dsh-plugin-local-infer 0.1.1 → 0.1.2

package/packages/providers/lib/acquire.js

@@ -0,0 +1,325 @@
+/**
+ * The per-engine acquisition data and the Get step. Each engine declares its
+ * git source (the upstream repo, the branch the rebase targets, and the build
+ * steps the checkout runs) as data here; the adapters' deploy steps and the
+ * /local-infer/get route both consume the same constants, so a build command
+ * has one home. The sync step clones a missing checkout, and a clean existing
+ * checkout is fetched and rebased onto the upstream branch; a checkout with
+ * local changes aborts, because a rebase would eat the user's work.
+ * @module dsh-plugin-local-infer/providers/acquire
+ */
+/**
+ * The remote name a user-provided fork is added under, so the checkout's own
+ * origin (the upstream it was cloned from) survives the fork override.
+ */
+export const FORK_REMOTE = 'dsh-fork';
+/**
+ * The per-engine git sources, keyed by engine id. NInfer and Infernix build
+ * through cmake (Infernix through its own preset); llama.cpp builds through
+ * the same cmake shape as NInfer; Hyper-Qwen is the compose-driven vLLM
+ * engine, so it carries no local build steps (the compose up pulls the
+ * image); Strata is the placeholder engine, and its repo and pull step land
+ * with the engine, the same way its docker image constant did.
+ */
+export const ENGINE_SOURCES = {
+    ninfer: {
+        repo: 'https://github.com/Neroued/ninfer',
+        branch: 'master',
+        buildSteps: [
+            'cmake -S {dir} -B {dir}/build -G Ninja -DCMAKE_BUILD_TYPE=Release && cmake --build {dir}/build -j',
+        ],
+    },
+    hyperqwen: {
+        repo: 'https://github.com/vllm-project/vllm',
+        branch: 'main',
+        buildSteps: [],
+    },
+    llamacpp: {
+        repo: 'https://github.com/ggml-org/llama.cpp',
+        branch: 'master',
+        buildSteps: [
+            'cmake -S {dir} -B {dir}/build -G Ninja -DCMAKE_BUILD_TYPE=Release && cmake --build {dir}/build -j',
+        ],
+    },
+    infernix: {
+        repo: 'https://github.com/Wallawalla47/Infernix',
+        branch: 'main',
+        buildSteps: ['cd {dir} && cmake --preset release && cmake --build build -j'],
+    },
+    strata: {
+        repo: 'https://github.com/strata/strata',
+        branch: 'main',
+        buildSteps: ['docker pull ghcr.io/strata/strata'],
+    },
+};
+/**
+ * engineSource reads the registered acquisition data for an engine id, or
+ * undefined when none is registered.
+ */
+export function engineSource(engineId) {
+    return ENGINE_SOURCES[engineId];
+}
+/**
+ * engineBuildSteps reads the build steps an engine runs in its checkout, so
+ * the adapter's deploy step and the Get build step share one list.
+ */
+export function engineBuildSteps(engineId) {
+    return ENGINE_SOURCES[engineId]?.buildSteps ?? [];
+}
+/**
+ * runLogged runs one shell command through the host's streaming surface when
+ * it has one (echoing the command and streaming its output through log), and
+ * through plain run otherwise. The streamed copy is also collected, so a
+ * caller can price the command's stdout (the dirty-check, for one).
+ */
+async function runLogged(host, command, log, options) {
+    let output = '';
+    const collect = (chunk) => {
+        output += chunk;
+        log(chunk);
+    };
+    if (options?.echo !== false) {
+        log(`$ ${command}\n`);
+    }
+    if (host.runStream !== undefined) {
+        const outcome = await host.runStream(command, collect);
+        return { ...outcome, output };
+    }
+    const outcome = await host.run(command);
+    return { ...outcome, output };
+}
+/**
+ * syncEngineRepo brings the engine checkout to the listed source. A missing
+ * checkout is a fresh clone of the fork (or the engine's upstream repo when
+ * no fork is provided). An existing checkout with local changes aborts, so
+ * the user's work is never rebased away; a clean checkout fetches the listed
+ * fork (added as the named fork remote when provided, origin when not) and
+ * rebases onto its branch. Every command echoes into log as it runs.
+ */
+export async function syncEngineRepo(host, request, log) {
+    const source = engineSource(request.engineId);
+    if (source === undefined) {
+        return {
+            ok: false,
+            detail: `no engine source is registered for "${request.engineId}"`,
+        };
+    }
+    const directory = request.directory;
+    const fork = request.fork !== undefined && request.fork.trim() !== '' ? request.fork : undefined;
+    const present = await host.run(`[ -d ${directory}/.git ]`);
+    if (present.ok) {
+        const status = await runLogged(host, `git -C ${directory} status --short`, log);
+        if (!status.ok) {
+            return { ok: false, detail: `git status failed: ${status.detail ?? 'unknown'}` };
+        }
+        if (status.output.trim() !== '') {
+            log(`!! ${directory} has local changes. Get aborted; commit or revert them first.\n`);
+            return {
+                ok: false,
+                aborted: true,
+                detail: 'the engine repository has local changes',
+            };
+        }
+        if (fork !== undefined) {
+            // Add the fork under its own remote name; re-Get updates the URL in
+            // place so an edited Engine Fork field takes effect on the next run.
+            const add = await runLogged(host, `git -C ${directory} remote add ${FORK_REMOTE} ${fork}`, log);
+            if (!add.ok) {
+                await runLogged(host, `git -C ${directory} remote set-url ${FORK_REMOTE} ${fork}`, log);
+            }
+        }
+        const remote = fork !== undefined ? FORK_REMOTE : 'origin';
+        const fetch = await runLogged(host, `git -C ${directory} fetch ${remote}`, log);
+        if (!fetch.ok) {
+            return { ok: false, detail: `git fetch ${remote} failed: ${fetch.detail ?? 'unknown'}` };
+        }
+        const rebase = await runLogged(host, `git -C ${directory} rebase ${remote}/${source.branch} || ` +
+            `{ git -C ${directory} rebase --abort; exit 1; }`, log);
+        if (!rebase.ok) {
+            return {
+                ok: false,
+                detail: `git rebase ${remote}/${source.branch} failed: ${rebase.detail ?? 'unknown'}`,
+            };
+        }
+        return { ok: true };
+    }
+    const clone = await runLogged(host, `git clone ${fork ?? source.repo} ${directory}`, log);
+    if (!clone.ok) {
+        return { ok: false, detail: `git clone failed: ${clone.detail ?? 'unknown'}` };
+    }
+    return { ok: true };
+}
+/**
+ * syncAndRebuildEngine brings the engine checkout to the listed source and
+ * rebuilds it only when it needs to. A missing checkout is a fresh clone of
+ * the fork (or the upstream repo) and always builds. An existing checkout is
+ * fetched and priced against the remote branch: when it is behind and clean,
+ * the primary branch is checked out, rebased onto the remote branch, and the
+ * working tree is cleaned (git clean -fxd) before the build; when it is not
+ * behind, or it carries local changes, the build is skipped so a launch does
+ * not rebuild a checkout that is current. The decision, the build command,
+ * and its output stream through log and are returned as the result's logs.
+ * After a rebuild, the engine binary is checked against the path; when a build
+ * that did not install it leaves it out, a symlink into /usr/local/bin is
+ * added (run when the trust gate is on, written to the logs when it is off).
+ */
+export async function syncAndRebuildEngine(host, request, log) {
+    const source = engineSource(request.engineId);
+    if (source === undefined) {
+        return {
+            ok: false,
+            rebuilt: false,
+            detail: `no engine source is registered for "${request.engineId}"`,
+            logs: '',
+        };
+    }
+    const directory = request.directory;
+    const fork = request.fork !== undefined && request.fork.trim() !== '' ? request.fork : undefined;
+    const remote = fork !== undefined ? FORK_REMOTE : 'origin';
+    const chunks = [];
+    const emit = (chunk) => {
+        chunks.push(chunk);
+        log(chunk);
+    };
+    const present = await host.run(`[ -d ${directory}/.git ]`);
+    let rebuilt = false;
+    if (!present.ok) {
+        // A missing checkout is a fresh clone of the fork (or the upstream repo),
+        // which always needs a build.
+        emit(`>> cloning ${fork ?? source.repo} into ${directory}\n`);
+        const clone = await runLogged(host, `git clone ${fork ?? source.repo} ${directory}`, emit);
+        if (!clone.ok) {
+            return {
+                ok: false,
+                rebuilt: false,
+                detail: `git clone failed: ${clone.detail ?? 'unknown'}`,
+                logs: chunks.join(''),
+            };
+        }
+        rebuilt = true;
+    }
+    else {
+        if (fork !== undefined) {
+            // Add the fork under its own remote name; a re-launch updates the URL in
+            // place so an edited fork field takes effect on the next run.
+            const add = await runLogged(host, `git -C ${directory} remote add ${FORK_REMOTE} ${fork}`, emit, { echo: false });
+            if (!add.ok) {
+                await runLogged(host, `git -C ${directory} remote set-url ${FORK_REMOTE} ${fork}`, emit, { echo: false });
+            }
+        }
+        const status = await runLogged(host, `git -C ${directory} status --short`, emit, { echo: false });
+        if (!status.ok) {
+            return {
+                ok: false,
+                rebuilt: false,
+                detail: `git status failed: ${status.detail ?? 'unknown'}`,
+                logs: chunks.join(''),
+            };
+        }
+        if (status.output.trim() !== '') {
+            emit(`!! ${directory} has local changes; the rebase and rebuild are skipped.\n`);
+            return { ok: true, rebuilt: false, logs: chunks.join('') };
+        }
+        // The fetch is best-effort: an offline host prices the checkout it has.
+        await runLogged(host, `git -C ${directory} fetch ${remote} || true`, emit, { echo: false });
+        // Price whether the current branch is behind the remote branch. A fetch
+        // that failed (offline) or a missing remote branch reads as not-behind, so
+        // the checkout builds as-is.
+        const count = await runLogged(host, `git -C ${directory} rev-list --count HEAD..${remote}/${source.branch} 2>/dev/null || echo 0`, emit, { echo: false });
+        const behind = (count.output.trim() || '0') !== '0';
+        if (behind) {
+            emit(`>> ${directory} is behind ${remote}/${source.branch}; rebasing and rebuilding.\n`);
+            const checkout = await runLogged(host, `git -C ${directory} checkout ${source.branch}`, emit);
+            if (!checkout.ok) {
+                return {
+                    ok: false,
+                    rebuilt: false,
+                    detail: `git checkout ${source.branch} failed: ${checkout.detail ?? 'unknown'}`,
+                    logs: chunks.join(''),
+                };
+            }
+            const rebase = await runLogged(host, `git -C ${directory} rebase ${remote}/${source.branch} || ` +
+                `{ git -C ${directory} rebase --abort; exit 1; }`, emit);
+            if (!rebase.ok) {
+                return {
+                    ok: false,
+                    rebuilt: false,
+                    detail: `git rebase ${remote}/${source.branch} failed: ${rebase.detail ?? 'unknown'}`,
+                    logs: chunks.join(''),
+                };
+            }
+            const clean = await runLogged(host, `git -C ${directory} clean -fxd`, emit);
+            if (!clean.ok) {
+                return {
+                    ok: false,
+                    rebuilt: false,
+                    detail: `git clean -fxd failed: ${clean.detail ?? 'unknown'}`,
+                    logs: chunks.join(''),
+                };
+            }
+            rebuilt = true;
+        }
+        else {
+            emit(`>> ${directory} is up to date with ${remote}/${source.branch}; the rebuild is skipped.\n`);
+        }
+    }
+    if (rebuilt) {
+        // Run the engine's build steps, capturing the command and its output.
+        for (const step of source.buildSteps) {
+            const command = step.replaceAll('{dir}', directory);
+            const outcome = await runLogged(host, command, emit);
+            if (!outcome.ok) {
+                return {
+                    ok: false,
+                    rebuilt: true,
+                    detail: `build failed: ${outcome.detail ?? 'unknown'}`,
+                    logs: chunks.join(''),
+                };
+            }
+        }
+        // A rebuild that did not install the binary (no make install) leaves it
+        // out of the path. Check, and add a symlink when the trust gate allows.
+        const binary = request.binary;
+        if (binary !== undefined) {
+            const which = await host.run(`command -v ${binary}`);
+            if (!which.ok) {
+                const link = `sudo ln -sf ${directory}/build/${binary} /usr/local/bin/${binary}`;
+                if (request.allowRunDeps === true) {
+                    const result = await runLogged(host, link, emit);
+                    if (!result.ok) {
+                        emit(`!! ${link} failed: ${result.detail ?? 'unknown'}\n`);
+                    }
+                }
+                else {
+                    emit(`!! ${binary} is not on the path; run ${link} (the provider command must be on the path to launch)\n`);
+                }
+            }
+        }
+    }
+    return { ok: true, rebuilt, logs: chunks.join('') };
+}
+/**
+ * buildEngineRepo runs the engine's build steps in the checkout, streaming
+ * each step's output through log. An engine with no build steps (the compose
+ * engines) passes through with a note; a failing step stops the build with
+ * the step's detail.
+ */
+export async function buildEngineRepo(host, engineId, directory, log) {
+    const source = engineSource(engineId);
+    if (source === undefined) {
+        return { ok: false, detail: `no engine source is registered for "${engineId}"` };
+    }
+    if (source.buildSteps.length === 0) {
+        log(`no local build steps for ${engineId}\n`);
+        return { ok: true };
+    }
+    for (const step of source.buildSteps) {
+        const command = step.replaceAll('{dir}', directory);
+        const outcome = await runLogged(host, command, log);
+        if (!outcome.ok) {
+            return { ok: false, detail: `build failed: ${outcome.detail ?? 'unknown'}` };
+        }
+    }
+    return { ok: true };
+}
+//# sourceMappingURL=acquire.js.map

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

@@ -8,7 +8,8 @@
  * @module dsh-plugin-local-infer/providers/adapters
  */
 import { EngineAdapter } from '../../core/lib/index.js';
-import type { Config, DebugLogger, DeployResult, EngineLaunchSignature, EnginePlatformConfig, HealthCheckOptions, HealthResult, HardwareProfile, LaunchPlan, LifecycleHost, ParameterSpec } from '../../core/lib/index.js';
+import type { Config, DebugLogger, DeployOptions, DeployResult, EngineLaunchSignature, EnginePlatformConfig, HealthCheckOptions, HealthResult, HardwareProfile, LaunchPlan, LaunchProgress, LifecycleHost, ParameterSpec } from '../../core/lib/index.js';
+export { INFERNIX_MODEL_ARTIFACTS } from './infernix.ts';
 /**
  * NInferModelArtifact is the per-model deploy data: the `.ninfer` artifact
  * file, the Hugging Face repo the model download pulls it from, and the
@@ -111,7 +112,7 @@ export declare class NInferAdapter extends EngineAdapter {
     };
     launchSignature(config?: Config): EngineLaunchSignature;
     renderLaunch(config: Config, _hw: HardwareProfile): LaunchPlan;
-    deploy(plan: LaunchPlan, platform?: EnginePlatformConfig): Promise<DeployResult>;
+    deploy(plan: LaunchPlan, platform?: EnginePlatformConfig, onProgress?: LaunchProgress, options?: DeployOptions): Promise<DeployResult>;
     healthCheck(plan: LaunchPlan, options?: HealthCheckOptions): Promise<HealthResult>;
     stop(plan: LaunchPlan): Promise<void>;
     private buildMode;
@@ -150,7 +151,7 @@ export declare class HyperQwenAdapter extends EngineAdapter {
     };
     launchSignature(config?: Config): EngineLaunchSignature;
     renderLaunch(config: Config, _hw: HardwareProfile): LaunchPlan;
-    deploy(plan: LaunchPlan): Promise<DeployResult>;
+    deploy(plan: LaunchPlan, platform?: EnginePlatformConfig, onProgress?: LaunchProgress, _options?: DeployOptions): Promise<DeployResult>;
     healthCheck(plan: LaunchPlan, options?: HealthCheckOptions): Promise<HealthResult>;
     stop(plan: LaunchPlan): Promise<void>;
     private port;
@@ -186,7 +187,7 @@ export declare class LlamaCppAdapter extends EngineAdapter {
     };
     launchSignature(config?: Config): EngineLaunchSignature;
     renderLaunch(config: Config, _hw: HardwareProfile): LaunchPlan;
-    deploy(plan: LaunchPlan, platform?: EnginePlatformConfig): Promise<DeployResult>;
+    deploy(plan: LaunchPlan, platform?: EnginePlatformConfig, onProgress?: LaunchProgress, options?: DeployOptions): Promise<DeployResult>;
     healthCheck(plan: LaunchPlan, options?: HealthCheckOptions): Promise<HealthResult>;
     stop(plan: LaunchPlan): Promise<void>;
     private buildMode;
@@ -223,10 +224,46 @@ export declare class StrataAdapter extends EngineAdapter {
     };
     launchSignature(config?: Config): EngineLaunchSignature;
     renderLaunch(config: Config, _hw: HardwareProfile): LaunchPlan;
-    deploy(plan: LaunchPlan): Promise<DeployResult>;
+    deploy(plan: LaunchPlan, _platform?: EnginePlatformConfig, onProgress?: LaunchProgress, _options?: DeployOptions): Promise<DeployResult>;
     healthCheck(plan: LaunchPlan, options?: HealthCheckOptions): Promise<HealthResult>;
     stop(plan: LaunchPlan): Promise<void>;
     private port;
     private planKey;
 }
+/**
+ * InfernixAdapter is the local-only engine for Infernix, the C++/CUDA fork
+ * of NInfer (https://github.com/wallawalla47/infernix) tuned for Qwen models
+ * on a single RTX 50-series GPU. `renderLaunch` renders a local
+ * `infernix-serve` process from the `INFERNIX_PARAMETERS` surface (the
+ * Infernix-only fields, from `--ngram-volume` to the quality routes, spell
+ * onto the argv only for this engine); `deploy` builds the engine in the
+ * engine directory (the engine's own cmake preset), downloads the model
+ * artifact (and the n-gram volume, for the multi-file Flash-Next repo) into
+ * the model directory, and starts the engine; `healthCheck` polls the health
+ * URL; `stop` signals the tracked process. The engine ships no docker image,
+ * so the adapter supports the local build mode only.
+ */
+export declare class InfernixAdapter extends EngineAdapter {
+    readonly id = "infernix";
+    readonly supports: {
+        docker: boolean;
+        local: boolean;
+    };
+    private readonly host;
+    private readonly children;
+    constructor(host?: LifecycleHost);
+    parameters(): readonly ParameterSpec[];
+    modelRef(config: Config): {
+        id: string;
+        name: string;
+    };
+    launchSignature(config?: Config): EngineLaunchSignature;
+    renderLaunch(config: Config, _hw: HardwareProfile): LaunchPlan;
+    deploy(plan: LaunchPlan, platform?: EnginePlatformConfig, onProgress?: LaunchProgress, options?: DeployOptions): Promise<DeployResult>;
+    healthCheck(plan: LaunchPlan, options?: HealthCheckOptions): Promise<HealthResult>;
+    stop(plan: LaunchPlan): Promise<void>;
+    private port;
+    private planKey;
+    private downloadModel;
+}
 //# sourceMappingURL=adapters.d.ts.map