Initial public ModelForge release
This commit is contained in:
Jens committed
commit
7082ab955a
490 files changed
+104252
No files matched your search
@@ -0,0 +1,186 @@
|
||||
import { cleanup, fireEvent, render, screen, waitFor } from "@testing-library/react";
|
||||
import { afterEach, describe, expect, it, vi } from "vitest";
|
||||
|
||||
import { CapabilityWorkspace } from "./CapabilityWorkspace";
|
||||
import type { CapabilityDeployment, SchedulerBudget } from "../types";
|
||||
|
||||
const deployment: CapabilityDeployment = {
|
||||
id: "deployment-1", capability: "rag.embedding", contract_version: 1,
|
||||
deployment_candidate_id: "candidate-1", production_approval_id: "approval-1",
|
||||
artifact_set_id: "artifact-set-1", runtime_profile_id: "profile-1",
|
||||
compute_node_id: "node-1", accelerator_id: "gpu-1",
|
||||
embedding_space: { id: "space-1", identity_digest: "a".repeat(64), dimension: 1024, normalized: true, migration_class: "requires_reindex", identity_facts: {}, created_at: "2026-08-25T12:00:00Z" },
|
||||
resource_envelope: { id: "envelope-1", runtime_probe_id: "probe-1", accelerator_kind: "NVIDIA RTX 4080 SUPER", accelerator_uuid: "GPU-test", environment_fingerprint: "b".repeat(64), concurrency: 1, batch_size: 8, max_sequence_length: 8192, baseline_vram_bytes: 1_100_000_000, resident_vram_bytes: 1_500_000_000, peak_vram_bytes: 2_700_000_000, required_vram_bytes: 1_600_000_000, cold_load_time_ms: 3690, inference_latency_ms: 1203, stale: false },
|
||||
residency: { id: "residency-1", state: "warm", worker_instance_id: "worker-1", load_count: 1, active_requests: 0, measured_resident_vram_bytes: 1_500_000_000, external_baseline_vram_bytes: 636_000_000, health: { functional: true }, resident_since: "2026-08-25T12:00:00Z", last_used_at: "2026-08-25T12:01:00Z" },
|
||||
channel: "stable", status: "stable", production: true, health_status: "healthy",
|
||||
routing_weight: 100, fallback_policy: {}, residency_policy: "keep_warm",
|
||||
keep_warm_seconds: 900, max_concurrency: 1, max_queue_depth: 16,
|
||||
config_fingerprint: "c".repeat(64), provenance: { runtime_probe_id: "probe-1", compatibility_assessment_id: "assessment-1" },
|
||||
rollback_policy: { mode: "drain_to_unavailable" }, promoted_at: "2026-08-25T12:00:00Z", created_at: "2026-08-25T12:00:00Z",
|
||||
};
|
||||
|
||||
const budget: SchedulerBudget = {
|
||||
compute_node_id: "node-1", node_name: "GPU Node", accelerator_id: "gpu-1",
|
||||
accelerator_uuid: "GPU-test", accelerator_name: "NVIDIA RTX 4080 SUPER",
|
||||
total_vram_bytes: 17_171_480_576, observed_used_vram_bytes: 2_200_000_000,
|
||||
external_vram_bytes: 700_000_000, resident_vram_bytes: 1_500_000_000,
|
||||
leased_vram_bytes: 0, safety_reserve_bytes: 1_073_741_824,
|
||||
schedulable_free_vram_bytes: 13_897_738_752, pressure: false, observed_at: "2026-08-25T12:00:00Z",
|
||||
pressure_state: "NORMAL", attribution_confidence: "KNOWN", invariant_delta_bytes: 0,
|
||||
policy_revision: "m10-v1",
|
||||
};
|
||||
|
||||
function mockServing(options: { deployments?: CapabilityDeployment[]; invokeError?: boolean; calls?: string[] } = {}): void {
|
||||
vi.stubGlobal("fetch", vi.fn((input: RequestInfo | URL, init?: RequestInit) => {
|
||||
const url = String(input);
|
||||
let payload: unknown = [];
|
||||
let status = 200;
|
||||
if (url.endsWith("/api/v1/capability-deployments")) payload = options.deployments ?? [deployment];
|
||||
if (url.endsWith("/api/v1/capability-estate")) payload = [
|
||||
{ capability: "rag.embedding", version: 1, category: "RAG", purpose: "Local retrieval", declared_stability: "stable", operational_state: options.deployments?.length === 0 ? "not_deployed" : "stable", current_deployment_id: "deployment-1", model: "Nomic", revision: "a".repeat(40), runtime: "sentence_transformers", node: "node-1", resource_class: "MEDIUM", measured_required_vram_bytes: 1_600_000_000, consumers: ["examplerag"], privacy_class: "confidential", evaluation_type: "retrieval", evaluation_state: "completed" },
|
||||
{ capability: "document.ocr", version: 1, category: "DOCUMENT", purpose: "Local OCR", declared_stability: "experimental", operational_state: "not_deployed", resource_class: "MEDIUM", consumers: ["examplerag"], privacy_class: "restricted", evaluation_type: "ocr", evaluation_state: "not_evaluated" },
|
||||
{ capability: "vision.embedding", version: 1, category: "VISION", purpose: "Visual retrieval", declared_stability: "experimental", operational_state: "not_deployed", resource_class: "HEAVY", consumers: ["examplevision"], privacy_class: "confidential", evaluation_type: "visual-retrieval", evaluation_state: "not_evaluated" },
|
||||
{ capability: "speech.transcription", version: 1, category: "AUDIO", purpose: "Local transcription", declared_stability: "experimental", operational_state: "not_deployed", resource_class: "HEAVY", consumers: [], privacy_class: "restricted", evaluation_type: "asr", evaluation_state: "not_evaluated" },
|
||||
];
|
||||
if (url.endsWith("/api/v1/scheduler")) payload = [budget];
|
||||
if (url.endsWith("/api/v1/deployment-candidates")) payload = [];
|
||||
if (url.includes("/api/v1/gateway/requests")) payload = [{ request_id: "request-1", capability: "rag.embedding@1", client: "acceptance", status: "completed", cold: false, queue_time_ms: 1, total_latency_ms: 20, created_at: "2026-08-25T12:00:00Z" }];
|
||||
if (url.endsWith("/api/v1/admin/production-approvals")) payload = [];
|
||||
if (url.includes("/api/v1/admin/scheduler/placements")) payload = [];
|
||||
if (url.endsWith("/api/v1/admin/scheduler/policy")) payload = { revision: "m10-v1", active: true, configuration: { lab_paused: false }, created_at: "2026-08-25T12:00:00Z" };
|
||||
if (url.includes("/unload")) { options.calls?.push(url); payload = { unloaded: true }; }
|
||||
if (url.endsWith("/api/v1/admin/service-clients") && init?.method !== "POST") payload = [];
|
||||
if (url.endsWith("/api/v1/admin/service-clients") && init?.method === "POST") payload = { id: "client-1", name: "modelforge-m5-acceptance", status: "active", allowed_capabilities: ["rag.embedding@1"], requests_per_minute: 60, max_concurrent_requests: 1, credential_prefix: "mf_test", created_at: "2026-08-25T12:00:00Z", credential: "mf_test_one_time_secret" };
|
||||
if (url.endsWith("/api/v1/capabilities/rag.embedding@1/invoke")) {
|
||||
if (options.invokeError) { status = 429; payload = { error: { message: "rate limit exceeded", details: { failure_code: "RATE_LIMITED" } } }; }
|
||||
else payload = { capability: "rag.embedding@1", dimension: 1024, normalized: true, embedding_space_id: "space-1", data: [Array(1024).fill(0.03125)], request_id: "request-2", execution: { cold: false, node: "GPU Node", residency: "warm", load_count: 1, timings: { gateway_ms: 1, queue_ms: 0, load_ms: 0, inference_ms: 12, total_ms: 14 } }, usage: { input_count: 1, input_tokens: 4 } };
|
||||
}
|
||||
return Promise.resolve(new Response(JSON.stringify(payload), { status }));
|
||||
}));
|
||||
}
|
||||
|
||||
afterEach(() => { cleanup(); vi.unstubAllGlobals(); });
|
||||
|
||||
describe("M5 capability operations UI", () => {
|
||||
it("shows the healthy stable capability and warm residency", async () => {
|
||||
mockServing(); render(<CapabilityWorkspace onError={vi.fn()} />);
|
||||
expect(await screen.findByText("rag.embedding@1")).toBeInTheDocument();
|
||||
expect(screen.getByText("healthy")).toBeInTheDocument();
|
||||
expect(screen.getAllByText("warm").length).toBeGreaterThan(0);
|
||||
expect(screen.getByText("requires_reindex")).toBeInTheDocument();
|
||||
});
|
||||
|
||||
it("shows unavailable honestly when promotion has not happened", async () => {
|
||||
mockServing({ deployments: [] }); render(<CapabilityWorkspace onError={vi.fn()} />);
|
||||
expect(await screen.findByText("unavailable")).toBeInTheDocument();
|
||||
expect(screen.getByText(/No stable deployment exists/)).toBeInTheDocument();
|
||||
});
|
||||
|
||||
it("renders real scheduler budget categories including external usage", async () => {
|
||||
mockServing(); render(<CapabilityWorkspace onError={vi.fn()} />);
|
||||
fireEvent.click(await screen.findByRole("tab", { name: "scheduler" }));
|
||||
expect(screen.getByText(/GPU SCHEDULING COCKPIT · GPU Node/)).toBeInTheDocument();
|
||||
expect(screen.getByText("External / unmanaged")).toBeInTheDocument();
|
||||
expect(screen.getByText("Safety reserve")).toBeInTheDocument();
|
||||
expect(screen.getByText("Schedulable free")).toBeInTheDocument();
|
||||
});
|
||||
|
||||
it("shows a newly created credential exactly in the one-time result", async () => {
|
||||
mockServing(); render(<CapabilityWorkspace onError={vi.fn()} />);
|
||||
fireEvent.click(await screen.findByRole("tab", { name: "clients" }));
|
||||
fireEvent.change(screen.getByLabelText("Operator API key"), { target: { value: "admin-test" } });
|
||||
fireEvent.click(screen.getByRole("button", { name: "Unlock operational view" }));
|
||||
fireEvent.click(await screen.findByRole("button", { name: /Create one-time credential/ }));
|
||||
expect(await screen.findByText("mf_test_one_time_secret")).toBeInTheDocument();
|
||||
expect(screen.getByText(/Only its hash is stored/)).toBeInTheDocument();
|
||||
});
|
||||
|
||||
it("invokes by capability without asking for a concrete model name", async () => {
|
||||
mockServing(); render(<CapabilityWorkspace onError={vi.fn()} />);
|
||||
fireEvent.click(await screen.findByRole("tab", { name: "playground" }));
|
||||
expect(screen.queryByLabelText(/model name/i)).not.toBeInTheDocument();
|
||||
expect(screen.getByRole("option", { name: "document.ocr@1" })).toBeInTheDocument();
|
||||
expect(screen.getByRole("option", { name: "vision.embedding@1" })).toBeInTheDocument();
|
||||
expect(screen.getByRole("option", { name: "speech.transcription@1" })).toBeInTheDocument();
|
||||
fireEvent.change(screen.getByLabelText("Service credential"), { target: { value: "client-secret" } });
|
||||
fireEvent.click(screen.getByRole("button", { name: "Invoke capability" }));
|
||||
expect(await screen.findByText("1 × 1024 finite values")).toBeInTheDocument();
|
||||
});
|
||||
|
||||
it("surfaces typed gateway failures", async () => {
|
||||
mockServing({ invokeError: true }); render(<CapabilityWorkspace onError={vi.fn()} />);
|
||||
fireEvent.click(await screen.findByRole("tab", { name: "playground" }));
|
||||
fireEvent.change(screen.getByLabelText("Service credential"), { target: { value: "client-secret" } });
|
||||
fireEvent.click(screen.getByRole("button", { name: "Invoke capability" }));
|
||||
expect(await screen.findByText("rate limit exceeded (RATE_LIMITED)")).toBeInTheDocument();
|
||||
});
|
||||
});
|
||||
|
||||
describe("M18 closure · capability tabs and guarded residency unload", () => {
|
||||
async function openScheduler(calls: string[]) {
|
||||
mockServing({ calls });
|
||||
render(<CapabilityWorkspace onError={vi.fn()} />);
|
||||
fireEvent.click(await screen.findByRole("tab", { name: "scheduler" }));
|
||||
fireEvent.change(screen.getByLabelText("Operator API key"), { target: { value: "admin-test" } });
|
||||
fireEvent.click(screen.getByRole("button", { name: "Unlock operational view" }));
|
||||
return await screen.findByRole("button", { name: "Unload idle" });
|
||||
}
|
||||
|
||||
it("pairs each capability tab with the panel it controls", async () => {
|
||||
mockServing(); render(<CapabilityWorkspace onError={vi.fn()} />);
|
||||
await screen.findByRole("tablist", { name: "Capability operations" });
|
||||
const selected = screen.getByRole("tab", { selected: true });
|
||||
const panel = screen.getByRole("tabpanel");
|
||||
expect(selected.getAttribute("aria-controls")).toBe(panel.getAttribute("id"));
|
||||
expect(panel.getAttribute("aria-labelledby")).toBe(selected.getAttribute("id"));
|
||||
const ids = screen.getAllByRole("tab").map((tab) => tab.getAttribute("id"));
|
||||
expect(new Set(ids).size).toBe(ids.length);
|
||||
});
|
||||
|
||||
it("guards residency unload with the console safety dialog instead of a native confirm", async () => {
|
||||
const confirmSpy = vi.spyOn(window, "confirm");
|
||||
const calls: string[] = [];
|
||||
fireEvent.click(await openScheduler(calls));
|
||||
const dialog = await screen.findByRole("alertdialog", { name: "Unload idle residency?" });
|
||||
expect(dialog).toBeInTheDocument();
|
||||
// The native dialog is gone for good: it cannot carry this impact evidence.
|
||||
expect(confirmSpy).not.toHaveBeenCalled();
|
||||
expect(screen.getByText("rag.embedding@1")).toBeInTheDocument();
|
||||
expect(screen.getByText("Resident VRAM released")).toBeInTheDocument();
|
||||
expect(calls).toHaveLength(0);
|
||||
confirmSpy.mockRestore();
|
||||
});
|
||||
|
||||
it("performs no mutation when the guarded unload is cancelled", async () => {
|
||||
const calls: string[] = [];
|
||||
fireEvent.click(await openScheduler(calls));
|
||||
await screen.findByRole("alertdialog", { name: "Unload idle residency?" });
|
||||
fireEvent.click(screen.getByRole("button", { name: "Cancel" }));
|
||||
await waitFor(() => expect(screen.queryByRole("alertdialog")).not.toBeInTheDocument());
|
||||
expect(calls).toHaveLength(0);
|
||||
});
|
||||
|
||||
it("requires acknowledgement and then unloads exactly once", async () => {
|
||||
const calls: string[] = [];
|
||||
fireEvent.click(await openScheduler(calls));
|
||||
await screen.findByRole("alertdialog", { name: "Unload idle residency?" });
|
||||
const confirm = screen.getByRole("button", { name: "Unload residency" });
|
||||
expect(confirm).toBeDisabled();
|
||||
fireEvent.click(screen.getByRole("checkbox", { name: /cold-load cost/ }));
|
||||
expect(confirm).toBeEnabled();
|
||||
fireEvent.click(confirm);
|
||||
fireEvent.click(confirm);
|
||||
expect(await screen.findByRole("status")).toHaveTextContent(/was unloaded/);
|
||||
// Exactly one mutation, even though confirm was clicked twice.
|
||||
expect(calls).toHaveLength(1);
|
||||
});
|
||||
|
||||
it("dismisses the guarded unload with Escape", async () => {
|
||||
const calls: string[] = [];
|
||||
fireEvent.click(await openScheduler(calls));
|
||||
await screen.findByRole("alertdialog", { name: "Unload idle residency?" });
|
||||
fireEvent.keyDown(window, { key: "Escape" });
|
||||
await waitFor(() => expect(screen.queryByRole("alertdialog")).not.toBeInTheDocument());
|
||||
expect(calls).toHaveLength(0);
|
||||
});
|
||||
});
|
||||
Reference in new issue
Block a user