a04c013f test_strata_mcp MacCatalog: the temp data folder under the account's real home (pwd), not HOME, which the fake root points at a temp folder Niko1221 2026-11-10 05:42:45 +01:00
  • 872a7bcb28 test_strata_mcp: the macOS expectedFailure marks come off (lane 3's zombie fix, f-0214-app, lands first: an unexpected success would fail) Niko1221 2026-11-10 05:33:51 +01:00
  • b6a7915c89 serve/test_server.py: LF line ends again (the hip_ordinal edit had written the whole file with CRLF) Niko1221 2026-11-10 05:32:39 +01:00
  • 0645f2edec tests: suites that read the box they run on (Boston, p2006, p1006, macOS) Niko1221 2026-11-10 04:21:43 +01:00
  • 804d2032d2 tests: a nvidia-smi that answers nothing comes first on PATH for the Python suites on POSIX, so a box with an NVIDIA card runs them as GitHub's runners do (p1004's P100: 64 stub-engine update tests failed against the stand-in packages' sm_86 code) Niko1221 2026-11-10 03:41:30 +01:00
  • 1c50fe076c make_model_release: the plan prints the exact strata launch lines with the size name a person types (<quant>-<id> when the quant's name is a built-in size, as feed_models.py decides) Niko1221 2026-11-10 02:40:41 +01:00
  • 9e50303078 feed_tool: an OPT-IN multi-threaded gzip for tar.gz packages (STRATA_PACK_THREADS >= 2: 4 MiB gzip members compressed in parallel, the same bytes for the same tree); unset, the archive is the single-stream tarfile gzip of before. Published 0.2.12 / 0.2.13 Terminals, install.sh and the 0.2.13 AppImage's Control install from such a feed (not the default until every reader is shown) Niko1221 2026-11-10 02:40:41 +01:00
  • 1eac64cd5d Merge commit '8cf002e1b' into f-0214-app Niko1221 2026-11-10 02:38:04 +01:00
  • 8cf002e1b0 test: drafting.test.mjs's second source check reads the first-load condition too Niko1221 2026-11-10 02:37:28 +01:00
  • c8d1d73122 Merge commit '95e9b7341' into f-0214-app Niko1221 2026-11-10 01:51:44 +01:00
  • 95e9b7341f setup: the first range probe of a parallel download is tried again after a hiccup (proposal, from lane 8) Niko1221 2026-11-10 01:49:32 +01:00
  • 7ea172159a strata_mcp: without /proc (macOS) an ended child that is not reaped yet is not "running" Niko1221 2026-11-10 01:49:32 +01:00
  • e481f3b6b9 Control: a model whose draft layer is off keeps the server's own "without the draft layer" step; the app's first-load test shape Niko1221 2026-11-10 01:49:32 +01:00
  • 7fddfd8292 Merge commit '81f7422c1' into f-0214-app Niko1221 2026-11-10 01:49:20 +01:00
  • 8793c8f85c Merge branch 'main' into f-0214-app Niko1221 2026-11-10 01:48:08 +01:00
  • a9d52eae2f Merge branch 'f-0214-appwin' into f-0214-app Niko1221 2026-11-10 01:42:03 +01:00
  • 81f7422c11 Desktop (Linux): the restarter closes inherited files in the new app only, not in itself Niko1221 2026-11-10 01:32:28 +01:00
  • a846190347 tests: the vision-build tests name an AVX2 CPU (a box without one, or a Mac, rebuilt the engine with an ISA floor); the feed command-line tests run setup.main() as on Linux on a Mac Niko1221 2026-11-10 01:27:18 +01:00
  • e9239a87dc tools tests on macOS: only setup.py and strata_mcp see linux (a module-level sys view), the CPU reads x86_64; the three MCP install-job tests are expected failures on macOS (a zombie child reads as running there: product bug, reported) Niko1221 2026-11-10 01:18:48 +01:00
  • cdbb312bbc test_setup_download: one stream means one stream without setup's one-byte probes (a server without ranges answers the probe with the whole file, which can still be sending when the stream starts: 2 of 10 runs on p2006) Niko1221 2026-11-10 01:16:54 +01:00
  • 99b3dc6243 tools/conftest.py: on macOS each tools test runs as on the Linux x86 runner (sys.platform, platform.system/machine, sysinfo.machine), not only setup.py's view; the Apple CPU read as ARM64 and strata_mcp took the Mac path Niko1221 2026-11-10 01:10:55 +01:00
  • 737ab4f262 test: the Strix app-vs-launch reserve test compares the two engine lines with each scratch world's folder written as <ROOT> Niko1221 2026-11-10 00:44:58 +01:00
  • 2e25954e57 ci: a diagnostic step on the Windows host job shows what Windows PowerShell 5.1 children return under a PowerShell 7 parent (hardware probe, setup cpu_info) Niko1221 2026-11-10 00:32:25 +01:00
  • f328c2cd27 test_setup_download: a killed download may have taken the one-stream path (one failed range probe); its .part is then read as finished pieces, as setup itself reads it Niko1221 2026-11-10 00:32:25 +01:00
  • 78451ea772 test: the Strix reserve field test of the hardware report no longer patches sys.platform (it broke shutil.which on Linux) Niko1221 2026-11-10 00:29:53 +01:00
  • 3801ec40a7 SYCL router: an fp32 router for a card without fp64 (Intel Arc A/B-series) - a model of another shape (Qwen4: top-8, 256 experts) routes through the generic router, whose double kernel the Arc runtime refused ('double' is not supported), ending the engine at its first token; a card with fp64 keeps the double kernels; the parity test covers the Qwen4 shapes and compares fp32 with double on the same inputs Niko1221 2026-11-10 00:16:36 +01:00
  • 8cfd842c72 tests: pass on GitHub's macOS runners Niko1221 2026-11-10 00:15:52 +01:00
  • 46e84e2398 Intel Arc: a model of the signed feed starts (Qwen4 day) - the Terminal's and the app's Arc run config takes a feed qwen4exp model (Unsloth layout, its one GGUF or its shards, no Flash-Next draft layer, drafter none -> drafting off, the neutral expert profile for another shape) and a model of its own shape on its engine component (component:strata-q4-<shape>; the app's Arc setup fetches that engine first); the built-in families' configs are unchanged Niko1221 2026-11-10 00:02:17 +01:00
  • 078b525636 make_model_release: --title (the model's title in the entry and the notes; default the file's general.name, as before), and case A with --write needs --min-engine instead of leaving [FILL: minimum engine version] in the entry Niko1221 2026-11-10 00:02:16 +01:00
  • b2f2684035 Shape engine: one that cannot answer strata --geometry is refused at install (owner 11-09: "it does not say which model shape it was built for" inside the approved start_failed line); the Strata engine is never asked Niko1221 2026-11-09 23:46:48 +01:00
  • 7937614e34 Desktop (Linux) + App: after the in-app update the update job reads done, and no old AppImage runtime is kept alive Niko1221 2026-11-09 23:43:01 +01:00
  • 8b92e9ace8 Packaging: SHAPE=<geometry> in the Linux engine build scripts (CUDA x86_64 and aarch64/Spark, HIP, SYCL): the engine of a model shape in one command per platform - configured for that geometry (or the header STRATA_CMAKE_EXTRA names), checked with strata --geometry, packaged with --shape, no DeepSeek engine in it; without SHAPE every command is unchanged Niko1221 2026-11-09 23:32:59 +01:00
  • 75a75e1a87 Settings > Engine, shape group Use dialog: keep the two reassurance sentences after the new first one Niko1221 2026-11-09 23:22:04 +01:00
  • b130a479dc Settings > Engine: the Use dialog of a Qwen4 model shape's engine says "{model} runs on engine {version} from its next start." Niko1221 2026-11-09 23:21:05 +01:00
  • 77f7835a8b SYCL engine: strata --geometry and strata --families, as the CUDA/HIP/Metal program answers them (the Intel program is its own main; without them a SYCL shape engine was installed with no geometry check) Niko1221 2026-11-09 22:03:41 +01:00
  • b5dc27e388 Desktop (Linux): the app comes back by itself after an in-app update Niko1221 2026-11-09 22:49:42 +01:00
  • 519c7e8c08 Control: a feed model whose entry says drafter: none starts with no drafter, and the Terminal plans none for it Niko1221 2026-11-09 22:49:42 +01:00
  • b473875485 App: the first-run engine card also takes a newer Strata Control from the feed, and the app switches to it Niko1221 2026-11-09 22:49:42 +01:00
  • af9b832fd3 App: the setup prepares the drafter Control's Auto picks for this card; the first load after it fetches a missing MTP layer Niko1221 2026-11-09 22:49:42 +01:00
  • 06f3beb0ca App: an engine version that failed its test start is not "up to date"; Update engine tries it again Niko1221 2026-11-09 22:49:41 +01:00
  • fdfb2db6de App + Control: run-config paths into a gone Control folder (AppImage /tmp/.mount_*) are repaired before updates and loads Niko1221 2026-11-09 22:49:41 +01:00
  • 9c09f839c3 App: exit 7 from Control's update lock is "an update is already running", and the update waits for the lock Niko1221 2026-11-09 22:49:41 +01:00
  • 28d6695774 Settings > Engine and strata components engines list the engine of a Qwen4 model shape as its own group, with Use and Delete Niko1221 2026-11-09 22:31:43 +01:00
  • afe6c8bc14 Strix reserve, follow-up: with no setup reserve in the run config the app takes the one setup would write (Control's setup_reserve_mib), else the floor of before Niko1221 2026-11-09 22:23:00 +01:00
  • 1f2f737e8c test_setup_pins: the stand-in pip stamp carries this system's marker, so the Requirements tests do not depend on the Python that runs them (pyyaml missing made the stamp unbelieved) Niko1221 2026-11-09 22:18:52 +01:00
  • ecccc359dd A Models-page Load on a unified-memory chip (Strix Halo) keeps the reserve setup wrote, as strata launch does Niko1221 2026-11-09 22:17:46 +01:00
  • 731b40fedc The This computer card places a unified-memory chip's model with the label's rule (carve-out only beside the RAM) Niko1221 2026-11-09 22:14:57 +01:00
  • db35f76272 test: Settings > About still shows an interrupted or failed update with no later finished one (also while a retry runs) Niko1221 2026-11-09 22:14:57 +01:00
  • 1eec11bd85 Read pictures unticked holds across a restart on a model set up without the picture files: a choice made with a reader present carries a marker Niko1221 2026-11-09 22:13:44 +01:00
  • fb721c37d8 The version list's fit word counts a unified-memory chip (Strix Halo) only for its carve-out, as Control's fit_label does Niko1221 2026-11-09 22:12:38 +01:00
  • 87ec2e7734 Settings > About checks for updates again after an in-app Strata or engine update finished, so the row no longer keeps "Update available" Niko1221 2026-11-09 22:11:18 +01:00
  • 8c98f9fdb9 Settings > About no longer shows an update failure that a later finished update of the same component made right Niko1221 2026-11-09 22:11:07 +01:00
  • 0a6a5ad254 Menus stay inside the window on WebView2 before 128 at Interface size 125% and up: the window's size is measured like the rectangles (a fixed box), not taken from innerWidth/innerHeight Niko1221 2026-11-09 22:10:41 +01:00
  • beed8a5174 tests: pass on GitHub's Windows runners (CRLF, PowerShell 7 module path, classic console, coarse clocks) Niko1221 2026-11-09 22:07:36 +01:00
  • 8e69b84dd6 gitattributes: chat templates (*.jinja) are LF in every checkout Niko1221 2026-11-09 22:07:36 +01:00
  • 80cde871b9 ci: the macOS check jobs keep their logs and upload them on a failure, as Linux and Windows do Niko1221 2026-11-09 22:05:44 +01:00