dsh-plugin-local-infer 0.1.2 → 0.1.4

package/packages/providers/lib/engines/strata/strataWindows.js

@@ -0,0 +1,37 @@
+/**
+ * The Strata Windows build profile. The Windows source route is the engine's
+ * own START-HERE.bat wrapper (docs/SOURCE_INSTALL.md, From source: Windows),
+ * not the Linux setup.sh. `--yes` accepts the recommended answer to every
+ * question, so the step runs unattended. The wrapper acquires a ready-made
+ * engine when one fits the PC, and compiles one when it does not: the
+ * compile needs the Visual Studio 2022 Build Tools with the C++ workload
+ * (the sources are C++20, so VS2019 is too old). When setup cannot find the
+ * toolchain, point STRATA_VCVARS at vcvars64.bat before running. The build
+ * keeps its dependencies and products in the source install and writes the
+ * launch scripts (run-<model>.bat) and run configs (strata-<model>.json)
+ * into the checkout, so the profile carries no build directory. Updates run
+ * through UPDATE.bat.
+ * @module dsh-plugin-local-infer/providers/engines/strata/strataWindows
+ */
+import { strataSourceLaunchArgs } from "./strata.js";
+import { SYSTEM_DEPS } from "../../systemDeps.js";
+export const STRATA_WINDOWS = {
+    buildSteps: ['cd {dir} && .\\START-HERE.bat --yes'],
+    // The engine command is the strata Terminal on every platform (a
+    // PowerShell command on Windows), so the name stays plain.
+    binary: 'strata',
+    // The Windows source route is the engine's own START-HERE.bat wrapper; the
+    // host deps are the same git-plus-curl list as the Linux route.
+    deps: [SYSTEM_DEPS.git, SYSTEM_DEPS.curl],
+};
+/**
+ * strataWindowsSourceLaunch renders the source-build launch argv for the
+ * Windows route: the engine's own START-HERE.bat wrapper with --yes
+ * (unattended) and the resolved model, family, vision, context, and port. The
+ * wrapper acquires a ready-made engine when one fits the PC, and compiles one
+ * (the Visual Studio 2022 C++ toolchain) when it does not.
+ */
+export function strataWindowsSourceLaunch(args) {
+    return strataSourceLaunchArgs('START-HERE.bat', args);
+}
+//# sourceMappingURL=strataWindows.js.map

package/packages/providers/lib/engines.d.ts

@@ -14,15 +14,18 @@ export interface EngineInfo {
     supports: {
         docker: boolean;
         local: boolean;
+        binary?: boolean;
     };
     parameters: readonly ParameterSpec[];
 }
 /**
  * The engines the Web card offers, in the order the card lists them. NInfer,
- * Hyper-Qwen, Infernix, and llama.cpp carry parameter surfaces (llama.cpp's
- * is the specified set the plan names, verified when the binary lands);
- * Strata is the placeholder engine, its surface empty until the engine's
- * flag set lands.
+ * Hyper-Qwen, Infernix, llama.cpp, Strata, and Basalt all carry parameter
+ * surfaces (llama.cpp's is the specified set the plan names, verified when the
+ * binary lands); Strata's surface is the `strata launch` flags and it supports
+ * all three build modes (docker, source build, and the binary install);
+ * Basalt's surface is the `basalt_config.json` settings and it is local-build
+ * only (the C++/CUDA engine ships no container image).
  */
 export declare const ENGINES: readonly EngineInfo[];
-//# sourceMappingURL=engines.d.ts.map
+//# sourceMappingURL=engines.d.ts.map

package/packages/providers/lib/engines.js

@@ -7,17 +7,20 @@
  * here, never a rewire of the card.
  * @module dsh-plugin-local-infer/providers/engines
  */
-import { HYPERQWEN_PARAMETERS } from "./hyperqwen.js";
-import { INFERNIX_PARAMETERS } from "./infernix.js";
-import { LLAMACPP_PARAMETERS } from "./llamacpp.js";
-import { NINFER_PARAMETERS } from "./ninfer.js";
-import { STRATA_PARAMETERS } from "./strata.js";
+import { BASALT_PARAMETERS } from "./engines/basalt/basalt.js";
+import { HYPERQWEN_PARAMETERS } from "./engines/hyperqwen/hyperqwen.js";
+import { INFERNIX_PARAMETERS } from "./engines/infernix/infernix.js";
+import { LLAMACPP_PARAMETERS } from "./engines/llamacpp/llamacpp.js";
+import { NINFER_PARAMETERS } from "./engines/ninfer/ninfer.js";
+import { STRATA_PARAMETERS } from "./engines/strata/strata.js";
 /**
  * The engines the Web card offers, in the order the card lists them. NInfer,
- * Hyper-Qwen, Infernix, and llama.cpp carry parameter surfaces (llama.cpp's
- * is the specified set the plan names, verified when the binary lands);
- * Strata is the placeholder engine, its surface empty until the engine's
- * flag set lands.
+ * Hyper-Qwen, Infernix, llama.cpp, Strata, and Basalt all carry parameter
+ * surfaces (llama.cpp's is the specified set the plan names, verified when the
+ * binary lands); Strata's surface is the `strata launch` flags and it supports
+ * all three build modes (docker, source build, and the binary install);
+ * Basalt's surface is the `basalt_config.json` settings and it is local-build
+ * only (the C++/CUDA engine ships no container image).
  */
 export const ENGINES = [
     {
@@ -47,8 +50,16 @@ export const ENGINES = [
     {
         id: 'strata',
         label: 'Strata',
-        supports: { docker: true, local: false },
+        supports: { docker: true, local: true, binary: true },
         parameters: STRATA_PARAMETERS,
     },
+    {
+        // Basalt is the C++/CUDA engine for the Qwen3.8-Flash-Next MoE model,
+        // Blackwell-only: a local source build, no container image.
+        id: 'basalt',
+        label: 'Basalt',
+        supports: { docker: false, local: true },
+        parameters: BASALT_PARAMETERS,
+    },
 ];
-//# sourceMappingURL=engines.js.map
+//# sourceMappingURL=engines.js.map

package/packages/providers/lib/guidelines.d.ts

@@ -1,9 +1,10 @@
 /**
  * Built-in hardware guidelines. The rules that every engine shares, expressed
  * as `Guideline` data over (config, hardware): the docker-availability gate,
- * the host RAM tier ceiling, and the WSL2 pin-memory boot abort. Each is
- * registered with `ctx.localInfer` by the provider `apply`, so a per-engine
- * override or an extra rule composes on top without touching these.
+ * the host RAM tier ceiling, the WSL2 pin-memory boot abort, and the Basalt
+ * Blackwell (sm_120) gate. Each is registered with `ctx.localInfer` by the
+ * provider `apply`, so a per-engine override or an extra rule composes on top
+ * without touching these.
  *
  * The rules that need the registered VramModel (the estimator-backed
  * "exceeds usable VRAM" and "spec needs draft weights" rules) live in the
@@ -26,4 +27,4 @@ export declare const HOST_STATE_IMAGE_BYTES = 195897344;
  * degrades before it breaks.
  */
 export declare const BUILTIN_GUIDELINES: readonly Guideline[];
-//# sourceMappingURL=guidelines.d.ts.map
+//# sourceMappingURL=guidelines.d.ts.map

package/packages/providers/lib/guidelines.js

@@ -1,9 +1,10 @@
 /**
  * Built-in hardware guidelines. The rules that every engine shares, expressed
  * as `Guideline` data over (config, hardware): the docker-availability gate,
- * the host RAM tier ceiling, and the WSL2 pin-memory boot abort. Each is
- * registered with `ctx.localInfer` by the provider `apply`, so a per-engine
- * override or an extra rule composes on top without touching these.
+ * the host RAM tier ceiling, the WSL2 pin-memory boot abort, and the Basalt
+ * Blackwell (sm_120) gate. Each is registered with `ctx.localInfer` by the
+ * provider `apply`, so a per-engine override or an extra rule composes on top
+ * without touching these.
  *
  * The rules that need the registered VramModel (the estimator-backed
  * "exceeds usable VRAM" and "spec needs draft weights" rules) live in the
@@ -11,7 +12,7 @@
  * a model, not against the hardware alone.
  * @module dsh-plugin-local-infer/providers/guidelines
  */
-import { infernixDraftModelPath } from "./infernix.js";
+import { infernixDraftModelPath } from "./engines/infernix/infernix.js";
 /**
  * One pinned Host StateImage, in bytes. The NInfer measured figure for the
  * 27B artifact (195,897,344 B, the README's Host StateImage size), used as
@@ -73,5 +74,17 @@ export const BUILTIN_GUIDELINES = [
             'draftModelPreset' in config &&
             infernixDraftModelPath(config) === null,
     },
+    {
+        // Basalt is Blackwell-only: the engine is compiled for sm_120 and its
+        // NVFP4 expert kernels need it. A basalt config on a card that reports a
+        // different SM number (or no card at all) cannot boot, so the rule names
+        // the fix before the engine starts. It is basalt-specific: only basalt
+        // configs carry the engine key, so every other engine stays silent.
+        id: 'basalt-needs-blackwell',
+        level: 'error',
+        message: 'Basalt needs an NVIDIA Blackwell GPU (SM 120). This machine reports a different SM number, or no GPU at all.',
+        anchors: ['gpu'],
+        when: (config, hw) => config.engine === 'basalt' && hw.gpu?.sm !== 120,
+    },
 ];
-//# sourceMappingURL=guidelines.js.map
+//# sourceMappingURL=guidelines.js.map

package/packages/providers/lib/index.d.ts

@@ -1,6 +1,6 @@
 /**
  * The concrete providers for the local-inference capability: the engine
- * adapters (ninfer, hyperqwen, llamacpp, infernix, strata), the NInfer
+ * adapters (ninfer, hyperqwen, llamacpp, infernix, strata, basalt), the NInfer
  * VramModel, the built-in presets, the built-in guidelines, and the per-OS
  * SystemDep install command data. Each registers itself with `ctx.localInfer`.
  *
@@ -18,36 +18,45 @@ export declare const inject: string[];
  * Phase 5 adapters. NInfer's surface is CLI flags (`toArgv`); Hyper-Qwen's is
  * `.env` keys (`toEnv`).
  */
-export { CTX_OPTIONS, HYPERQWEN_PARAMETERS, KV_OPTIONS, SPEC_OPTIONS as HYPERQWEN_SPEC_OPTIONS, } from './hyperqwen.ts';
-export { CHAT_TEMPLATE_OPTIONS, DEVICE_PROFILE_OPTIONS, KV_DTYPE_OPTIONS, NINFER_PARAMETERS, SPEC_OPTIONS as NINFER_SPEC_OPTIONS, } from './ninfer.ts';
-export { NINFER_VRAM_MODELS, ninferVramModel } from './vram.ts';
+export { CTX_OPTIONS, HYPERQWEN_PARAMETERS, KV_OPTIONS, SPEC_OPTIONS as HYPERQWEN_SPEC_OPTIONS, } from './engines/hyperqwen/hyperqwen.ts';
+export { CHAT_TEMPLATE_OPTIONS, DEVICE_PROFILE_OPTIONS, KV_DTYPE_OPTIONS, NINFER_PARAMETERS, SPEC_OPTIONS as NINFER_SPEC_OPTIONS, } from './engines/ninfer/ninfer.ts';
+export { BASALT_VRAM_MODELS, NINFER_VRAM_MODELS, basaltVramModel, ninferVramModel, } from './vram.ts';
 export { BUILTIN_GUIDELINES, HOST_STATE_IMAGE_BYTES } from './guidelines.ts';
 export { ENGINES } from './engines.ts';
 export type { EngineInfo } from './engines.ts';
-export { ENGINE_SOURCES, FORK_REMOTE, buildEngineRepo, engineBuildSteps, engineSource, syncAndRebuildEngine, syncEngineRepo, } from './acquire.ts';
-export type { AcquireLog, EngineSource, EngineStepResult, SyncRebuildRequest, SyncRebuildResult, } from './acquire.ts';
-export { ENGINE_DEP_PLANS, SYSTEM_DEPS } from './deps.ts';
+export { ENGINE_SOURCES, FORK_REMOTE, buildEngineRepo, engineBuildSteps, engineSource, syncAndRebuildEngine, syncEngineForDeploy, syncEngineRepo, } from './acquire.ts';
+export type { AcquireLog, EngineDeploySync, EngineSource, EngineStepResult, SyncRebuildRequest, SyncRebuildResult, SyncStepResult, } from './acquire.ts';
+export { ENGINE_DEP_PLANS, SYSTEM_DEPS, engineLocalDeps } from './deps.ts';
 export { ENGINE_PLATFORM_DEFAULTS, enginePlatformDefaults } from './platform.ts';
 export type { EnginePlatformDefaults } from './platform.ts';
 export { BUILTIN_PRESETS, PRESET_SOURCES, extractJsonBlock } from './presets.ts';
 export { DEFAULT_MODELS, defaultModelFor, defaultModelsForEngine } from './model-defaults.ts';
-export { HyperQwenAdapter, LLAMACPP_DOCKER_IMAGE, NINFER_DOCKER_IMAGE, NINFER_MODEL_ARTIFACTS, NInferAdapter, STRATA_DOCKER_IMAGE, StrataAdapter, LlamaCppAdapter, customArgEnv, customArgFlags, envFromSpecs, flagsFromSpecs, nodeLifecycleHost, pollHealth, } from './adapters.ts';
-export type { NInferModelArtifact } from './adapters.ts';
-export { KV_CACHE_TYPE_OPTIONS, LLAMACPP_MODEL_ARTIFACTS, LLAMACPP_PARAMETERS, llamaCppModelKey, } from './llamacpp.ts';
-export type { LlamaCppModelArtifact } from './llamacpp.ts';
-export { DFLASH2_CUSTOM_VALUE, DFLASH2_GATE_MESSAGE, INFERNIX_DFLASH2_MODELS, INFERNIX_KV_DTYPE_OPTIONS, INFERNIX_MODEL_ARTIFACTS, INFERNIX_NGRAM_RESIDENCY_OPTIONS, INFERNIX_NGRAM_VOLUMES, INFERNIX_PARAMETERS, INFERNIX_REASONING_LOOP_OPTIONS, INFERNIX_SPEC_OPTIONS, SPEC_GATE_MESSAGE as INFERNIX_SPEC_GATE_MESSAGE, dflash2PresetOptions, infernixArtifactFor, infernixDraftModelPath, infernixGetDownloads, infernixModelKey, } from './infernix.ts';
-export type { InfernixDFlash2Model, InfernixModelArtifact, InfernixNgramVolume, } from './infernix.ts';
-export { STRATA_PARAMETERS } from './strata.ts';
+export { createNodeLifecycleHost, customArgEnv, customArgFlags, envFromSpecs, flagsFromSpecs, nodeLifecycleHost, pollHealth, resolveLaunchArgv, } from './adapters.ts';
+export type { SudoHook } from './adapters.ts';
+export { HyperQwenAdapter } from './engines/hyperqwen/hyperqwenAdapter.ts';
+export { InfernixAdapter } from './engines/infernix/infernixAdapter.ts';
+export { LLAMACPP_DOCKER_IMAGE, LlamaCppAdapter } from './engines/llamacpp/llamacppAdapter.ts';
+export { NINFER_DOCKER_IMAGE, NINFER_MODEL_ARTIFACTS, NInferAdapter } from './engines/ninfer/ninferAdapter.ts';
+export { STRATA_DOCKER_IMAGE, StrataAdapter } from './engines/strata/strataAdapter.ts';
+export type { NInferModelArtifact } from './engines/ninfer/ninferAdapter.ts';
+export { KV_CACHE_TYPE_OPTIONS, LLAMACPP_MODEL_ARTIFACTS, LLAMACPP_PARAMETERS, llamaCppModelKey, } from './engines/llamacpp/llamacpp.ts';
+export type { LlamaCppModelArtifact } from './engines/llamacpp/llamacpp.ts';
+export { DFLASH2_CUSTOM_VALUE, DFLASH2_GATE_MESSAGE, INFERNIX_DFLASH2_MODELS, INFERNIX_KV_DTYPE_OPTIONS, INFERNIX_MODEL_ARTIFACTS, INFERNIX_NGRAM_RESIDENCY_OPTIONS, INFERNIX_NGRAM_VOLUMES, INFERNIX_PARAMETERS, INFERNIX_REASONING_LOOP_OPTIONS, INFERNIX_SPEC_OPTIONS, SPEC_GATE_MESSAGE as INFERNIX_SPEC_GATE_MESSAGE, dflash2PresetOptions, infernixArtifactFor, infernixDraftModelPath, infernixGetDownloads, infernixModelKey, } from './engines/infernix/infernix.ts';
+export type { InfernixDFlash2Model, InfernixModelArtifact, InfernixNgramVolume, } from './engines/infernix/infernix.ts';
+export { STRATA_PARAMETERS } from './engines/strata/strata.ts';
+export { BASALT_PARAMETERS } from './engines/basalt/basalt.ts';
+export { BasaltAdapter } from './engines/basalt/basaltAdapter.ts';
 /**
- * Mount the providers. The NInfer measured VramModels register here so the
- * service's estimate prices a named model; the built-in hardware guidelines
- * register so the guideline engine warns on docker, host RAM, and WSL2
- * before the engine boots; the per-engine dep plans register so the deps,
- * installString, and runInstall entry points resolve a plan for every engine
- * the card offers; the built-in presets register so the defaults resolver
- * resolves a sane starting config for the detected hardware class; and the
- * engine adapters register so the launch, stop, and register entry points
- * route a plan to the adapter that rendered it.
+ * Mount the providers. The NInfer measured VramModels and the Basalt
+ * estimated VramModel register here so the service's estimate prices a named
+ * model; the built-in hardware guidelines register so the guideline engine
+ * warns on docker, host RAM, WSL2, and the Basalt Blackwell gate before the
+ * engine boots; the per-engine dep plans register so the deps, installString,
+ * and runInstall entry points resolve a plan for every engine the card
+ * offers; the built-in presets register so the defaults resolver resolves a
+ * sane starting config for the detected hardware class; and the engine
+ * adapters register so the launch, stop, and register entry points route a
+ * plan to the adapter that rendered it.
  */
 export declare function apply(ctx: Context): void;
 declare const plugin: {
@@ -56,4 +65,4 @@ declare const plugin: {
     apply: typeof apply;
 };
 export default plugin;
-//# sourceMappingURL=index.d.ts.map
+//# sourceMappingURL=index.d.ts.map

package/packages/providers/lib/index.js

@@ -1,6 +1,6 @@
 /**
  * The concrete providers for the local-inference capability: the engine
- * adapters (ninfer, hyperqwen, llamacpp, infernix, strata), the NInfer
+ * adapters (ninfer, hyperqwen, llamacpp, infernix, strata, basalt), the NInfer
  * VramModel, the built-in presets, the built-in guidelines, and the per-OS
  * SystemDep install command data. Each registers itself with `ctx.localInfer`.
  *
@@ -10,12 +10,18 @@
  * that the defaults resolver resolves against.
  * @module dsh-plugin-local-infer/providers
  */
-import { createNodeLifecycleHost, HyperQwenAdapter, InfernixAdapter, LlamaCppAdapter, NInferAdapter, StrataAdapter, } from "./adapters.js";
+import { createNodeLifecycleHost } from "./adapters.js";
+import { BasaltAdapter } from "./engines/basalt/basaltAdapter.js";
+import { HyperQwenAdapter } from "./engines/hyperqwen/hyperqwenAdapter.js";
+import { InfernixAdapter } from "./engines/infernix/infernixAdapter.js";
+import { LlamaCppAdapter } from "./engines/llamacpp/llamacppAdapter.js";
+import { NInferAdapter } from "./engines/ninfer/ninferAdapter.js";
+import { StrataAdapter } from "./engines/strata/strataAdapter.js";
 import { createDshDebugLogger } from "./debug.js";
 import { ENGINE_DEP_PLANS } from "./deps.js";
 import { BUILTIN_GUIDELINES } from "./guidelines.js";
 import { BUILTIN_PRESETS } from "./presets.js";
-import { NINFER_VRAM_MODELS } from "./vram.js";
+import { BASALT_VRAM_MODELS, NINFER_VRAM_MODELS } from "./vram.js";
 export const name = 'local-infer-providers';
 export const inject = ['localInfer'];
 /**
@@ -23,36 +29,47 @@ export const inject = ['localInfer'];
  * Phase 5 adapters. NInfer's surface is CLI flags (`toArgv`); Hyper-Qwen's is
  * `.env` keys (`toEnv`).
  */
-export { CTX_OPTIONS, HYPERQWEN_PARAMETERS, KV_OPTIONS, SPEC_OPTIONS as HYPERQWEN_SPEC_OPTIONS, } from "./hyperqwen.js";
-export { CHAT_TEMPLATE_OPTIONS, DEVICE_PROFILE_OPTIONS, KV_DTYPE_OPTIONS, NINFER_PARAMETERS, SPEC_OPTIONS as NINFER_SPEC_OPTIONS, } from "./ninfer.js";
-export { NINFER_VRAM_MODELS, ninferVramModel } from "./vram.js";
+export { CTX_OPTIONS, HYPERQWEN_PARAMETERS, KV_OPTIONS, SPEC_OPTIONS as HYPERQWEN_SPEC_OPTIONS, } from "./engines/hyperqwen/hyperqwen.js";
+export { CHAT_TEMPLATE_OPTIONS, DEVICE_PROFILE_OPTIONS, KV_DTYPE_OPTIONS, NINFER_PARAMETERS, SPEC_OPTIONS as NINFER_SPEC_OPTIONS, } from "./engines/ninfer/ninfer.js";
+export { BASALT_VRAM_MODELS, NINFER_VRAM_MODELS, basaltVramModel, ninferVramModel, } from "./vram.js";
 export { BUILTIN_GUIDELINES, HOST_STATE_IMAGE_BYTES } from "./guidelines.js";
 export { ENGINES } from "./engines.js";
-export { ENGINE_SOURCES, FORK_REMOTE, buildEngineRepo, engineBuildSteps, engineSource, syncAndRebuildEngine, syncEngineRepo, } from "./acquire.js";
-export { ENGINE_DEP_PLANS, SYSTEM_DEPS } from "./deps.js";
+export { ENGINE_SOURCES, FORK_REMOTE, buildEngineRepo, engineBuildSteps, engineSource, syncAndRebuildEngine, syncEngineForDeploy, syncEngineRepo, } from "./acquire.js";
+export { ENGINE_DEP_PLANS, SYSTEM_DEPS, engineLocalDeps } from "./deps.js";
 export { ENGINE_PLATFORM_DEFAULTS, enginePlatformDefaults } from "./platform.js";
 export { BUILTIN_PRESETS, PRESET_SOURCES, extractJsonBlock } from "./presets.js";
 export { DEFAULT_MODELS, defaultModelFor, defaultModelsForEngine } from "./model-defaults.js";
-export { HyperQwenAdapter, LLAMACPP_DOCKER_IMAGE, NINFER_DOCKER_IMAGE, NINFER_MODEL_ARTIFACTS, NInferAdapter, STRATA_DOCKER_IMAGE, StrataAdapter, LlamaCppAdapter, customArgEnv, customArgFlags, envFromSpecs, flagsFromSpecs, nodeLifecycleHost, pollHealth, } from "./adapters.js";
-export { KV_CACHE_TYPE_OPTIONS, LLAMACPP_MODEL_ARTIFACTS, LLAMACPP_PARAMETERS, llamaCppModelKey, } from "./llamacpp.js";
-export { DFLASH2_CUSTOM_VALUE, DFLASH2_GATE_MESSAGE, INFERNIX_DFLASH2_MODELS, INFERNIX_KV_DTYPE_OPTIONS, INFERNIX_MODEL_ARTIFACTS, INFERNIX_NGRAM_RESIDENCY_OPTIONS, INFERNIX_NGRAM_VOLUMES, INFERNIX_PARAMETERS, INFERNIX_REASONING_LOOP_OPTIONS, INFERNIX_SPEC_OPTIONS, SPEC_GATE_MESSAGE as INFERNIX_SPEC_GATE_MESSAGE, dflash2PresetOptions, infernixArtifactFor, infernixDraftModelPath, infernixGetDownloads, infernixModelKey, } from "./infernix.js";
-export { STRATA_PARAMETERS } from "./strata.js";
+export { createNodeLifecycleHost, customArgEnv, customArgFlags, envFromSpecs, flagsFromSpecs, nodeLifecycleHost, pollHealth, resolveLaunchArgv, } from "./adapters.js";
+export { HyperQwenAdapter } from "./engines/hyperqwen/hyperqwenAdapter.js";
+export { InfernixAdapter } from "./engines/infernix/infernixAdapter.js";
+export { LLAMACPP_DOCKER_IMAGE, LlamaCppAdapter } from "./engines/llamacpp/llamacppAdapter.js";
+export { NINFER_DOCKER_IMAGE, NINFER_MODEL_ARTIFACTS, NInferAdapter } from "./engines/ninfer/ninferAdapter.js";
+export { STRATA_DOCKER_IMAGE, StrataAdapter } from "./engines/strata/strataAdapter.js";
+export { KV_CACHE_TYPE_OPTIONS, LLAMACPP_MODEL_ARTIFACTS, LLAMACPP_PARAMETERS, llamaCppModelKey, } from "./engines/llamacpp/llamacpp.js";
+export { DFLASH2_CUSTOM_VALUE, DFLASH2_GATE_MESSAGE, INFERNIX_DFLASH2_MODELS, INFERNIX_KV_DTYPE_OPTIONS, INFERNIX_MODEL_ARTIFACTS, INFERNIX_NGRAM_RESIDENCY_OPTIONS, INFERNIX_NGRAM_VOLUMES, INFERNIX_PARAMETERS, INFERNIX_REASONING_LOOP_OPTIONS, INFERNIX_SPEC_OPTIONS, SPEC_GATE_MESSAGE as INFERNIX_SPEC_GATE_MESSAGE, dflash2PresetOptions, infernixArtifactFor, infernixDraftModelPath, infernixGetDownloads, infernixModelKey, } from "./engines/infernix/infernix.js";
+export { STRATA_PARAMETERS } from "./engines/strata/strata.js";
+export { BASALT_PARAMETERS } from "./engines/basalt/basalt.js";
+export { BasaltAdapter } from "./engines/basalt/basaltAdapter.js";
 /**
- * Mount the providers. The NInfer measured VramModels register here so the
- * service's estimate prices a named model; the built-in hardware guidelines
- * register so the guideline engine warns on docker, host RAM, and WSL2
- * before the engine boots; the per-engine dep plans register so the deps,
- * installString, and runInstall entry points resolve a plan for every engine
- * the card offers; the built-in presets register so the defaults resolver
- * resolves a sane starting config for the detected hardware class; and the
- * engine adapters register so the launch, stop, and register entry points
- * route a plan to the adapter that rendered it.
+ * Mount the providers. The NInfer measured VramModels and the Basalt
+ * estimated VramModel register here so the service's estimate prices a named
+ * model; the built-in hardware guidelines register so the guideline engine
+ * warns on docker, host RAM, WSL2, and the Basalt Blackwell gate before the
+ * engine boots; the per-engine dep plans register so the deps, installString,
+ * and runInstall entry points resolve a plan for every engine the card
+ * offers; the built-in presets register so the defaults resolver resolves a
+ * sane starting config for the detected hardware class; and the engine
+ * adapters register so the launch, stop, and register entry points route a
+ * plan to the adapter that rendered it.
  */
 export function apply(ctx) {
     const localInfer = ctx.localInfer;
     for (const model of NINFER_VRAM_MODELS) {
         localInfer.registerVramModel(model);
     }
+    for (const model of BASALT_VRAM_MODELS) {
+        localInfer.registerVramModel(model);
+    }
     for (const guideline of BUILTIN_GUIDELINES) {
         localInfer.registerGuideline(guideline);
     }
@@ -66,14 +83,19 @@ export function apply(ctx) {
     // (the HARNESS_LOG_LEVEL env var or the DSH logger level config), the host
     // logs every command it issues to an engine, from build to launch, through
     // the cordis logger's debug method. A normal run wires the no-debug host, so
-    // the command path stays quiet.
-    const host = createNodeLifecycleHost(createDshDebugLogger(ctx));
+    // the command path stays quiet. The sudo hook reports a waiting
+    // `sudo ln -sf ...` deploy link through the service's notifySudoNeeded, so
+    // the route-registered observer can show the CLI note while the card waits.
+    const host = createNodeLifecycleHost(createDshDebugLogger(ctx), {
+        notify: (command) => localInfer.notifySudoNeeded(command),
+    });
     localInfer.registerEngine(new NInferAdapter(host));
     localInfer.registerEngine(new HyperQwenAdapter(host));
     localInfer.registerEngine(new LlamaCppAdapter(host));
     localInfer.registerEngine(new InfernixAdapter(host));
     localInfer.registerEngine(new StrataAdapter(host));
+    localInfer.registerEngine(new BasaltAdapter(host));
 }
 const plugin = { name, inject, apply };
 export default plugin;
-//# sourceMappingURL=index.js.map
+//# sourceMappingURL=index.js.map

package/packages/providers/lib/model-defaults.d.ts

@@ -4,8 +4,13 @@
  * Hyper-Qwen W4A16 27B and AWQ 35B-A3B artifacts, the llama.cpp GGUF 27B and
  * 35B-A3B artifacts, and the Infernix 27B NVIDIA NVFP4 and the Qwen3.8-
  * Flash-Next Dense8 conversion (whose whole-repo download carries the FP8
- * n-gram volume the `--ngram-volume` flag reads). Strata does not focus on
- * the dense Qwen models, so it carries no entry.
+ * n-gram volume the `--ngram-volume` flag reads). Strata carries the two
+ * models the README's Models section links (Gyro and DeepSeek V4.1-Flash),
+ * each named by the model the engine accepts and the repo it downloads.
+ * Basalt carries the two `.basalt` artifacts the engine serves (the
+ * Qwen3.8-Flash-Next NVFP4 conversion and the DeepSeek-V4.1-Flash UD-IQ1_M
+ * pack, each named by the repo and file the adapter's deploy step
+ * downloads).
  *
  * The list is the single source the card's Model dropdown, the model
  * presence check, and the hf CLI download route read. Adding a model is a
@@ -22,8 +27,8 @@ import type { DefaultModel } from 'dsh-plugin-local-infer/core';
 export declare const DEFAULT_MODELS: readonly DefaultModel[];
 /**
  * defaultModelsForEngine reads the default list for one engine id, in list
- * order. An engine with no entry (Strata) reads as an empty list, which the
- * card renders as no Model row.
+ * order. An engine with no entry reads as an empty list, which the card
+ * renders as no Model row.
  */
 export declare function defaultModelsForEngine(engineId: string): readonly DefaultModel[];
 /**
@@ -32,4 +37,4 @@ export declare function defaultModelsForEngine(engineId: string): readonly Defau
  * undefined entry with a 404, so a model no default names never downloads.
  */
 export declare function defaultModelFor(engineId: string, modelId: string): DefaultModel | undefined;
-//# sourceMappingURL=model-defaults.d.ts.map
+//# sourceMappingURL=model-defaults.d.ts.map

package/packages/providers/lib/model-defaults.js

@@ -4,14 +4,20 @@
  * Hyper-Qwen W4A16 27B and AWQ 35B-A3B artifacts, the llama.cpp GGUF 27B and
  * 35B-A3B artifacts, and the Infernix 27B NVIDIA NVFP4 and the Qwen3.8-
  * Flash-Next Dense8 conversion (whose whole-repo download carries the FP8
- * n-gram volume the `--ngram-volume` flag reads). Strata does not focus on
- * the dense Qwen models, so it carries no entry.
+ * n-gram volume the `--ngram-volume` flag reads). Strata carries the two
+ * models the README's Models section links (Gyro and DeepSeek V4.1-Flash),
+ * each named by the model the engine accepts and the repo it downloads.
+ * Basalt carries the two `.basalt` artifacts the engine serves (the
+ * Qwen3.8-Flash-Next NVFP4 conversion and the DeepSeek-V4.1-Flash UD-IQ1_M
+ * pack, each named by the repo and file the adapter's deploy step
+ * downloads).
  *
  * The list is the single source the card's Model dropdown, the model
  * presence check, and the hf CLI download route read. Adding a model is a
  * data edit here, never a route or card rewire.
  * @module dsh-plugin-local-infer/providers/model-defaults
  */
+import { BASALT_DEFAULT_MODEL_ID, BASALT_MODEL_FILE, BASALT_MODEL_REPO, BASALT_DEEPSEEK_MODEL_FILE, BASALT_DEEPSEEK_MODEL_ID, BASALT_DEEPSEEK_MODEL_REPO, } from "./engines/basalt/basalt.js";
 /**
  * DEFAULT_MODELS is the default model list, in the order the card lists them.
  * The 27B keys reuse the VramModel registry ids; the 35B-A3B entries use the
@@ -88,11 +94,47 @@ export const DEFAULT_MODELS = [
         url: 'https://huggingface.co/neroued/Qwen3.6-35B-A3B-NInfer',
         file: 'qwen3_6_35b_a3b.ninfer',
     },
+    // The two models the Strata README's Models section links: Gyro (the
+    // smallest way to run Qwen3.8-Flash-Next) and DeepSeek V4.1-Flash (the
+    // bigger model that runs on one 12 GB card). No single file: the engine
+    // downloads the model itself (into the container volume or its data dir),
+    // so the entry names the repo only.
+    {
+        engineId: 'strata',
+        modelId: 'gyro-s',
+        label: 'Strata Gyro-S (Qwen3.8-Flash-Next)',
+        url: 'https://huggingface.co/agentionai/Qwen3.8-Flash-Next-Gyro-GGUF',
+    },
+    {
+        engineId: 'strata',
+        modelId: 'ist-das-2.00bpw',
+        label: 'Strata DeepSeek V4.1-Flash (2.00 bpw)',
+        url: 'https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF',
+    },
+    // The two `.basalt` artifacts the engine serves: for each, the weights,
+    // tokenizer, chat template, drafter, vision projector, and default
+    // sampling all come from the one file. The repos and files ride the
+    // adapter constants, so the download route and the deploy step share one
+    // source.
+    {
+        engineId: 'basalt',
+        modelId: BASALT_DEFAULT_MODEL_ID,
+        label: 'Basalt Qwen3.8-Flash-Next (NVFP4 DQ6_K)',
+        url: `https://huggingface.co/${BASALT_MODEL_REPO}`,
+        file: BASALT_MODEL_FILE,
+    },
+    {
+        engineId: 'basalt',
+        modelId: BASALT_DEEPSEEK_MODEL_ID,
+        label: 'Basalt DeepSeek-V4.1-Flash (UD-IQ1_M)',
+        url: `https://huggingface.co/${BASALT_DEEPSEEK_MODEL_REPO}`,
+        file: BASALT_DEEPSEEK_MODEL_FILE,
+    },
 ];
 /**
  * defaultModelsForEngine reads the default list for one engine id, in list
- * order. An engine with no entry (Strata) reads as an empty list, which the
- * card renders as no Model row.
+ * order. An engine with no entry reads as an empty list, which the card
+ * renders as no Model row.
  */
 export function defaultModelsForEngine(engineId) {
     return DEFAULT_MODELS.filter((entry) => entry.engineId === engineId);
@@ -105,4 +147,4 @@ export function defaultModelsForEngine(engineId) {
 export function defaultModelFor(engineId, modelId) {
     return DEFAULT_MODELS.find((entry) => entry.engineId === engineId && entry.modelId === modelId);
 }
-//# sourceMappingURL=model-defaults.js.map
+//# sourceMappingURL=model-defaults.js.map

package/packages/providers/lib/platform.d.ts

@@ -19,13 +19,16 @@ export interface EnginePlatformDefaults {
  * The per-engine platform defaults, keyed by engine id. The build command is
  * the engine's first acquisition build step; the docker tag is the upstream
  * image the adapter pulls. Hyper-Qwen is compose-driven, so its docker tag is
- * the vLLM OpenAI server image; Infernix ships no docker image, so it carries
- * a build command only.
+ * the vLLM OpenAI server image; Infernix and Basalt ship no docker image, so
+ * they carry a build command only.
  */
 export declare const ENGINE_PLATFORM_DEFAULTS: Record<string, EnginePlatformDefaults>;
 /**
  * enginePlatformDefaults reads the registered platform defaults for an engine
- * id, or an empty object when none is registered.
+ * id on an OS, or an empty object when none is registered. The build command
+ * is the engine's first build step for the OS (the Windows vcpkg route on a
+ * Windows host), so the card's placeholder seeds the OS-appropriate default.
+ * The docker tag is OS-independent and rides the registered entry.
  */
-export declare function enginePlatformDefaults(engineId: string): EnginePlatformDefaults;
-//# sourceMappingURL=platform.d.ts.map
+export declare function enginePlatformDefaults(engineId: string, os?: 'linux' | 'windows' | 'macos'): EnginePlatformDefaults;
+//# sourceMappingURL=platform.d.ts.map

package/packages/providers/lib/platform.js

@@ -5,7 +5,9 @@
  * defaults are a data edit here, never a card rewire.
  * @module dsh-plugin-local-infer/providers/platform
  */
-import { LLAMACPP_DOCKER_IMAGE, NINFER_DOCKER_IMAGE, STRATA_DOCKER_IMAGE, } from "./adapters.js";
+import { LLAMACPP_DOCKER_IMAGE } from "./engines/llamacpp/llamacppAdapter.js";
+import { NINFER_DOCKER_IMAGE } from "./engines/ninfer/ninferAdapter.js";
+import { STRATA_DOCKER_IMAGE } from "./engines/strata/strataAdapter.js";
 import { engineBuildSteps } from "./acquire.js";
 // The local build command is the engine's first acquisition build step (the
 // cmake build the deploy step runs in the checkout), read through
@@ -15,12 +17,14 @@ import { engineBuildSteps } from "./acquire.js";
 const NINFER_BUILD_COMMAND = engineBuildSteps('ninfer')[0];
 const LLAMACPP_BUILD_COMMAND = engineBuildSteps('llamacpp')[0];
 const INFERNIX_BUILD_COMMAND = engineBuildSteps('infernix')[0];
+const STRATA_BUILD_COMMAND = engineBuildSteps('strata')[0];
+const BASALT_BUILD_COMMAND = engineBuildSteps('basalt')[0];
 /**
  * The per-engine platform defaults, keyed by engine id. The build command is
  * the engine's first acquisition build step; the docker tag is the upstream
  * image the adapter pulls. Hyper-Qwen is compose-driven, so its docker tag is
- * the vLLM OpenAI server image; Infernix ships no docker image, so it carries
- * a build command only.
+ * the vLLM OpenAI server image; Infernix and Basalt ship no docker image, so
+ * they carry a build command only.
  */
 export const ENGINE_PLATFORM_DEFAULTS = {
     ninfer: {
@@ -45,14 +49,30 @@ export const ENGINE_PLATFORM_DEFAULTS = {
             : {}),
     },
     strata: {
+        ...(STRATA_BUILD_COMMAND !== undefined
+            ? { defaultBuildCommand: STRATA_BUILD_COMMAND }
+            : {}),
         defaultDockerTag: STRATA_DOCKER_IMAGE,
     },
+    basalt: {
+        // The CMake build command the deploy step runs in the checkout; the
+        // engine ships no container image, so no docker tag.
+        ...(BASALT_BUILD_COMMAND !== undefined ? { defaultBuildCommand: BASALT_BUILD_COMMAND } : {}),
+    },
 };
 /**
  * enginePlatformDefaults reads the registered platform defaults for an engine
- * id, or an empty object when none is registered.
+ * id on an OS, or an empty object when none is registered. The build command
+ * is the engine's first build step for the OS (the Windows vcpkg route on a
+ * Windows host), so the card's placeholder seeds the OS-appropriate default.
+ * The docker tag is OS-independent and rides the registered entry.
  */
-export function enginePlatformDefaults(engineId) {
-    return ENGINE_PLATFORM_DEFAULTS[engineId] ?? {};
+export function enginePlatformDefaults(engineId, os = 'linux') {
+    const base = ENGINE_PLATFORM_DEFAULTS[engineId] ?? {};
+    const buildCommand = engineBuildSteps(engineId, os)[0];
+    return {
+        ...base,
+        ...(buildCommand !== undefined ? { defaultBuildCommand: buildCommand } : {}),
+    };
 }
-//# sourceMappingURL=platform.js.map
+//# sourceMappingURL=platform.js.map

package/packages/providers/lib/presets.d.ts

@@ -2,10 +2,12 @@
  * Built-in preset data for the local-inference defaults layer. The presets are
  * the offline floor the DefaultsResolver resolves against: the NInfer
  * measured configs (the `config-calculator.html` figures and the
- * `qwen-ninfer` launcher, on a 24 GiB sm86 card) and the Hyper-Qwen `.env`
+ * `qwen-ninfer` launcher, on a 24 GiB sm86 card), the Hyper-Qwen `.env`
  * profiles (the single-user speed, long-context, and huge-context tiers from
- * `.env.example`). Each preset is keyed by its exact (engine, model, hardware
- * class) triple, so it is only trusted for the hardware it was sourced for.
+ * `.env.example`), and the Basalt Blackwell long-context starting config
+ * (keyed to the wildcard class, since Basalt runs on any Blackwell card).
+ * Each preset is keyed by its exact (engine, model, hardware class) triple, so
+ * it is only trusted for the hardware it was sourced for.
  *
  * The `PRESET_SOURCES` list names the mod-search fetch path: the URL the
  * bridge pulls and the pure extractor that turns the fetched page into a
@@ -32,4 +34,4 @@ export declare function extractJsonBlock(content: string): Config;
  * block the docs embed, so a page without one degrades to an empty config.
  */
 export declare const PRESET_SOURCES: readonly PresetSource[];
-//# sourceMappingURL=presets.d.ts.map
+//# sourceMappingURL=presets.d.ts.map