feat(hub): drop offline-mode (HF_HUB_OFFLINE) support

deploy/makermodslab-station.service

@@ -11,15 +11,14 @@
 #   sudo systemctl restart makermodslab-station
 #   sudo systemctl stop makermodslab-station   # e.g. before running `makermodslab --dev`
 #
-# The service runs the offline posture (HF_HUB_OFFLINE=1 via makermodslab-station):
-# every hardware flow works, Hub calls fail fast. For an upload/Hub session,
-# stop the service and run `.venv/bin/makermodslab --lan` in a proxied shell — or
-# uncomment the Environment lines below to run the service itself through a
-# proxy (then it is NOT offline-deterministic; see INSTALL.md "server
-# posture").
+# The service runs the LAN posture (makermodslab-station == `makermodslab --lan`):
+# it binds 0.0.0.0 so other machines on the network can reach it, and never
+# opens a browser. The Hub is expected to be reachable — login, dataset upload,
+# and cloud training all need it — so give the station outbound network. If it
+# sits behind a proxy, uncomment the Environment lines below.
 
 [Unit]
-Description=MakerMods Lab robot station (headless, LAN-served, Hub-offline)
+Description=MakerMods Lab robot station (headless, LAN-served)
 After=network-online.target
 Wants=network-online.target
 
@@ -30,7 +29,7 @@ ExecStart=/home/makermods/MakerModsLab/.venv/bin/makermodslab-station
 Restart=on-failure
 RestartSec=3
 # Serial ports: the unit user must be in dialout (usermod -aG dialout <user>).
-# Proxy for an online posture (see note above) — normally leave commented:
+# Proxy, if the station reaches the Hub through one — normally leave commented:
 #Environment=https_proxy=http://127.0.0.1:7897
 #Environment=http_proxy=http://127.0.0.1:7897
 #Environment=no_proxy=localhost,127.0.0.1,192.168.0.0/16

frontend/src/components/landing/ManageCachesDialog.tsx

@@ -7,14 +7,13 @@ import {
   DialogTitle,
 } from "@/components/ui/dialog";
 import { Button } from "@/components/ui/button";
-import { Loader2, Trash2, HardDrive, AlertTriangle } from "lucide-react";
+import { Loader2, Trash2, HardDrive } from "lucide-react";
 import { useApi } from "@/contexts/ApiContext";
 import {
   DatasetItem,
   deleteDataset,
   getDatasetInfo,
 } from "@/lib/replayApi";
-import { listRunnerHardware } from "@/lib/jobsApi";
 
 interface Props {
   open: boolean;
@@ -49,25 +48,15 @@ const ManageCachesDialog: React.FC<Props> = ({
   // Repo ids currently being cleared (per-row spinner + disabled buttons).
   const [clearing, setClearing] = useState<Set<string>>(new Set());
   const [error, setError] = useState<string | null>(null);
-  // HF_HUB_OFFLINE on the backend: a cleared cache can't be re-downloaded.
-  const [offline, setOffline] = useState(false);
 
-  // On open: reset transient state and fetch sizes + the offline signal.
+  // On open: reset transient state and fetch sizes.
   useEffect(() => {
     if (!open) return;
     setError(null);
     setClearing(new Set());
     setSizes({});
 
     let cancelled = false;
-    listRunnerHardware(baseUrl, fetchWithHeaders)
-      .then((h) => {
-        if (!cancelled) setOffline(!!h.offline);
-      })
-      .catch(() => {
-        if (!cancelled) setOffline(false);
-      });
-
     for (const d of cached) {
       getDatasetInfo(baseUrl, fetchWithHeaders, d.repo_id)
         .then((info) => {
@@ -137,18 +126,6 @@ const ManageCachesDialog: React.FC<Props> = ({
         </DialogHeader>
 
         <div className="space-y-4">
-          {offline && (
-            <div className="flex items-start gap-2 rounded-md border border-amber-500/40 bg-amber-500/10 p-2 text-xs text-amber-700 dark:text-amber-200">
-              <AlertTriangle className="mt-0.5 h-4 w-4 shrink-0" />
-              <span>
-                This backend is in offline mode (
-                <code className="text-amber-800 dark:text-amber-100">HF_HUB_OFFLINE</code>). A
-                cleared cache can&apos;t be re-downloaded from the Hub until
-                offline mode is switched off.
-              </span>
-            </div>
-          )}
-
           {cached.length === 0 ? (
             <p className="rounded-md border border-border p-3 text-sm text-muted-foreground">
               No HF datasets are cached locally.

frontend/src/components/training/TrainingConfigurator.tsx

@@ -299,8 +299,6 @@ const TrainingConfigurator: React.FC<TrainingConfiguratorProps> = ({
   const [authenticated, setAuthenticated] = useState<boolean>(false);
   const [flavors, setFlavors] = useState<RunnerFlavor[]>([]);
   const [hardwareLoading, setHardwareLoading] = useState(true);
-  // HF_HUB_OFFLINE on the backend: Hub writes (incl. dataset upload) disabled.
-  const [offline, setOffline] = useState<boolean>(false);
 
   // Whether the user has hand-toggled the AMP switch this session — once they
   // have, their choice wins over the per-policy default below, same as any
@@ -356,12 +354,10 @@ const TrainingConfigurator: React.FC<TrainingConfiguratorProps> = ({
       .then((data) => {
         setAuthenticated(data.authenticated);
         setFlavors(data.flavors);
-        setOffline(!!data.offline);
       })
       .catch(() => {
         setAuthenticated(false);
         setFlavors([]);
-        setOffline(false);
       })
       .finally(() => setHardwareLoading(false));
   }, [baseUrl, fetchWithHeaders, auth.status]);
@@ -581,35 +577,21 @@ const TrainingConfigurator: React.FC<TrainingConfiguratorProps> = ({
     config.steps <= resumeSeed.step
       ? `Total steps must be greater than the checkpoint's step (${resumeSeed.step.toLocaleString()}).`
       : null;
-  // A local-only dataset on a cloud run is uploadable — unless the backend is
-  // in offline mode, in which case uploads are impossible and Start is a hard
-  // block. (needsUpload is already gated on isCloud.)
-  const uploadBlockedOffline = needsUpload && offline;
-  // A local run continued on the cloud has to push its checkpoint to the Hub
-  // first, which offline mode makes impossible — a hard block, exactly like the
-  // dataset case above.
-  const checkpointUploadBlockedOffline = needsCheckpointUpload && offline;
   const startDisabled =
     isStarting ||
     uploading ||
     !datasetRepoId ||
     localBlocked ||
     (targetRequiresAuth && !authenticated) ||
     targetMissingFlavor ||
-    uploadBlockedOffline ||
-    checkpointUploadBlockedOffline ||
     resumeStepError != null;
   const startTooltip = localBlocked
     ? "Another local training is already running"
     : targetRequiresAuth && !authenticated
       ? "Log in to Hugging Face to use cloud compute"
       : targetMissingFlavor
         ? "Select a hardware flavor"
-        : uploadBlockedOffline
-          ? "Offline mode is on — the dataset can't be uploaded to the Hub"
-          : checkpointUploadBlockedOffline
-            ? "Offline mode is on — the checkpoint can't be uploaded to the Hub"
-            : undefined;
+        : undefined;
 
   return (
     <div className="w-full">
@@ -679,7 +661,6 @@ const TrainingConfigurator: React.FC<TrainingConfiguratorProps> = ({
           <LocalDatasetCloudNotice
             repoId={datasetRepoId}
             sizeBytes={datasetSizeBytes}
-            offline={offline}
             uploading={uploading}
             errorMessage={uploadError}
           />
@@ -694,7 +675,6 @@ const TrainingConfigurator: React.FC<TrainingConfiguratorProps> = ({
           <LocalCheckpointCloudNotice
             runName={resumeSeed.name}
             step={resumeSeed.step}
-            offline={offline}
           />
         </div>
       ) : null}

frontend/src/components/training/config/LocalCheckpointCloudNotice.tsx

@@ -1,14 +1,11 @@
 import React from "react";
-import { AlertTriangle, Lock, UploadCloud, WifiOff } from "lucide-react";
+import { AlertTriangle, Lock, UploadCloud } from "lucide-react";
 
 interface LocalCheckpointCloudNoticeProps {
   /** The run being continued, for a notice that names what moves. */
   runName: string;
   /** The checkpoint step the continuation resumes from; null ⇒ the latest. */
   step: number | null;
-  /** Backend is in HF_HUB_OFFLINE mode: nothing can be uploaded, so this
-   * continuation can't run on the cloud at all. */
-  offline: boolean;
 }
 
 /**
@@ -29,36 +26,10 @@ interface LocalCheckpointCloudNoticeProps {
 const LocalCheckpointCloudNotice: React.FC<LocalCheckpointCloudNoticeProps> = ({
   runName,
   step,
-  offline,
 }) => {
   const stepLabel =
     step != null ? `step ${step.toLocaleString()}` : "its latest checkpoint";
 
-  if (offline) {
-    return (
-      <div className="rounded-lg border border-amber-500/40 bg-amber-500/10 p-4 text-sm text-amber-700 dark:text-amber-100">
-        <div className="flex items-start gap-2">
-          <WifiOff className="w-4 h-4 mt-0.5 shrink-0 text-amber-600 dark:text-amber-300" />
-          <div>
-            <div className="font-semibold">
-              This checkpoint is only on this machine
-            </div>
-            <p className="mt-1 text-amber-700/80 dark:text-amber-200/80">
-              Hugging Face Cloud continues from the Hub, but the server is in
-              offline mode (
-              <code className="text-amber-700 dark:text-amber-100">
-                HF_HUB_OFFLINE
-              </code>
-              ), so {stepLabel} of{" "}
-              <span className="font-medium">{runName}</span> can't be uploaded.
-              Switch off offline mode, or continue this run locally.
-            </p>
-          </div>
-        </div>
-      </div>
-    );
-  }
-
   return (
     <div className="rounded-lg border border-amber-500/40 bg-amber-500/10 p-4 text-sm text-amber-700 dark:text-amber-100">
       <div className="flex items-start gap-2">

frontend/src/components/training/config/LocalDatasetCloudNotice.tsx

@@ -1,15 +1,12 @@
 import React from "react";
-import { AlertTriangle, Loader2, UploadCloud, WifiOff } from "lucide-react";
+import { AlertTriangle, Loader2, UploadCloud } from "lucide-react";
 
 interface LocalDatasetCloudNoticeProps {
   /** The local-only dataset the cloud run would train on. */
   repoId: string;
   /** Approximate on-disk size in bytes, if the info endpoint had it cheaply.
    * null when unknown (Hub-only detail, or the fetch hasn't resolved). */
   sizeBytes: number | null;
-  /** Backend is in HF_HUB_OFFLINE mode: uploads are disabled, so training
-   * can't proceed for this dataset at all. */
-  offline: boolean;
   /** True while this dataset's upload is in flight (drives the progress line). */
   uploading: boolean;
   /** Last upload error, shown in-place so the user doesn't lose it to a toast. */
@@ -33,39 +30,16 @@ const formatSize = (bytes: number): string => {
  * targeted at a dataset that exists only on this machine. HF Jobs trains from
  * the Hub, so the dataset must be uploaded first — the Start button becomes
  * "Upload & start training" and the upload is chained before the job launches
- * (see Training.tsx). In offline mode the upload is impossible, so this turns
- * into a hard block instead.
+ * (see Training.tsx).
  */
 const LocalDatasetCloudNotice: React.FC<LocalDatasetCloudNoticeProps> = ({
   repoId,
   sizeBytes,
-  offline,
   uploading,
   errorMessage,
 }) => {
   const sizeLabel = sizeBytes != null ? formatSize(sizeBytes) : null;
 
-  if (offline) {
-    return (
-      <div className="rounded-lg border border-amber-500/40 bg-amber-500/10 p-4 text-sm text-amber-700 dark:text-amber-100">
-        <div className="flex items-start gap-2">
-          <WifiOff className="w-4 h-4 mt-0.5 shrink-0 text-amber-600 dark:text-amber-300" />
-          <div>
-            <div className="font-semibold">
-              This dataset is only on this machine
-            </div>
-            <p className="mt-1 text-amber-700/80 dark:text-amber-200/80">
-              Hugging Face Cloud trains from the Hub, but the server is in
-              offline mode (<code className="text-amber-700 dark:text-amber-100">HF_HUB_OFFLINE</code>
-              ), so <span className="font-medium">{repoId}</span> can't be
-              uploaded. Switch off offline mode, or run this training locally.
-            </p>
-          </div>
-        </div>
-      </div>
-    );
-  }
-
   return (
     <div className="rounded-lg border border-amber-500/40 bg-amber-500/10 p-4 text-sm text-amber-700 dark:text-amber-100">
       <div className="flex items-start gap-2">

frontend/src/lib/jobsApi.ts

@@ -311,18 +311,12 @@ export interface RunnerHardwareResponse {
   authenticated: boolean;
   username: string | null;
   flavors: RunnerFlavor[];
-  // True when the backend is in HF_HUB_OFFLINE mode: every Hub write is
-  // disabled, so a local-only dataset can't be uploaded for a cloud run. The
-  // training page uses this to keep Start disabled + explain why. Absent on
-  // older backends → treated as online (false).
-  offline?: boolean;
 }
 
 const EMPTY_HARDWARE: RunnerHardwareResponse = {
   authenticated: false,
   username: null,
   flavors: [],
-  offline: false,
 };
 
 export async function listRunnerHardware(

frontend/src/lib/modelsApi.ts

@@ -70,8 +70,8 @@ export async function getModelInfo(
 /** POST /models/upload — push a local run's final checkpoint to the Hub as a
  * PUBLIC, MakerModsLab-tagged model repo. `id` is the local run id; `repoId` optionally
  * overrides the default namespaced repo id. MUTATES the Hub. Throws ApiError
- * (400 offline, 403 no write access, 404 no checkpoint, 502 Hub failure) with
- * the backend message in `.detail`. Returns {repo_id, url, tags}. */
+ * (403 no write access, 404 no checkpoint, 502 Hub failure) with the backend
+ * message in `.detail`. Returns {repo_id, url, tags}. */
 export async function uploadModel(
   baseUrl: string,
   fetcher: Fetcher,

frontend/src/lib/replayApi.ts

@@ -276,9 +276,9 @@ export interface HubSettings {
   tags: string[];
 }
 
-/** Read the current visibility + tags of a Hub dataset. Throws ApiError (400
- * offline, 403 no read/write access, 502 Hub failure) with the backend message
- * in `.detail`. */
+/** Read the current visibility + tags of a Hub dataset. Throws ApiError (403
+ * no read/write access, 502 Hub failure) with the backend message in
+ * `.detail`. */
 export async function getDatasetHubSettings(
   baseUrl: string,
   fetcher: Fetcher,
@@ -294,8 +294,8 @@ export async function getDatasetHubSettings(
 }
 
 /** Flip a Hub dataset's visibility (public <-> private). MUTATES the live repo.
- * Throws ApiError (400 offline, 403 no write access, 502 Hub failure) with the
- * backend message in `.detail`. */
+ * Throws ApiError (403 no write access, 502 Hub failure) with the backend
+ * message in `.detail`. */
 export async function setDatasetVisibility(
   baseUrl: string,
   fetcher: Fetcher,
@@ -313,8 +313,8 @@ export async function setDatasetVisibility(
 
 /** Replace a Hub dataset card's `tags:`. The backend re-adds the required org
  * tags, so the returned list may include tags beyond the ones passed. MUTATES
- * the live card. Throws ApiError (400 offline, 403 no write access, 502 Hub
- * failure) with the backend message in `.detail`. */
+ * the live card. Throws ApiError (403 no write access, 502 Hub failure) with
+ * the backend message in `.detail`. */
 export async function setDatasetTags(
   baseUrl: string,
   fetcher: Fetcher,