Strata plugin 0.8.1

Strata 0.8.1

Updates the plugin for DSH 0.2.2 and Cordis 4.1. The self-contained package now installs and verifies the pinned Strata 0.1.55 / runtime 0.8.0 components before packaging them; it does not restamp the legacy 0.5.0 archive as a new runtime.

Includes the Cores kept free setting, Off reasoning support, and Medium defaults for Qwen profiles. Preserves the 0.8.0 managed runtime migration and native DeepSeek effort choices. Model files and saved settings remain owned by the user’s DSH home.

Strata plugin 0.7.1

Strata 0.7.1

Fixes the offline package's shared-runtime version marker. Both GitHub URL installation and the self-contained archive retain runtime 0.5.0 independently of the plugin version. Use this release instead of the 0.7.0 offline archive.

Portable configs can now import prepared Flash-Next NVFP4 GPTQ and NVFP4/Q8 Mix packs. Import finds each component by exact byte size and MD5, restores the pack layout using links, and attaches its separate MTP runtime and selected vision projector. Dense-only GGUFs are no longer treated as complete standalone models. MTP settings recognize validated prepared packs even when their GGUFs omit the original model-name metadata.

Imported tuning settings remain active without automatic preset replacement. Portable CPU affinity applies to the whole server process tree on Linux. Imports of another KV variant reuse verified file identities during the plugin session; changed files are checked again.

The bundled runtime is unchanged. NVFP4 describes model weights here; this release does not add NVFP4 or FP8 KV. Stock Flash-Next serving uses MTP; DFlash2 remains CLI-only in the tested engine. Model weights stay in their original folders.

Strata plugin 0.7.0

Strata 0.7.0

Portable configs can now import prepared Flash-Next NVFP4 GPTQ and NVFP4/Q8 Mix packs. Import finds each component by exact byte size and MD5, restores the pack layout using links, and attaches its separate MTP runtime and selected vision projector. Dense-only GGUFs are no longer treated as complete standalone models. MTP settings recognize validated prepared packs even when their GGUFs omit the original model-name metadata.

Imported tuning settings remain active without automatic preset replacement. Portable CPU affinity applies to the whole server process tree on Linux. Imports of another KV variant reuse verified file identities during the plugin session; changed files are checked again.

The bundled runtime is unchanged. NVFP4 describes model weights here; this release does not add NVFP4 or FP8 KV. Stock Flash-Next serving uses MTP; DFlash2 remains CLI-only in the tested engine. Model weights stay in their original folders.

Strata plugin 0.6.6

Strata 0.6.6

DeepSeek profiles now share one official engine. The plugin reuses installed signed runtimes from earlier profiles through Strata's native verified installer, preserving the originals and avoiding duplicate downloads when model files move or another profile is added. Existing managed profiles migrate on their next load or tune; explicit custom runtime overrides stay unchanged.

Model preparation now finishes before tuning requests GPU admission. Preparing, waiting and measuring are separate UI phases. A preparation job no longer pins an idle Strata GPU server or displays premature token-rate measurements. Model files and native tuning policy are unchanged.

Strata plugin 0.6.5

Strata 0.6.5

Local DeepSeek 4.1 Flash imports now use Strata's unchanged native setup. Complete supported shard sets are staged without copying weights; nearby Engram, DSpark and metadata files are verified and reused. The native installer obtains the separate official DeepSeek engine, prepares its runtime pack and writes its own configuration. Engine-specific library paths are preserved.

Existing DeepSeek profiles are no longer blocked by a cached importer warning. The editor explains native KV management, DSpark and text-only local setup. Update and restart DSH, then load the existing DeepSeek profile or re-import its folder.

Native engine code and binaries are not modified by this update. DeepSeek's official engine is installed through its upstream signed feed when needed.

Strata plugin 0.6.4

Strata 0.6.4

Strata now cooperates with the optional dsh-model-scheduler for GPU handoff with NInfer and other registered providers. A request holds its shared lease through generation and checkpoint saving; native tuning reserves the same GPU. The settings card shows whether shared scheduling is active.

Native startup failures now report a concise reason and competing GPU usage, preserving full diagnostics in a private startup.log. Failed loads and cancelled queued requests release admission safely for retries. The Strata engine, native memory sizing, context and KV settings are unchanged.

Install dsh-model-scheduler separately and restart DSH to enable shared handoff. Without it, unload another provider's model before starting Strata on the same GPU.

Strata plugin 0.6.3

Strata 0.6.3

DeepSeek profiles now expose their own native Thorough tuner through Auto-tune and Save and auto-tune. The unchanged upstream tuner and its helpers are included in the repository and plugin package. Successful DSpark/CPU-thread settings persist across restarts; failed or cancelled measurements keep the last saved settings. Activity shows native measured decode rates.

Qwen keeps its existing tuner. Strata engine binaries, model weights and vision support are unchanged. Existing profiles and caches remain in the plugin data directory when updating.

Strata plugin 0.6.2

Strata 0.6.0

Adds portable JSON model configuration export/import. Required pieces are matched recursively by exact size and MD5, including model shards, vision projectors and separate MTP runtimes. Original weights stay in place. Selected model settings and supported tuning flags are retained.

The bundled Strata engine is unchanged. Qwen 3.6 settings explain its fixed FP32 KV and embedded-MTP requirement.

Strata plugin 0.6.1

Strata 0.6.0

Adds portable JSON model configuration export/import. Required pieces are matched recursively by exact size and MD5, including model shards, vision projectors and separate MTP runtimes. Original weights stay in place. Selected model settings and supported tuning flags are retained.

The bundled Strata engine is unchanged. Qwen 3.6 settings explain its fixed FP32 KV and embedded-MTP requirement.

Strata plugin 0.6.0

Strata 0.6.0

Adds portable JSON model configuration export/import. Required pieces are matched recursively by exact size and MD5, including model shards, vision projectors and separate MTP runtimes. Original weights stay in place. Selected model settings and supported tuning flags are retained.

The bundled Strata engine is unchanged. Qwen 3.6 settings explain its fixed FP32 KV and embedded-MTP requirement.