Strata for DSH
A local model provider for DeepSeek Harness. 0.6.0 identifies built-in MTP and supports selecting or automatically preparing a separate Flash-Next MTP runtime. It automatically downloads the verified Strata runtime after a GitHub installation. It separates Models, Files and Activity into focused views. Each model has a Settings view that opens in place, with Load beside it. Strata supports native SSD checkpoints and a top-level Purge all cache button.
Install
Requires DSH 0.2.1-rc.2 Web profile, Node 22.19+, and pnpm on DSH's PATH. The included engines target Linux x86-64, NVIDIA Blackwell sm_120, CUDA 13, glibc 2.43 or newer. Other GPUs and operating systems have not been validated with this bundle.
Download the .tgz from Releases, then enter its absolute local path in Plugins → Add plugin. Enable it and open Settings → Models. Strata, portable Python, CUDA libraries, the GGUF helper and the Gyro L fork are included. Model weights are separate. No compiler or separate Strata installation is needed.
The GitHub address is also accepted by DSH:
https://github.com/averyjohnkucan-arch/dsh-strata
A GitHub source installation automatically downloads and verifies the runtime from this repository when the plugin is enabled. Progress appears in the Strata card; if it fails, use Retry installation. Prefer the release archive for an offline installation. Updating an existing plugin requires uninstalling and reinstalling it, restarting DSH, then reloading the browser page. Profiles and original weights are retained in the plugin's data directory.
Import models
- Choose Import files, then select any shard or a folder. The host file picker works on the DSH machine, including mounted drives.
- The importer recursively reads GGUF headers and groups shards. Files lists complete models, incomplete sets, each file's purpose and its size. Original weights stay where they are.
- A single vision projector under the model folder is selected automatically. Otherwise choose Select mmproj file or Continue without vision. Strata checks compatibility when loading.
- Choose KV precision (K / V) and Context, then Save and auto-tune. Context defaults to Native, using the model's declared GGUF context. Smaller choices remain available in 32,768-token increments. Saved explicit choices are retained.
- When tuning finishes, choose Load and select the model in DSH's conversation picker.
The KV choices are FP16/FP16, FP16/Q8, Q8/Q8, Q8/Q5, Q5/Q5, Q5/Q4, Q4/Q4. Unsupported choices are disabled. The importer never silently substitutes precision. The bundled TQP fork is selected automatically for Gyro L's TQP5/6/7 encodings; Gyro S uses the regular Strata engine. Explicit runtime overrides remain available under Advanced.
Control-vector GGUFs appear under Files → Components, never as loadable models. Earlier accidental vector profiles are preserved in component-profiles.json and removed from model routes; original GGUF files are untouched. Select reveals bulk tuning controls. Logs remain under Activity.
Thinking presets
The conversation selector offers Low, Medium, High and, for compatible original Flash-Next models, Uncensored. The native requests are low, medium, xhigh, and xhigh with the bundled experimental refusal-direction projection enabled. Ordinary choices explicitly disable projection. The vector loads once with the model; switching presets applies to the next request without a model reload. Each compatible request includes a current-mode runtime notice, so the assistant can distinguish the new setting from earlier replies. Changing projection also changes this prompt suffix; the prior conversation text remains available. Native KV state is separated by projection mode, and SSD checkpoint names are partitioned by it.
The projection came unchanged from upstream Strata, not from this plugin's author. It is a model-specific experiment, not a guarantee that every refusal disappears. Disabling it does not undo changes already baked into an abliterated model. See projection provenance and behavior.
Tuning and checkpoints
Save and auto-tune calls Strata's native tune auto --mode thorough. Strata owns the search, resource allocation and presets. The plugin adds model identity, context, KV, transport and projector paths. It does not set numeric worker, PCIe or prefill performance presets. Thorough tuning can take hours. Failed or cancelled tuning retains the previous result.
After a successful ordinary conversation answer, Strata writes its native session checkpoint to SSD. The plugin keeps up to 16 GiB per model, evicting the checkpoint with the oldest last-use time first. Files and last-use times survive process restarts and system reboots. On the first conversation after loading a model, its matching checkpoint is restored before generation. Strata checks model and cache identity; incompatible checkpoints are skipped. Its own RAM cache handles further conversations while that model stays loaded.
Disk checkpoints work with FP16/FP16, Q8/Q8 and Q4/Q4 (and legacy Q8/Q4). The fork currently rejects mixed KV snapshots; the editor displays that limitation. An individual checkpoint larger than 16 GiB is not retained. Restoring KV avoids some repeated prompt processing; model weights still need loading into RAM/VRAM. No model-switch speed improvement is promised without a workload measurement.
Purge all cache, at the top of the provider card, waits for an active answer, stops loaded plugin models and removes the plugin-owned cache directory, including SSD checkpoints and native RAM caches. Model weights, prepared packs, profiles, tuning presets and DSH chat history are retained. A subsequent load rebuilds its caches.
Runtime and storage
Regular Strata is based on v0.2.7 / Control 0.1.52 / engine 0.1.52. Both bundled engines include a small native checkpoint fix for models without an allocated MTP draft. Their source patches, upstream revisions, licenses and binary checksums are under vendor/. The Gyro L fork also includes its recorded TQP CPU changes. These engines were built for the tested sm_120 machine; HIP, SYCL, Windows and other CUDA architectures were not tested.
Local preparation supports qwen4exp Flash-Next, compatible qwen35moe GGUFs and native-catalog deepseek41 shard sets. Complete shards establish presence, not numerical compatibility. Engram/PLE, MTP, DFlash and vision components appear in the inventory. DFlash is inventoried only. Native Qwen loads embedded MTP, but Strata's native engine drafts only for greedy requests without sampling or penalties. It uses the native engine's FP32 KV format; external native Qwen donors are unsupported. Existing explicit Flash-Next MTP configurations remain readable.
Flash-Next requires a separate prepared Strata MTP directory containing experts.bin, dense.bin, dense.txt, and draft_vocab.bin. Import attaches an unambiguous discovered runtime; otherwise Strata fetches its approximately 5 GB draft from the original checkpoint and prepares a shared runtime on first use. To avoid that download, choose an existing prepared MTP folder in the model's settings. Raw MTP GGUFs from other converters are not automatically interchangeable with this runtime format. Changing MTP invalidates the previous model-specific tuning result.
Data lives under $DSH_HOME/dsh-strata-v2, normally ~/.dsh/dsh-strata-v2. cache/checkpoints// holds private session files. This writable plugin data is separate from its replaceable installed package, so updates do not erase checkpoints. The optional dataDir setting relocates it. Checkpoints contain conversation state; protect this directory as you would chat history. The host API uses DSH's authenticated owner connection, including after a plugin/server restart.
Packing and tokenizer extraction use vendored upstream tools. Shards are staged with symlinks; source weights are read-only. Upstream compatibility packing may convert dense tensors to BF16 and records the conversions.
Development
client/: provider card, file picker and compact settings; rebuildclient.jsafter edits.lib/: adapter, lifecycle, checkpoints and workflow.python/: inspection, preparation and release installation.vendor/: pinned Control/helper sources, engine metadata, patches and licenses. Engine binaries are in the pinned release runtime.packaging/: verified, self-contained release packaging.
pnpm install --ignore-scripts
pnpm run build
pnpm run check
pnpm test
The release workflow tests the wrapper and packages verified runtime inputs. GPU regressions and real-model validation are recorded separately in docs/VALIDATION.md; CI does not have a GPU. Building the browser bundle requires no access to model files.
Portable model configs
Open a model’s ⋯ → Export config JSON menu. Each model shard, selected vision projector and selected MTP runtime piece is identified by its exact byte size and full-file MD5. Hashing large weights takes time; Activity shows progress and offers Download config JSON when complete. Model weights are not exported.
Choose Import config, select the JSON, and enter the folder to search. Subfolders are searched by size first, then MD5. The plugin links matching files into a private layout and creates a new model profile, so filenames and drive locations can change without copying weights. Native binaries, credentials and machine-specific runtime paths are excluded. Supported saved tuning flags are retained; run Thorough tuning again after changing hardware.
The Strata engine is unchanged. Qwen 3.6 remains FP32 KV with embedded MTP only; its settings explain these native limitations.
During auto-tuning, the Models and Activity views show Prefill and Decode in tokens per second from the latest completed native measurement. Rates refresh as trial logs arrive; pending measurements show “Measuring…”. These are live trial samples, not the final aggregate tuning score.
Rename a saved model under Settings → Model name → Save. The saved name appears in the model list and DSH model picker. Renaming preserves its model files, profile identity and saved tuning.
DeepSeek 1017 auto-tuning
For a prepared DeepSeek profile, choose ⋯ → Auto-tune DeepSeek or Settings → Save and auto-tune. The DeepSeek download flow exposes the same native-tuning option after setup. DeepSeek calls its own Thorough tuner, not the Qwen preset tuner. It measures DSpark and CPU-thread choices, keeps the native tuner’s greater-than-5% acceptance rule, and preserves your selected context and automatic RAM-store sizing. Activity shows its measured decode rate; this tuner does not measure prefill.
The exact tuner is included at vendor/gyro-l-control/tools/dsv41/tune.py with its tokenizer/smoke helpers and upstream license/provenance. Its SHA-256 is bb75873d33774cc5d27525e4af0693eb9d6cb2a72dc9d92c50426d3651b1fab0, matching the pinned Control 0.1.52 tool. No tuner download or manual script path is needed.
Successful settings and the measurement report are saved in the model profile and reused after restarting DSH. Failed/cancelled measurements preserve the last saved configuration. An exit code of zero alone is not considered a successful tune. DeepSeek saves launch arguments rather than a Qwen preset ID. Updating the plugin preserves existing profiles; the engine, vision support and model weights are unchanged. Local DeepSeek GGUF preparation remains separate from the Qwen importer; DeepSeek profiles from Strata’s download/setup flow carry the required prepared configuration.
Switching between local providers
For shared GPU handoff, install dsh-model-scheduler and restart DSH. Strata and NInfer register with that optional service; it queues competing requests and unloads the previous cooperating model before loading the next one. Strata holds its lease through generation and checkpoint saving, and native auto-tuning reserves the same GPU. The Strata card says Shared GPU scheduler when connected.
Without that plugin, the card says Independent GPU queue. Strata serializes its own requests, but does not unload another provider's server. Finish the other answer and unload that model before loading Strata on the same GPU. Neither mode kills unknown GPU processes. Context, KV precision and native engine memory sizing remain unchanged.
A failed native startup now reports a short reason and, for allocation failures, GPU usage captured before loading. Full startup output is kept privately at models//startup.log in Strata's data directory. Failed or cancelled startup releases admission so the next request can retry.
Importing local DeepSeek 4.1 Flash
Import a folder containing a complete native Strata set: IST-DAS IQ1_M, IQ2_XXS or IQ2_XS, or Unsloth UD-Q8_K_XL, using the publisher filenames (upstream legacy names are also accepted). Nearby Engram, DSpark and metadata files are reused. Strata's unchanged native setup verifies hashes, acquires its separate official DeepSeek engine, creates the pack and writes its own launch settings. Missing companion files are acquired by that native flow. Existing model weights are linked into plugin staging rather than copied.
The first hash check can take several minutes for hundreds of GB. A failed or cancelled setup leaves the saved profile unprepared and can be retried; source weights are never removed. The old cached “local preparation unavailable” warning no longer blocks existing DeepSeek profiles. Re-importing is optional after updating: the existing saved model can be loaded again.
DeepSeek owns its KV representation and DSpark automatic policy. The editor does not show Qwen KV controls for it. This local path is text only; it does not attach Flash Next mmproj files. Its context selection is preserved and the native setup validates the supported range. File preparation skips the separate GPU smoke process so an Import/Prepare operation cannot collide with another provider; the subsequent scheduled model load validates readiness through the actual DeepSeek engine. Native auto-tuning remains available after preparation.
DeepSeek uses its own persistent slot cache under cache/deepseek/, with a 16 GiB native slot budget by default. The plugin does not send Qwen's explicit SAVE/RESTORE protocol to that engine. Purge all cache covers the default DeepSeek slot directory as well.
Shared DeepSeek engine (0.6.6)
Every managed DeepSeek profile uses one official runtime under runtime/deepseek/engine-dsv41 in the plugin data directory. The first use checks existing per-profile installations and migrates a valid signed package through Strata's native installer. It copies the runtime into the shared location, preserves the original, and does not fetch it again. Only when no usable runtime exists does the native signed-feed downloader run; all profiles then share that resumable download. Installation is locked across bridge processes.
Existing prepared profiles switch to the shared runtime when next loaded or tuned. Model paths, prepared packs, contexts and tuning results stay profile-specific. Explicit custom runtime overrides are preserved. Moving/copying GGUF files or adding another profile does not require downloading another engine. Changing weight files may still require native verification and packing.
Tuning reports Preparing model, then Waiting for GPU, then measurement. Preparation uses CPU/disk and does not reserve the GPU. Token rates appear only when the measurement stage starts.
