dsh-plugin-local-infer 0.1.0 → 0.1.2

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

@@ -5,9 +5,9 @@
  * concrete EngineAdapter (Phase 5) composes it into a LaunchPlan. The registry
  * is the single list the Web card offers, so adding an engine is a data edit
  * here, never a rewire of the card.
- * @module @dsh/dsh-plugin-local-infer/providers/engines
+ * @module dsh-plugin-local-infer/providers/engines
  */
-import type { ParameterSpec } from '@dsh/dsh-plugin-local-infer/core';
+import type { ParameterSpec } from 'dsh-plugin-local-infer/core';
 export interface EngineInfo {
     id: string;
     label: string;
@@ -19,9 +19,10 @@ export interface EngineInfo {
 }
 /**
  * The engines the Web card offers, in the order the card lists them. NInfer,
- * Hyper-Qwen, 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, 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.
  */
 export declare const ENGINES: readonly EngineInfo[];
 //# sourceMappingURL=engines.d.ts.map

package/packages/providers/lib/engines.js

@@ -5,17 +5,19 @@
  * concrete EngineAdapter (Phase 5) composes it into a LaunchPlan. The registry
  * is the single list the Web card offers, so adding an engine is a data edit
  * here, never a rewire of the card.
- * @module @dsh/dsh-plugin-local-infer/providers/engines
+ * @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";
 /**
  * The engines the Web card offers, in the order the card lists them. NInfer,
- * Hyper-Qwen, 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, 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.
  */
 export const ENGINES = [
     {
@@ -36,6 +38,12 @@ export const ENGINES = [
         supports: { docker: true, local: true },
         parameters: LLAMACPP_PARAMETERS,
     },
+    {
+        id: 'infernix',
+        label: 'Infernix',
+        supports: { docker: false, local: true },
+        parameters: INFERNIX_PARAMETERS,
+    },
     {
         id: 'strata',
         label: 'Strata',

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

@@ -9,9 +9,9 @@
  * "exceeds usable VRAM" and "spec needs draft weights" rules) live in the
  * core `modelGuidelines` and run in the service, because they price against
  * a model, not against the hardware alone.
- * @module @dsh/dsh-plugin-local-infer/providers/guidelines
+ * @module dsh-plugin-local-infer/providers/guidelines
  */
-import type { Guideline } from '@dsh/dsh-plugin-local-infer/core';
+import type { Guideline } from 'dsh-plugin-local-infer/core';
 /**
  * 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

package/packages/providers/lib/guidelines.js

@@ -9,8 +9,9 @@
  * "exceeds usable VRAM" and "spec needs draft weights" rules) live in the
  * core `modelGuidelines` and run in the service, because they price against
  * a model, not against the hardware alone.
- * @module @dsh/dsh-plugin-local-infer/providers/guidelines
+ * @module dsh-plugin-local-infer/providers/guidelines
  */
+import { infernixDraftModelPath } from "./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
@@ -57,5 +58,20 @@ export const BUILTIN_GUIDELINES = [
         anchors: ['pinMemory'],
         when: (config, hw) => hw.wsl2 && config.spec === 'dflash2' && config.pinMemory !== true,
     },
+    {
+        // The Infernix dflash2 backend always loads a drafter through
+        // --draft-model, so a config that selects dflash2 but names no drafter
+        // path (no preset, no Custom path) cannot boot. The rule is Infernix-
+        // specific: only Infernix configs carry the draftModelPreset key, so a
+        // NInfer dflash2 (which rides the model's own companion weights) never
+        // fires it.
+        id: 'infernix-dflash2-needs-draft-model',
+        level: 'error',
+        message: 'dflash2 needs a draft model. Select a DFlash2 model, or switch to Custom and set the Draft model path.',
+        anchors: ['draftModelPreset', 'draftModel'],
+        when: (config) => config.spec === 'dflash2' &&
+            'draftModelPreset' in config &&
+            infernixDraftModelPath(config) === null,
+    },
 ];
 //# sourceMappingURL=guidelines.js.map

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

@@ -4,9 +4,9 @@
  * `docker-compose.yml`, and the `single-user`/`batch` start scripts. The
  * concepts mirror the NInfer surface; the spelling is a `.env` key the
  * `HyperQwenAdapter` (Phase 5) writes into the compose environment.
- * @module @dsh/dsh-plugin-local-infer/providers/hyperqwen
+ * @module dsh-plugin-local-infer/providers/hyperqwen
  */
-import type { EnumOption, ParameterSpec } from '@dsh/dsh-plugin-local-infer/core';
+import type { EnumOption, ParameterSpec } from 'dsh-plugin-local-infer/core';
 /**
  * The speculative-decoding backends Hyper-Qwen serves. The launcher's own
  * default is `mtp`; the single-user speed profile sets `dflash2`.

package/packages/providers/lib/hyperqwen.js

@@ -4,7 +4,7 @@
  * `docker-compose.yml`, and the `single-user`/`batch` start scripts. The
  * concepts mirror the NInfer surface; the spelling is a `.env` key the
  * `HyperQwenAdapter` (Phase 5) writes into the compose environment.
- * @module @dsh/dsh-plugin-local-infer/providers/hyperqwen
+ * @module dsh-plugin-local-infer/providers/hyperqwen
  */
 /**
  * The speculative-decoding backends Hyper-Qwen serves. The launcher's own

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

@@ -1,14 +1,14 @@
 /**
  * The concrete providers for the local-inference capability: the engine
- * adapters (ninfer, hyperqwen, llamacpp, strata), 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`.
+ * adapters (ninfer, hyperqwen, llamacpp, infernix, strata), 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`.
  *
  * Phase 3 registers the NInfer measured VramModels and the built-in hardware
  * guidelines. The engine adapters land in Phase 5. Phase 7 registers the
  * built-in preset data (the NInfer measured and Hyper-Qwen profile presets)
  * that the defaults resolver resolves against.
- * @module @dsh/dsh-plugin-local-infer/providers
+ * @module dsh-plugin-local-infer/providers
  */
 import type { Context } from '@deepseek-ai/cordis';
 export declare const name = "local-infer-providers";
@@ -24,13 +24,19 @@ export { NINFER_VRAM_MODELS, 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_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';
 /**
  * Mount the providers. The NInfer measured VramModels register here so the

package/packages/providers/lib/index.js

@@ -1,16 +1,16 @@
 /**
  * The concrete providers for the local-inference capability: the engine
- * adapters (ninfer, hyperqwen, llamacpp, strata), 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`.
+ * adapters (ninfer, hyperqwen, llamacpp, infernix, strata), 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`.
  *
  * Phase 3 registers the NInfer measured VramModels and the built-in hardware
  * guidelines. The engine adapters land in Phase 5. Phase 7 registers the
  * built-in preset data (the NInfer measured and Hyper-Qwen profile presets)
  * that the defaults resolver resolves against.
- * @module @dsh/dsh-plugin-local-infer/providers
+ * @module dsh-plugin-local-infer/providers
  */
-import { createNodeLifecycleHost, HyperQwenAdapter, LlamaCppAdapter, NInferAdapter, StrataAdapter, } from "./adapters.js";
+import { createNodeLifecycleHost, HyperQwenAdapter, InfernixAdapter, LlamaCppAdapter, NInferAdapter, StrataAdapter, } from "./adapters.js";
 import { createDshDebugLogger } from "./debug.js";
 import { ENGINE_DEP_PLANS } from "./deps.js";
 import { BUILTIN_GUIDELINES } from "./guidelines.js";
@@ -28,11 +28,14 @@ export { CHAT_TEMPLATE_OPTIONS, DEVICE_PROFILE_OPTIONS, KV_DTYPE_OPTIONS, NINFER
 export { NINFER_VRAM_MODELS, 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_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";
 /**
  * Mount the providers. The NInfer measured VramModels register here so the
@@ -68,6 +71,7 @@ export function apply(ctx) {
     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));
 }
 const plugin = { name, inject, apply };

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

@@ -0,0 +1,170 @@
+/**
+ * Infernix parameter surface. The declarative `ParameterSpec` data for the
+ * Infernix engine (https://github.com/wallawalla47/infernix), a C++/CUDA
+ * fork of NInfer tuned for Qwen models on a single RTX 50-series GPU,
+ * sourced from the project's `docs/cli.md` and `docs/serving.md`. The
+ * spelling is an `infernix-serve` CLI flag the `InfernixAdapter` passes onto
+ * the process.
+ *
+ * The surface carries the flags the NInfer surface shares (KV format,
+ * context, speculation, sampling) plus the Infernix-only fields: the
+ * `--ngram-volume` PLE table for Qwen3.8-Flash-Next, the `--draft-model`
+ * DFlash2 companion drafter, the n-gram copy proposal set, the offloaded-
+ * expert memory tiers, and the quality routes. Those fields live only on
+ * this surface, so the form renderer shows them only while the Infernix
+ * engine is selected, and the adapters of the other engines never spell
+ * them. Adding a flag here is a data edit, never a core rewire.
+ * @module dsh-plugin-local-infer/providers/infernix
+ */
+import type { Config, EnumOption, ParameterSpec } from 'dsh-plugin-local-infer/core';
+/**
+ * The KV cache quantization formats Infernix serves, in the order the form
+ * dropdown lists them. The set differs from NInfer's: the 2-bit and 4-bit-
+ * K vector-quantized tiers (vq2, k4v2) replace the KVarN rk tiers, and both
+ * hold three to four times the context of INT8 in the same VRAM.
+ */
+export declare const INFERNIX_KV_DTYPE_OPTIONS: readonly EnumOption[];
+/**
+ * The speculative-decoding backends. `none` loads no speculative backend at
+ * all, which is the smallest resident profile.
+ */
+export declare const INFERNIX_SPEC_OPTIONS: readonly EnumOption[];
+/**
+ * The PLE n-gram table residency modes for Qwen3.8-Flash-Next. `auto` takes
+ * pinned RAM only when the RAM ledger keeps every expert in RAM beside it.
+ */
+export declare const INFERNIX_NGRAM_RESIDENCY_OPTIONS: readonly EnumOption[];
+/**
+ * The reasoning-loop guard modes. The guard ends a reply, or closes the
+ * thinking, when the open thinking keeps repeating whole passages.
+ */
+export declare const INFERNIX_REASONING_LOOP_OPTIONS: readonly EnumOption[];
+/**
+ * The gate message the form shows while the speculation fields are greyed
+ * out. It names the fix (pick a backend) and the backends that unlock the
+ * fields, so a first-time user knows the next click, not just the constraint.
+ */
+export declare const SPEC_GATE_MESSAGE = "Pick a speculation backend first (mtp, dflash, or dflash2)";
+/**
+ * The gate message the form shows while the dflash2-only fields are greyed
+ * out. It names the fix (pick dflash2) and the reason, so a first-time user
+ * knows the next click, not just the constraint.
+ */
+export declare const DFLASH2_GATE_MESSAGE = "Pick the dflash2 speculation backend first; the DFlash2 drafter loads with dflash2";
+/**
+ * InfernixModelArtifact is the per-model deploy data: the `.infernix` (or
+ * `.ninfer`) artifact files, the Hugging Face repo the model download pulls
+ * them from, and the model id the server advertises on its OpenAI endpoint.
+ * A multi-file artifact (Qwen3.8-Flash-Next) names every shard in `files`
+ * and the engine loads the first shard from the model directory. The data
+ * is the surface; adding a model is a data edit here, never an adapter
+ * rewire.
+ */
+export interface InfernixModelArtifact {
+    file: string;
+    files?: string[];
+    repo: string;
+    modelId: string;
+}
+/**
+ * The Infernix model artifacts, keyed by the config model key. The 27B
+ * NVIDIA NVFP4 entry is the default the DEFAULT_MODELS list names; an id no
+ * entry names falls back to it, so a bare config always renders a plan. The
+ * Flash-Next entry is the recommended Dense8 conversion: its download pulls
+ * the whole repo, which carries the FP8 n-gram volume beside the shards, so
+ * the `--ngram-volume` flag points at a file the same download left on disk.
+ */
+export declare const INFERNIX_MODEL_ARTIFACTS: Record<string, InfernixModelArtifact>;
+/**
+ * InfernixDFlash2Model is one common DFlash2 drafter the form's DFlash2 model
+ * dropdown offers: the `.infernix` artifact file the hf CLI pulls, the Hugging
+ * Face repo it lives in, and the label the dropdown shows. The DFlash2 drafter
+ * is a companion model that runs beside a Flash-Next conversion, so it
+ * downloads separately from the main model and loads through `--draft-model`.
+ * The data is the surface; adding a drafter is a data edit here, never an
+ * adapter rewire.
+ */
+export interface InfernixDFlash2Model {
+    file: string;
+    repo: string;
+    label: string;
+}
+/**
+ * The common DFlash2 drafters the Infernix DFlash2 model dropdown offers, in
+ * dropdown order. The Flash-Next DFlash2 is Inco AI's drafter (the 776 MB
+ * `.infernix` artifact) that the Infernix README recommends beside any
+ * Flash-Next conversion. The list is the single source the dropdown, the
+ * `--draft-model` path resolver, and the Get download step read.
+ */
+export declare const INFERNIX_DFLASH2_MODELS: readonly InfernixDFlash2Model[];
+/**
+ * The dropdown value the DFlash2 model field's Custom entry carries. A
+ * selection of this value (or no selection) makes the effective `--draft-model`
+ * path fall through to the Draft model path field, so a drafter no list entry
+ * names still loads.
+ */
+export declare const DFLASH2_CUSTOM_VALUE = "custom";
+/**
+ * dflash2PresetOptions renders the DFlash2 model dropdown options: one entry
+ * per common drafter (the label, the repo as the note) plus the Custom entry
+ * last. The option value is the drafter's artifact file, so the selection
+ * resolves back to the model data by file name.
+ */
+export declare function dflash2PresetOptions(): readonly EnumOption[];
+/**
+ * InfernixNgramVolume is the PLE n-gram table a model key reads: the `.ngram`
+ * file and the Hugging Face repo the Get step pulls it from. The volume is kept
+ * apart from the model shards and read from an NVMe drive, so the
+ * `--ngram-volume` flag points at a file the same download leaves on disk.
+ */
+export interface InfernixNgramVolume {
+    file: string;
+    repo: string;
+}
+/**
+ * The PLE n-gram volumes, keyed by the config model key. The Flash-Next entry
+ * is the FP8 volume (the recommended one), which ships beside the Dense8 model
+ * shards in the same repo, so a Get of the Flash-Next model pulls the table
+ * into the model directory.
+ */
+export declare const INFERNIX_NGRAM_VOLUMES: Record<string, InfernixNgramVolume>;
+/**
+ * infernixDraftModelPath resolves the effective `--draft-model` path a config
+ * carries: the selected DFlash2 preset's artifact (the model directory plus the
+ * drafter file, the layout the hf CLI download leaves) when a preset is
+ * selected, else the Custom Draft model path field. Returns null when neither
+ * names a path, which is the misconfiguration the dflash2 guideline names.
+ */
+export declare function infernixDraftModelPath(config: Config): string | null;
+/**
+ * infernixGetDownloads renders the hf CLI download commands a Get run pulls
+ * for a config: the selected DFlash2 drafter (into the model directory, the
+ * layout the `--draft-model` path expects) when a preset is selected, and the
+ * model key's PLE n-gram volume (into the model directory) when the key names
+ * one. The commands are the single source the Get route streams, so a Get
+ * leaves the model directory ready to launch.
+ */
+export declare function infernixGetDownloads(config: Config, modelDir: string): string[];
+/**
+ * The Infernix `ParameterSpec` surface, in form order. The `group` field is
+ * the form section; specs without a group render under `General`. The
+ * Infernix-only fields (ngramVolume, draftModel, the n-gram proposal set,
+ * the expert memory tiers, the quality routes) appear only on this surface,
+ * so the form shows them only while the Infernix engine is selected.
+ */
+export declare const INFERNIX_PARAMETERS: readonly ParameterSpec[];
+/**
+ * The Infernix model key a config carries: the string model value, falling
+ * back to the 27B artifact key so a bare config always renders a plan. A
+ * value that names a local artifact file (a directory-scanned model, a
+ * `.ninfer` or `.infernix` name) passes through unchanged.
+ */
+export declare function infernixModelKey(config: Config): string;
+/**
+ * infernixArtifactFor resolves the artifact a model key deploys. A key that
+ * names a local artifact file passes through the artifact map's 27B entry
+ * for the model id, and an id no entry names falls back to the 27B entry, so
+ * a bare config always renders a plan.
+ */
+export declare function infernixArtifactFor(modelKey: string): InfernixModelArtifact;
+//# sourceMappingURL=infernix.d.ts.map