diff --git a/studio/backend/auth/authentication.py b/studio/backend/auth/authentication.py
index b39f915764..409dee29a9 100644
--- a/studio/backend/auth/authentication.py
+++ b/studio/backend/auth/authentication.py
@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import os
 import secrets
 from datetime import datetime, timedelta, timezone
 from typing import Optional, Tuple
@@ -21,7 +22,28 @@ ALGORITHM = "HS256"
 ACCESS_TOKEN_EXPIRE_MINUTES = 60
 REFRESH_TOKEN_EXPIRE_DAYS = 7
 
-security = HTTPBearer()  # Reads Authorization: Bearer <token>
+security = HTTPBearer(auto_error = False)  # Reads Authorization: Bearer <token>
+
+
+def is_auth_disabled() -> bool:
+    """
+    Return True when auth should be bypassed.
+
+    For the HF Spaces branch, auth is disabled by default so users can access
+    Studio without login/signup. You can explicitly re-enable auth with:
+      UNSLOTH_STUDIO_ENABLE_AUTH=1
+    """
+    raw_enable = os.getenv("UNSLOTH_STUDIO_ENABLE_AUTH", "").strip().lower()
+    if raw_enable in {"1", "true", "yes", "on"}:
+        return False
+
+    raw = os.getenv("UNSLOTH_STUDIO_DISABLE_AUTH", "").strip().lower()
+    if raw in {"1", "true", "yes", "on"}:
+        return True
+    if raw in {"0", "false", "no", "off"}:
+        return False
+
+    return True
 
 
 def _get_secret_for_subject(subject: str) -> str:
@@ -103,7 +125,7 @@ def reload_secret() -> None:
 
 
 async def get_current_subject(
-    credentials: HTTPAuthorizationCredentials = Depends(security),
+    credentials: Optional[HTTPAuthorizationCredentials] = Depends(security),
 ) -> str:
     """Validate JWT and require the password-change flow to be completed."""
     return await _get_current_subject(
@@ -113,7 +135,7 @@ async def get_current_subject(
 
 
 async def get_current_subject_allow_password_change(
-    credentials: HTTPAuthorizationCredentials = Depends(security),
+    credentials: Optional[HTTPAuthorizationCredentials] = Depends(security),
 ) -> str:
     """Validate JWT but allow access to the password-change endpoint."""
     return await _get_current_subject(
@@ -123,7 +145,7 @@ async def get_current_subject_allow_password_change(
 
 
 async def _get_current_subject(
-    credentials: HTTPAuthorizationCredentials,
+    credentials: Optional[HTTPAuthorizationCredentials],
     *,
     allow_password_change: bool,
 ) -> str:
@@ -136,6 +158,15 @@ async def _get_current_subject(
         async def secure_endpoint(current_subject: str = Depends(get_current_subject)):
             ...
     """
+    if is_auth_disabled():
+        return "hf-space-user"
+
+    if credentials is None:
+        raise HTTPException(
+            status_code = status.HTTP_401_UNAUTHORIZED,
+            detail = "Missing bearer token",
+        )
+
     token = credentials.credentials
     subject = _decode_subject_without_verification(token)
     if subject is None:
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 8054bd2a19..1874deca66 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -1010,8 +1010,8 @@ class LlamaCppBackend:
             self._is_vision = is_vision
             self._model_identifier = model_identifier
 
-            # Wait for llama-server to become healthy
-            if not self._wait_for_health(timeout = 120.0):
+            # Wait for llama-server to become healthy (no timeout — let it take as long as needed)
+            if not self._wait_for_health():
                 self._kill_process()
                 raise RuntimeError(
                     "llama-server failed to start. "
@@ -1129,39 +1129,49 @@ class LlamaCppBackend:
         """atexit handler to ensure llama-server is terminated."""
         self._kill_process()
 
-    def _wait_for_health(self, timeout: float = 120.0, interval: float = 0.5) -> bool:
+    def _wait_for_health(self, interval: float = 0.5) -> bool:
         """
         Poll llama-server's /health endpoint until it responds 200.
 
+        No timeout — waits indefinitely so large models on slow hardware
+        (e.g. HF Spaces) have time to load. Logs elapsed time every 30s.
         Also monitors subprocess for early exit/crash.
         """
-        deadline = time.monotonic() + timeout
+        start = time.monotonic()
+        last_log = start
         url = f"http://127.0.0.1:{self._port}/health"
 
-        while time.monotonic() < deadline:
+        while True:
             # Check if process crashed
             if self._process.poll() is not None:
                 # Give the drain thread a moment to collect final output
                 if self._stdout_thread is not None:
                     self._stdout_thread.join(timeout = 2)
                 output = "\n".join(self._stdout_lines[-50:])
+                elapsed = time.monotonic() - start
                 logger.error(
-                    f"llama-server exited with code {self._process.returncode}. "
-                    f"Output: {output[:2000]}"
+                    f"llama-server exited with code {self._process.returncode} "
+                    f"after {elapsed:.1f}s. Output: {output[:2000]}"
                 )
                 return False
 
             try:
                 resp = httpx.get(url, timeout = 2.0)
                 if resp.status_code == 200:
+                    elapsed = time.monotonic() - start
+                    logger.info(f"llama-server healthy after {elapsed:.1f}s")
                     return True
             except (httpx.ConnectError, httpx.TimeoutException):
                 pass
 
-            time.sleep(interval)
+            # Periodic progress logging
+            now = time.monotonic()
+            if now - last_log >= 30.0:
+                elapsed = now - start
+                logger.info(f"Waiting for llama-server health check... {elapsed:.0f}s elapsed")
+                last_log = now
 
-        logger.error(f"llama-server health check timed out after {timeout}s")
-        return False
+            time.sleep(interval)
 
     # ── Message building (OpenAI format) ──────────────────────────
 
diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index 6ff7fd2cbf..fd9bfd1cc9 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -262,28 +262,56 @@ class InferenceOrchestrator:
         except (EOFError, OSError, ValueError):
             return None
 
-    def _wait_response(self, expected_type: str, timeout: float = 120.0) -> dict:
+    def _wait_response(self, expected_type: str, timeout: Optional[float] = None) -> dict:
         """Block until a response of the expected type arrives.
 
         Also handles 'status' and 'error' events during the wait.
         Returns the matching response dict.
-        Raises RuntimeError on timeout or subprocess crash.
+        Raises RuntimeError on subprocess crash.
+        If timeout is None, waits indefinitely (logs progress every 30s).
         """
-        deadline = time.monotonic() + timeout
+        start = time.monotonic()
+        last_log = start
 
-        while time.monotonic() < deadline:
-            remaining = max(0.1, deadline - time.monotonic())
-            resp = self._read_resp(timeout = min(remaining, 1.0))
+        while True:
+            if timeout is not None:
+                remaining = timeout - (time.monotonic() - start)
+                if remaining <= 0:
+                    raise RuntimeError(
+                        f"Timeout waiting for '{expected_type}' response after {timeout}s"
+                    )
+                poll_timeout = min(remaining, 1.0)
+            else:
+                poll_timeout = 1.0
+
+            resp = self._read_resp(timeout = poll_timeout)
 
             if resp is None:
                 # Check subprocess health
                 if not self._ensure_subprocess_alive():
-                    raise RuntimeError("Inference subprocess crashed during wait")
+                    elapsed = time.monotonic() - start
+                    raise RuntimeError(
+                        f"Inference subprocess crashed during wait after {elapsed:.1f}s"
+                    )
+                # Periodic progress logging
+                now = time.monotonic()
+                if now - last_log >= 30.0:
+                    elapsed = now - start
+                    logger.info(
+                        "Waiting for '%s' response... %.0fs elapsed",
+                        expected_type,
+                        elapsed,
+                    )
+                    last_log = now
                 continue
 
             rtype = resp.get("type", "")
 
             if rtype == expected_type:
+                elapsed = time.monotonic() - start
+                logger.info(
+                    "Received '%s' response after %.1fs", expected_type, elapsed
+                )
                 return resp
 
             if rtype == "error":
@@ -301,10 +329,6 @@ class InferenceOrchestrator:
                 expected_type,
             )
 
-        raise RuntimeError(
-            f"Timeout waiting for '{expected_type}' response after {timeout}s"
-        )
-
     def _drain_queue(self) -> list:
         """Drain all pending responses."""
         events = []
@@ -614,7 +638,7 @@ class InferenceOrchestrator:
                 needed_major,
             )
             self._spawn_subprocess(sub_config)
-            resp = self._wait_response("loaded", timeout = 180)
+            resp = self._wait_response("loaded")
 
             # Update local state from response
             if resp.get("success"):
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 2eb46f3217..432377a9b6 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -156,6 +156,7 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
                 except Exception as e:
                     logger.warning("Could not read adapter_config.json: %s", e)
 
+        load_start = time.monotonic()
         success = backend.load_model(
             config = mc,
             max_seq_length = config.get("max_seq_length", 2048),
@@ -163,6 +164,8 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
             hf_token = hf_token,
             trust_remote_code = config.get("trust_remote_code", False),
         )
+        load_elapsed = time.monotonic() - load_start
+        logger.info("Inference model load took %.1fs (success=%s)", load_elapsed, success)
 
         if success:
             # Build model_info for the parent to mirror
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index ccd805b7ac..31c528fbad 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -415,6 +415,7 @@ def run_training_process(
 
         # ── 4c. Load training model (uses VRAM — dataset already formatted) ──
         _send_status(event_queue, "Loading model...")
+        load_start = time.monotonic()
         success = trainer.load_model(
             model_name = model_name,
             max_seq_length = config["max_seq_length"],
@@ -425,6 +426,8 @@ def run_training_process(
             is_dataset_audio = config.get("is_dataset_audio", False),
             trust_remote_code = config.get("trust_remote_code", False),
         )
+        load_elapsed = time.monotonic() - load_start
+        logger.info("Training model load took %.1fs (success=%s)", load_elapsed, success)
         if not success or trainer.should_stop:
             if trainer.should_stop:
                 event_queue.put(
diff --git a/studio/backend/models/auth.py b/studio/backend/models/auth.py
index 73d21130ae..edba6ac250 100644
--- a/studio/backend/models/auth.py
+++ b/studio/backend/models/auth.py
@@ -34,6 +34,10 @@ class AuthStatusResponse(BaseModel):
         ...,
         description = "True if the seeded admin must still change the default password",
     )
+    auth_disabled: bool = Field(
+        default = False,
+        description = "True when auth is bypassed (HF Spaces mode).",
+    )
 
 
 class ChangePasswordRequest(BaseModel):
diff --git a/studio/backend/routes/auth.py b/studio/backend/routes/auth.py
index db37ed837d..92b6df08c6 100644
--- a/studio/backend/routes/auth.py
+++ b/studio/backend/routes/auth.py
@@ -20,6 +20,7 @@ from auth.authentication import (
     create_refresh_token,
     get_current_subject,
     get_current_subject_allow_password_change,
+    is_auth_disabled,
     refresh_access_token,
 )
 
@@ -34,6 +35,14 @@ async def auth_status() -> AuthStatusResponse:
     - initialized = False -> frontend should wait for the seeded admin bootstrap.
     - initialized = True  -> frontend should show login or force the first password change.
     """
+    if is_auth_disabled():
+        return AuthStatusResponse(
+            initialized = True,
+            default_username = storage.DEFAULT_ADMIN_USERNAME,
+            requires_password_change = False,
+            auth_disabled = True,
+        )
+
     return AuthStatusResponse(
         initialized = storage.is_initialized(),
         default_username = storage.DEFAULT_ADMIN_USERNAME,
@@ -42,6 +51,7 @@ async def auth_status() -> AuthStatusResponse:
         )
         if storage.is_initialized()
         else True,
+        auth_disabled = False,
     )
 
 
diff --git a/studio/frontend/src/app/auth-guards.ts b/studio/frontend/src/app/auth-guards.ts
index 1dcdfcb143..27453e2bfe 100644
--- a/studio/frontend/src/app/auth-guards.ts
+++ b/studio/frontend/src/app/auth-guards.ts
@@ -10,6 +10,22 @@ import {
   refreshSession,
 } from "@/features/auth";
 
+type AuthStatus = {
+  initialized: boolean;
+  requires_password_change: boolean;
+  auth_disabled?: boolean;
+};
+
+async function getAuthStatus(): Promise<AuthStatus | null> {
+  try {
+    const res = await fetch("/api/auth/status");
+    if (!res.ok) return null;
+    return (await res.json()) as AuthStatus;
+  } catch {
+    return null;
+  }
+}
+
 async function hasActiveSession(): Promise<boolean> {
   if (hasAuthToken()) return true;
   if (!hasRefreshToken()) return false;
@@ -17,28 +33,22 @@ async function hasActiveSession(): Promise<boolean> {
 }
 
 async function checkAuthInitialized(): Promise<boolean> {
-  try {
-    const res = await fetch("/api/auth/status");
-    if (!res.ok) return true; // fallback to login on error
-    const data = (await res.json()) as { initialized: boolean };
-    return data.initialized;
-  } catch {
-    return true; // fallback to login on error
-  }
+  const status = await getAuthStatus();
+  if (status?.auth_disabled) return true;
+  return status?.initialized ?? true; // fallback to login on error
 }
 
 async function checkPasswordChangeRequired(): Promise<boolean> {
-  try {
-    const res = await fetch("/api/auth/status");
-    if (!res.ok) return mustChangePassword();
-    const data = (await res.json()) as { requires_password_change: boolean };
-    return data.requires_password_change || mustChangePassword();
-  } catch {
-    return mustChangePassword();
-  }
+  const status = await getAuthStatus();
+  if (status?.auth_disabled) return false;
+  if (!status) return mustChangePassword();
+  return status.requires_password_change || mustChangePassword();
 }
 
 export async function requireAuth(): Promise<void> {
+  const status = await getAuthStatus();
+  if (status?.auth_disabled) return;
+
   if (await hasActiveSession()) {
     if (await checkPasswordChangeRequired()) {
       throw redirect({ to: "/change-password" });
@@ -52,11 +62,21 @@ export async function requireAuth(): Promise<void> {
 }
 
 export async function requireGuest(): Promise<void> {
+  const status = await getAuthStatus();
+  if (status?.auth_disabled) {
+    throw redirect({ to: getPostAuthRoute() });
+  }
+
   if (!(await hasActiveSession())) return;
   throw redirect({ to: getPostAuthRoute() });
 }
 
 export async function requirePasswordChangeFlow(): Promise<void> {
+  const status = await getAuthStatus();
+  if (status?.auth_disabled) {
+    throw redirect({ to: getPostAuthRoute() });
+  }
+
   const requiresPasswordChange = await checkPasswordChangeRequired();
 
   if (requiresPasswordChange) return;
diff --git a/studio/frontend/src/features/training/hooks/use-training-runtime-lifecycle.ts b/studio/frontend/src/features/training/hooks/use-training-runtime-lifecycle.ts
index 5965e07eaa..f904238b08 100644
--- a/studio/frontend/src/features/training/hooks/use-training-runtime-lifecycle.ts
+++ b/studio/frontend/src/features/training/hooks/use-training-runtime-lifecycle.ts
@@ -15,6 +15,15 @@ import type { TrainingRuntimeStore } from "../types/runtime";
 const STATUS_POLL_INTERVAL_MS = 3000;
 const METRICS_POLL_INTERVAL_MS = 5000;
 const STREAM_RECONNECT_DELAY_MS = 1500;
+const AUTH_STATUS_RETRY_INTERVAL_MS = 3000;
+const AUTH_STATUS_TIMEOUT_MS = 3000;
+const INITIAL_HYDRATE_TIMEOUT_MS = 4000;
+
+function wait(ms: number): Promise<void> {
+  return new Promise((resolve) => {
+    setTimeout(resolve, ms);
+  });
+}
 
 function shouldUseLiveSync(state: TrainingRuntimeStore): boolean {
   return state.isTrainingRunning || state.phase === "training";
@@ -26,9 +35,46 @@ export function useTrainingRuntimeLifecycle(): void {
     let openingStream = false;
     let streamController: AbortController | null = null;
     let reconnectTimer: ReturnType<typeof setTimeout> | null = null;
+    /** HF Spaces / auth-disabled: no JWT, but train APIs allow anonymous access. */
+    let authDisabled = false;
 
     const runtimeStore = useTrainingRuntimeStore;
 
+    const canUseTrainApi = () => hasAuthToken() || authDisabled;
+    let authProbeInFlight = false;
+    let lastAuthProbeStartedAt = 0;
+
+    const maybeRefreshAuthMode = async (force = false) => {
+      if (disposed || hasAuthToken() || authDisabled) return;
+      if (authProbeInFlight) return;
+
+      const now = Date.now();
+      if (
+        !force &&
+        now - lastAuthProbeStartedAt < AUTH_STATUS_RETRY_INTERVAL_MS
+      ) {
+        return;
+      }
+
+      authProbeInFlight = true;
+      lastAuthProbeStartedAt = now;
+      const controller = new AbortController();
+      const timeout = setTimeout(() => {
+        controller.abort();
+      }, AUTH_STATUS_TIMEOUT_MS);
+      try {
+        const res = await fetch("/api/auth/status", { signal: controller.signal });
+        if (!res.ok) return;
+        const data = (await res.json()) as { auth_disabled?: boolean };
+        authDisabled = Boolean(data.auth_disabled);
+      } catch {
+        // Keep previous mode and retry later.
+      } finally {
+        clearTimeout(timeout);
+        authProbeInFlight = false;
+      }
+    };
+
     const clearReconnect = () => {
       if (reconnectTimer) {
         clearTimeout(reconnectTimer);
@@ -46,7 +92,10 @@ export function useTrainingRuntimeLifecycle(): void {
     };
 
     const pollMetrics = async () => {
-      if (!hasAuthToken()) return;
+      if (!canUseTrainApi()) {
+        void maybeRefreshAuthMode();
+        return;
+      }
       const gen = runtimeStore.getState().resetGeneration;
       try {
         const metrics = await getTrainingMetrics();
@@ -55,14 +104,17 @@ export function useTrainingRuntimeLifecycle(): void {
         }
         runtimeStore.getState().applyMetrics(metrics);
       } catch (error) {
-        if (!isAbortError(error) && !disposed && hasAuthToken()) {
+        if (!isAbortError(error) && !disposed && canUseTrainApi()) {
           runtimeStore.getState().setSseConnected(false);
         }
       }
     };
 
     const pollStatus = async () => {
-      if (!hasAuthToken()) return;
+      if (!canUseTrainApi()) {
+        void maybeRefreshAuthMode();
+        return;
+      }
       const gen = runtimeStore.getState().resetGeneration;
       try {
         const status = await getTrainingStatus();
@@ -79,7 +131,7 @@ export function useTrainingRuntimeLifecycle(): void {
           stopStream();
         }
       } catch (error) {
-        if (!isAbortError(error) && !disposed && hasAuthToken()) {
+        if (!isAbortError(error) && !disposed && canUseTrainApi()) {
           runtimeStore.getState().setSseConnected(false);
         }
       }
@@ -153,7 +205,11 @@ export function useTrainingRuntimeLifecycle(): void {
     const hydrate = async () => {
       runtimeStore.getState().setHydrating(true);
       try {
-        await Promise.all([pollStatus(), pollMetrics()]);
+        await maybeRefreshAuthMode(true);
+        await Promise.race([
+          Promise.allSettled([pollStatus(), pollMetrics()]).then(() => undefined),
+          wait(INITIAL_HYDRATE_TIMEOUT_MS),
+        ]);
       } finally {
         if (!disposed) {
           runtimeStore.getState().setHydrating(false);
@@ -162,13 +218,14 @@ export function useTrainingRuntimeLifecycle(): void {
       }
     };
 
-    void hydrate();
+    let statusTimer: ReturnType<typeof setInterval> | null = null;
+    let metricsTimer: ReturnType<typeof setInterval> | null = null;
 
-    const statusTimer = setInterval(() => {
+    void hydrate();
+    statusTimer = setInterval(() => {
       void pollStatus();
     }, STATUS_POLL_INTERVAL_MS);
-
-    const metricsTimer = setInterval(() => {
+    metricsTimer = setInterval(() => {
       const state = runtimeStore.getState();
       if (shouldUseLiveSync(state) || state.currentStep > 0) {
         void pollMetrics();
@@ -177,8 +234,8 @@ export function useTrainingRuntimeLifecycle(): void {
 
     return () => {
       disposed = true;
-      clearInterval(statusTimer);
-      clearInterval(metricsTimer);
+      if (statusTimer) clearInterval(statusTimer);
+      if (metricsTimer) clearInterval(metricsTimer);
       stopStream();
     };
   }, []);
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index a141c64425..20011a298a 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -140,15 +140,15 @@ def _bootstrap_uv() -> bool:
     global UV_NEEDS_SYSTEM
     if not shutil.which("uv"):
         return False
-    # Probe: try a dry-run install targeting the current Python explicitly.
-    # Without --python, uv can ignore the activated venv on some platforms.
+    # Probe: try a dry-run install without --system.
+    # If uv can't find a venv it exits with code 2.
     probe = subprocess.run(
-        ["uv", "pip", "install", "--dry-run", "--python", sys.executable, "pip"],
+        ["uv", "pip", "install", "--dry-run", "pip"],
         stdout = subprocess.PIPE,
         stderr = subprocess.STDOUT,
     )
     if probe.returncode != 0:
-        # Retry with --system (some envs need it when uv can't find a venv)
+        # Retry with --system to confirm it works
         probe_sys = subprocess.run(
             ["uv", "pip", "install", "--dry-run", "--system", "pip"],
             stdout = subprocess.PIPE,
@@ -204,10 +204,6 @@ def _build_uv_cmd(args: tuple[str, ...]) -> list[str]:
     cmd = ["uv", "pip", "install"]
     if UV_NEEDS_SYSTEM:
         cmd.append("--system")
-    # Always pass --python so uv targets the correct environment.
-    # Without this, uv can ignore an activated venv and install into
-    # the system Python (observed on Colab and similar environments).
-    cmd.extend(["--python", sys.executable])
     cmd.extend(_translate_pip_args_for_uv(args))
     cmd.append("--torch-backend=auto")
     return cmd
diff --git a/studio/setup.sh b/studio/setup.sh
index 3e56aebf38..ae47f7e362 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -41,26 +41,13 @@ if [[ "$keynames" == *$'\nCOLAB_'* ]]; then
 fi
 
 # ── Detect whether frontend needs building ──
-# Skip if dist/ exists AND no tracked input is newer than dist/.
-# Checks top-level config/entry files and src/, public/ recursively.
-# This handles: PyPI installs (dist/ bundled), repeat runs (no changes),
-# and upgrades/pulls (source newer than dist/ triggers rebuild).
-_NEED_FRONTEND_BUILD=true
-if [ -d "$SCRIPT_DIR/frontend/dist" ]; then
-    # Check all top-level files (package.json, bun.lock, vite.config.ts, index.html, etc.)
-    _changed=$(find "$SCRIPT_DIR/frontend" -maxdepth 1 -type f \
-        -newer "$SCRIPT_DIR/frontend/dist" -print -quit 2>/dev/null)
-    # Check src/ and public/ recursively (|| true guards against set -e when dirs are missing)
-    if [ -z "$_changed" ]; then
-        _changed=$(find "$SCRIPT_DIR/frontend/src" "$SCRIPT_DIR/frontend/public" \
-            -type f -newer "$SCRIPT_DIR/frontend/dist" -print -quit 2>/dev/null) || true
-    fi
-    if [ -z "$_changed" ]; then
-        _NEED_FRONTEND_BUILD=false
-    fi
-fi
-if [ "$_NEED_FRONTEND_BUILD" = false ]; then
-    echo "✅ Frontend already built and up to date -- skipping Node/npm check."
+# Only skip when BOTH conditions are true:
+#   1. We're inside site-packages (PyPI / pip install, not editable)
+#   2. dist/ already exists (pre-built in the wheel)
+# Otherwise always (re)build — handles upgrades, editable installs, and
+# pip-from-source where dist/ was never built.
+if [[ "$SCRIPT_DIR" == */site-packages/* ]] && [ -d "$SCRIPT_DIR/frontend/dist" ]; then
+    echo "✅ Frontend pre-built (PyPI) — skipping Node/npm check."
 else
 NEED_NODE=true
 if command -v node &>/dev/null && command -v npm &>/dev/null; then
@@ -159,27 +146,12 @@ run_quiet "npm run build" npm run build
 
 _restore_gitignores
 trap - EXIT
-
-# Validate CSS output -- catch truncated Tailwind builds
-_MAX_CSS=$(find "$SCRIPT_DIR/frontend/dist/assets" -name '*.css' -exec wc -c {} + 2>/dev/null | sort -n | tail -1 | awk '{print $1}')
-if [ -z "$_MAX_CSS" ]; then
-    echo "⚠️  WARNING: No CSS files were emitted. The frontend build may have failed."
-elif [ "$_MAX_CSS" -lt 100000 ]; then
-    echo "⚠️  WARNING: Largest CSS file is only $((_MAX_CSS / 1024))KB (expected >100KB)."
-    echo "   Tailwind may not have scanned all source files. Check for .gitignore interference."
-fi
-
+cd "$SCRIPT_DIR/backend/core/data_recipe/oxc-validator"
+run_quiet "npm install (oxc validator runtime)" npm install
 cd "$SCRIPT_DIR"
 echo "✅ Frontend built to frontend/dist"
 
-fi  # end frontend build check
-
-# ── oxc-validator runtime (needs npm -- skip if not available) ──
-if [ -d "$SCRIPT_DIR/backend/core/data_recipe/oxc-validator" ] && command -v npm &>/dev/null; then
-    cd "$SCRIPT_DIR/backend/core/data_recipe/oxc-validator"
-    run_quiet "npm install (oxc validator runtime)" npm install
-    cd "$SCRIPT_DIR"
-fi
+fi  # end frontend dist check
 
 # ── 6. Python venv + deps ──
 
@@ -251,261 +223,188 @@ install_python_stack() {
     python "$SCRIPT_DIR/install_python_stack.py"
 }
 
-# Create venv under ~/.unsloth/studio/ (shared location, not in repo).
-# All platforms (including Colab) use the same isolated venv so that
-# studio dependencies are never installed into the system Python.
-STUDIO_HOME="$HOME/.unsloth/studio"
-VENV_DIR="$STUDIO_HOME/.venv"
-VENV_T5_DIR="$STUDIO_HOME/.venv_t5"
-mkdir -p "$STUDIO_HOME"
-
-# Clean up legacy in-repo venvs if they exist
-[ -d "$REPO_ROOT/.venv" ] && rm -rf "$REPO_ROOT/.venv"
-[ -d "$REPO_ROOT/.venv_overlay" ] && rm -rf "$REPO_ROOT/.venv_overlay"
-[ -d "$REPO_ROOT/.venv_t5" ] && rm -rf "$REPO_ROOT/.venv_t5"
-
-rm -rf "$VENV_DIR"
-rm -rf "$VENV_T5_DIR"
-# Try creating venv with pip; fall back to --without-pip + bootstrap
-# (some environments like Colab have broken ensurepip)
-if ! "$BEST_PY" -m venv "$VENV_DIR" 2>/dev/null; then
-    "$BEST_PY" -m venv --without-pip "$VENV_DIR"
-    source "$VENV_DIR/bin/activate"
-    curl -sS https://bootstrap.pypa.io/get-pip.py | python > /dev/null
+if [ "$IS_COLAB" = true ]; then
+    # Colab: install packages directly without venv
+    install_python_stack
 else
+    # Local: create venv under studio home (shared location, not in repo)
+    # Configurable via UNSLOTH_STUDIO_HOME; defaults to ~/.unsloth/studio
+    STUDIO_HOME="${UNSLOTH_STUDIO_HOME:-$HOME/.unsloth/studio}"
+    echo "   Studio home: $STUDIO_HOME"
+    # Persist for future `unsloth studio` runs (survives shell restarts)
+    mkdir -p "$HOME/.unsloth"
+    echo "$STUDIO_HOME" > "$HOME/.unsloth/studio_home"
+    VENV_DIR="$STUDIO_HOME/.venv"
+    VENV_T5_DIR="$STUDIO_HOME/.venv_t5"
+    mkdir -p "$STUDIO_HOME"
+
+    # Clean up legacy in-repo venvs if they exist
+    [ -d "$REPO_ROOT/.venv" ] && rm -rf "$REPO_ROOT/.venv"
+    [ -d "$REPO_ROOT/.venv_overlay" ] && rm -rf "$REPO_ROOT/.venv_overlay"
+    [ -d "$REPO_ROOT/.venv_t5" ] && rm -rf "$REPO_ROOT/.venv_t5"
+
+    rm -rf "$VENV_DIR"
+    rm -rf "$VENV_T5_DIR"
+    "$BEST_PY" -m venv "$VENV_DIR"
     source "$VENV_DIR/bin/activate"
-fi
+    cd "$SCRIPT_DIR"
+    install_python_stack
 
-# ── Ensure uv is available (much faster than pip) ──
-USE_UV=false
-if command -v uv &>/dev/null; then
-    USE_UV=true
-elif curl -LsSf https://astral.sh/uv/install.sh | sh > /dev/null 2>&1; then
-    export PATH="$HOME/.local/bin:$PATH"
-    command -v uv &>/dev/null && USE_UV=true
-fi
-
-# Helper: install a package, preferring uv with pip fallback
-fast_install() {
-    if [ "$USE_UV" = true ]; then
-        uv pip install --python "$(command -v python)" "$@" && return 0
-    fi
-    python -m pip install "$@"
-}
-
-cd "$SCRIPT_DIR"
-install_python_stack
-
-# ── 6b. Pre-install transformers 5.x into .venv_t5/ ──
-# Models like GLM-4.7-Flash need transformers>=5.3.0. Instead of pip-installing
-# at runtime (slow, ~10-15s), we pre-install into a separate directory.
-# The training subprocess just prepends .venv_t5/ to sys.path -- instant switch.
-echo ""
-echo "   Pre-installing transformers 5.x for newer model support..."
-mkdir -p "$VENV_T5_DIR"
-run_quiet "install transformers 5.x" fast_install --target "$VENV_T5_DIR" --no-deps "transformers==5.3.0"
-run_quiet "install huggingface_hub for t5" fast_install --target "$VENV_T5_DIR" --no-deps "huggingface_hub==1.7.1"
-run_quiet "install hf_xet for t5" fast_install --target "$VENV_T5_DIR" --no-deps "hf_xet==1.4.2"
-# tiktoken is needed by Qwen-family tokenizers. Install with deps since
-# regex/requests may be missing on Windows.
-run_quiet "install tiktoken for t5" fast_install --target "$VENV_T5_DIR" "tiktoken"
-echo "✅ Transformers 5.x pre-installed to $VENV_T5_DIR/"
-
-# ── 7. WSL: pre-install GGUF build dependencies ──
-# On WSL, sudo requires a password and can't be entered during GGUF export
-# (runs in a non-interactive subprocess). Install build deps here instead.
-if grep -qi microsoft /proc/version 2>/dev/null; then
+    # ── 6b. Pre-install transformers 5.x into .venv_t5/ ──
+    # Models like GLM-4.7-Flash need transformers>=5.3.0. Instead of pip-installing
+    # at runtime (slow, ~10-15s), we pre-install into a separate directory.
+    # The training subprocess just prepends .venv_t5/ to sys.path — instant switch.
     echo ""
-    echo "⚠️  WSL detected -- installing build dependencies for GGUF export..."
-    _GGUF_DEPS="pciutils build-essential cmake curl git libcurl4-openssl-dev"
+    echo "   Pre-installing transformers 5.x for newer model support..."
+    mkdir -p "$VENV_T5_DIR"
+    run_quiet "pip install transformers 5.x" pip install --target "$VENV_T5_DIR" --no-deps "transformers==5.3.0"
+    run_quiet "pip install huggingface_hub for t5" pip install --target "$VENV_T5_DIR" --no-deps "huggingface_hub==1.3.0"
+    echo "✅ Transformers 5.x pre-installed to $VENV_T5_DIR/"
 
-    # Try without sudo first (works when already root)
-    apt-get update -y >/dev/null 2>&1 || true
-    apt-get install -y $_GGUF_DEPS >/dev/null 2>&1 || true
-
-    # Check which packages are still missing
-    _STILL_MISSING=""
-    for _pkg in $_GGUF_DEPS; do
-        case "$_pkg" in
-            build-essential) command -v gcc >/dev/null 2>&1 || _STILL_MISSING="$_STILL_MISSING $_pkg" ;;
-            pciutils) command -v lspci >/dev/null 2>&1 || _STILL_MISSING="$_STILL_MISSING $_pkg" ;;
-            libcurl4-openssl-dev) dpkg -s "$_pkg" >/dev/null 2>&1 || _STILL_MISSING="$_STILL_MISSING $_pkg" ;;
-            *) command -v "$_pkg" >/dev/null 2>&1 || _STILL_MISSING="$_STILL_MISSING $_pkg" ;;
-        esac
-    done
-    _STILL_MISSING=$(echo "$_STILL_MISSING" | sed 's/^ *//')
-
-    if [ -z "$_STILL_MISSING" ]; then
+    # ── 7. WSL: pre-install GGUF build dependencies ──
+    # On WSL, sudo requires a password and can't be entered during GGUF export
+    # (runs in a non-interactive subprocess). Install build deps here instead.
+    if grep -qi microsoft /proc/version 2>/dev/null; then
+        echo ""
+        echo "⚠️  WSL detected — installing build dependencies for GGUF export..."
+        echo "   You may be prompted for your password."
+        sudo apt-get update -y
+        sudo apt-get install -y build-essential cmake curl git libcurl4-openssl-dev
         echo "✅ GGUF build dependencies installed"
-    elif command -v sudo >/dev/null 2>&1; then
-        echo ""
-        echo "   !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!"
-        echo "   WARNING: We require sudo elevated permissions to install:"
-        echo "   $_STILL_MISSING"
-        echo "   If you accept, we'll run sudo now, and it'll prompt your password."
-        echo "   !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!"
-        echo ""
-        printf "   Accept? [Y/n] "
-        if [ -r /dev/tty ]; then
-            read -r REPLY </dev/tty || REPLY="y"
-        else
-            REPLY="y"
-        fi
-        case "$REPLY" in
-            [nN]*)
-                echo ""
-                echo "   Please install these packages first, then re-run Unsloth Studio setup:"
-                echo "   sudo apt-get update -y && sudo apt-get install -y $_STILL_MISSING"
-                _SKIP_GGUF_BUILD=true
-                ;;
-            *)
-                sudo apt-get update -y
-                sudo apt-get install -y $_STILL_MISSING
-                echo "✅ GGUF build dependencies installed"
-                ;;
-        esac
-    else
-        echo "   sudo is not available on this system."
-        echo "   Please install as root, then re-run setup:"
-        echo "   apt-get install -y $_STILL_MISSING"
-        _SKIP_GGUF_BUILD=true
     fi
 fi
 
 # ── 8. Build llama.cpp binaries for GGUF inference + export ──
-# Builds at ~/.unsloth/llama.cpp — a single shared location under the user's
-# home directory. This is used by both the inference server and the GGUF
-# export pipeline (unsloth-zoo).
-#   - llama-server: for GGUF model inference
-#   - llama-quantize: for GGUF export quantization (symlinked to root for check_llama_cpp())
-UNSLOTH_HOME="$HOME/.unsloth"
-mkdir -p "$UNSLOTH_HOME"
-LLAMA_CPP_DIR="$UNSLOTH_HOME/llama.cpp"
-LLAMA_SERVER_BIN="$LLAMA_CPP_DIR/build/bin/llama-server"
-if [ "${_SKIP_GGUF_BUILD:-}" = true ]; then
-    echo ""
-    echo "Skipping llama-server build (missing dependencies)"
-    echo "   Install the missing packages and re-run setup to enable GGUF inference."
-else
-rm -rf "$LLAMA_CPP_DIR"
-{
-    # Check prerequisites
-    if ! command -v cmake &>/dev/null; then
-        echo ""
-        echo "⚠️  cmake not found — skipping llama-server build (GGUF inference won't be available)"
-        echo "   Install cmake and re-run setup.sh to enable GGUF inference."
-    elif ! command -v git &>/dev/null; then
-        echo ""
-        echo "⚠️  git not found — skipping llama-server build (GGUF inference won't be available)"
-    else
-        echo ""
-        echo "Building llama-server for GGUF inference..."
-
-        BUILD_OK=true
-        run_quiet "clone llama.cpp" git clone --depth 1 https://github.com/ggml-org/llama.cpp.git "$LLAMA_CPP_DIR" || BUILD_OK=false
-
-        if [ "$BUILD_OK" = true ]; then
-            # Skip tests/examples we don't need (faster build)
-            CMAKE_ARGS="-DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_NATIVE=ON"
-
-            # Use ccache if available (dramatically faster rebuilds)
-            if command -v ccache &>/dev/null; then
-                CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache -DCMAKE_CUDA_COMPILER_LAUNCHER=ccache"
-                echo "   Using ccache for faster compilation"
-            fi
-
-            # Detect CUDA: check nvcc on PATH, then common install locations
-            NVCC_PATH=""
-            if command -v nvcc &>/dev/null; then
-                NVCC_PATH="$(command -v nvcc)"
-            elif [ -x /usr/local/cuda/bin/nvcc ]; then
-                NVCC_PATH="/usr/local/cuda/bin/nvcc"
-                export PATH="/usr/local/cuda/bin:$PATH"
-            elif ls /usr/local/cuda-*/bin/nvcc &>/dev/null 2>&1; then
-                # Pick the newest cuda-XX.X directory
-                NVCC_PATH="$(ls -d /usr/local/cuda-*/bin/nvcc 2>/dev/null | sort -V | tail -1)"
-                export PATH="$(dirname "$NVCC_PATH"):$PATH"
-            fi
-
-            if [ -n "$NVCC_PATH" ]; then
-                echo "   Building with CUDA support (nvcc: $NVCC_PATH)..."
-                CMAKE_ARGS="$CMAKE_ARGS -DGGML_CUDA=ON"
-
-                # Detect GPU compute capability and limit CUDA architectures
-                # Without this, cmake builds for ALL default archs (very slow)
-                CUDA_ARCHS=""
-                if command -v nvidia-smi &>/dev/null; then
-                    # Read all GPUs, deduplicate (handles mixed-GPU hosts)
-                    _raw_caps=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null || true)
-                    while IFS= read -r _cap; do
-                        _cap=$(echo "$_cap" | tr -d '[:space:]')
-                        if [[ "$_cap" =~ ^([0-9]+)\.([0-9]+)$ ]]; then
-                            _arch="${BASH_REMATCH[1]}${BASH_REMATCH[2]}"
-                            # Append if not already present
-                            case ";$CUDA_ARCHS;" in
-                                *";$_arch;"*) ;;
-                                *) CUDA_ARCHS="${CUDA_ARCHS:+$CUDA_ARCHS;}$_arch" ;;
-                            esac
-                        fi
-                    done <<< "$_raw_caps"
-                fi
-
-                if [ -n "$CUDA_ARCHS" ]; then
-                    echo "   GPU compute capabilities: ${CUDA_ARCHS//;/, } -- limiting build to detected archs"
-                    CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCHS}"
-                else
-                    echo "   Could not detect GPU arch -- building for all default CUDA architectures (slower)"
-                fi
-
-                # Multi-threaded nvcc compilation (uses all CPU cores per .cu file)
-                CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_FLAGS=--threads=0"
-            elif [ -d /usr/local/cuda ] || nvidia-smi &>/dev/null; then
-                echo "   CUDA driver detected but nvcc not found — building CPU-only"
-                echo "   To enable GPU: install cuda-toolkit or add nvcc to PATH"
-            else
-                echo "   Building CPU-only (no CUDA detected)..."
-            fi
-
-            NCPU=$(nproc 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || echo 4)
-
-            # Use Ninja if available (faster parallel builds than Make)
-            CMAKE_GENERATOR_ARGS=""
-            if command -v ninja &>/dev/null; then
-                CMAKE_GENERATOR_ARGS="-G Ninja"
-            fi
-
-            run_quiet "cmake llama.cpp" cmake $CMAKE_GENERATOR_ARGS -S "$LLAMA_CPP_DIR" -B "$LLAMA_CPP_DIR/build" $CMAKE_ARGS || BUILD_OK=false
-        fi
-
-        if [ "$BUILD_OK" = true ]; then
-            run_quiet "build llama-server" cmake --build "$LLAMA_CPP_DIR/build" --config Release --target llama-server -j"$NCPU" || BUILD_OK=false
-        fi
-
-        # Also build llama-quantize (needed by unsloth-zoo's GGUF export pipeline)
-        if [ "$BUILD_OK" = true ]; then
-            run_quiet "build llama-quantize" cmake --build "$LLAMA_CPP_DIR/build" --config Release --target llama-quantize -j"$NCPU" || true
-            # Symlink to llama.cpp root — check_llama_cpp() looks for the binary there
-            QUANTIZE_BIN="$LLAMA_CPP_DIR/build/bin/llama-quantize"
-            if [ -f "$QUANTIZE_BIN" ]; then
-                ln -sf build/bin/llama-quantize "$LLAMA_CPP_DIR/llama-quantize"
-            fi
-        fi
-
-        if [ "$BUILD_OK" = true ]; then
-            if [ -f "$LLAMA_SERVER_BIN" ]; then
-                echo "✅ llama-server built at $LLAMA_SERVER_BIN"
-            else
-                echo "⚠️  llama-server binary not found after build — GGUF inference won't be available"
-            fi
-            if [ -f "$LLAMA_CPP_DIR/llama-quantize" ]; then
-                echo "✅ llama-quantize available for GGUF export"
-            fi
-        else
-            echo "⚠️  llama-server build failed — GGUF inference won't be available, but everything else works"
-        fi
-    fi
-}
-fi  # end _SKIP_GGUF_BUILD check
+# Disabled: llama.cpp build is commented out for now.
+# UNCOMMENT the block below to re-enable.
+#
+# # Builds at ~/.unsloth/llama.cpp — a single shared location under the user's
+# # home directory. This is used by both the inference server and the GGUF
+# # export pipeline (unsloth-zoo).
+# #   - llama-server: for GGUF model inference
+# #   - llama-quantize: for GGUF export quantization (symlinked to root for check_llama_cpp())
+# UNSLOTH_HOME="$HOME/.unsloth"
+# mkdir -p "$UNSLOTH_HOME"
+# LLAMA_CPP_DIR="$UNSLOTH_HOME/llama.cpp"
+# LLAMA_SERVER_BIN="$LLAMA_CPP_DIR/build/bin/llama-server"
+# rm -rf "$LLAMA_CPP_DIR"
+# {
+#     # Check prerequisites
+#     if ! command -v cmake &>/dev/null; then
+#         echo ""
+#         echo "⚠️  cmake not found — skipping llama-server build (GGUF inference won't be available)"
+#         echo "   Install cmake and re-run setup.sh to enable GGUF inference."
+#     elif ! command -v git &>/dev/null; then
+#         echo ""
+#         echo "⚠️  git not found — skipping llama-server build (GGUF inference won't be available)"
+#     else
+#         echo ""
+#         echo "Building llama-server for GGUF inference..."
+#
+#         BUILD_OK=true
+#         run_quiet "clone llama.cpp" git clone --depth 1 https://github.com/ggml-org/llama.cpp.git "$LLAMA_CPP_DIR" || BUILD_OK=false
+#
+#         if [ "$BUILD_OK" = true ]; then
+#             # Skip tests/examples we don't need (faster build)
+#             CMAKE_ARGS="-DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_NATIVE=ON"
+#
+#             # Use ccache if available (dramatically faster rebuilds)
+#             if command -v ccache &>/dev/null; then
+#                 CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache -DCMAKE_CUDA_COMPILER_LAUNCHER=ccache"
+#                 echo "   Using ccache for faster compilation"
+#             fi
+#
+#             # Detect CUDA: check nvcc on PATH, then common install locations
+#             NVCC_PATH=""
+#             if command -v nvcc &>/dev/null; then
+#                 NVCC_PATH="$(command -v nvcc)"
+#             elif [ -x /usr/local/cuda/bin/nvcc ]; then
+#                 NVCC_PATH="/usr/local/cuda/bin/nvcc"
+#                 export PATH="/usr/local/cuda/bin:$PATH"
+#             elif ls /usr/local/cuda-*/bin/nvcc &>/dev/null 2>&1; then
+#                 # Pick the newest cuda-XX.X directory
+#                 NVCC_PATH="$(ls -d /usr/local/cuda-*/bin/nvcc 2>/dev/null | sort -V | tail -1)"
+#                 export PATH="$(dirname "$NVCC_PATH"):$PATH"
+#             fi
+#
+#             if [ -n "$NVCC_PATH" ]; then
+#                 echo "   Building with CUDA support (nvcc: $NVCC_PATH)..."
+#                 CMAKE_ARGS="$CMAKE_ARGS -DGGML_CUDA=ON"
+#
+#                 # Detect GPU compute capability and limit CUDA architectures
+#                 # Without this, cmake builds for ALL default archs (very slow)
+#                 CUDA_ARCHS=""
+#                 if command -v nvidia-smi &>/dev/null; then
+#                     # Read all GPUs, deduplicate (handles mixed-GPU hosts)
+#                     _raw_caps=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null || true)
+#                     while IFS= read -r _cap; do
+#                         _cap=$(echo "$_cap" | tr -d '[:space:]')
+#                         if [[ "$_cap" =~ ^([0-9]+)\.([0-9]+)$ ]]; then
+#                             _arch="${BASH_REMATCH[1]}${BASH_REMATCH[2]}"
+#                             # Append if not already present
+#                             case ";$CUDA_ARCHS;" in
+#                                 *";$_arch;"*) ;;
+#                                 *) CUDA_ARCHS="${CUDA_ARCHS:+$CUDA_ARCHS;}$_arch" ;;
+#                             esac
+#                         fi
+#                     done <<< "$_raw_caps"
+#                 fi
+#
+#                 if [ -n "$CUDA_ARCHS" ]; then
+#                     echo "   GPU compute capabilities: ${CUDA_ARCHS//;/, } -- limiting build to detected archs"
+#                     CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCHS}"
+#                 else
+#                     echo "   Could not detect GPU arch -- building for all default CUDA architectures (slower)"
+#                 fi
+#
+#                 # Multi-threaded nvcc compilation (uses all CPU cores per .cu file)
+#                 CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_FLAGS=--threads=0"
+#             elif [ -d /usr/local/cuda ] || nvidia-smi &>/dev/null; then
+#                 echo "   CUDA driver detected but nvcc not found — building CPU-only"
+#                 echo "   To enable GPU: install cuda-toolkit or add nvcc to PATH"
+#             else
+#                 echo "   Building CPU-only (no CUDA detected)..."
+#             fi
+#
+#             NCPU=$(nproc 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || echo 4)
+#
+#             # Use Ninja if available (faster parallel builds than Make)
+#             CMAKE_GENERATOR_ARGS=""
+#             if command -v ninja &>/dev/null; then
+#                 CMAKE_GENERATOR_ARGS="-G Ninja"
+#             fi
+#
+#             run_quiet "cmake llama.cpp" cmake $CMAKE_GENERATOR_ARGS -S "$LLAMA_CPP_DIR" -B "$LLAMA_CPP_DIR/build" $CMAKE_ARGS || BUILD_OK=false
+#         fi
+#
+#         if [ "$BUILD_OK" = true ]; then
+#             run_quiet "build llama-server" cmake --build "$LLAMA_CPP_DIR/build" --config Release --target llama-server -j"$NCPU" || BUILD_OK=false
+#         fi
+#
+#         # Also build llama-quantize (needed by unsloth-zoo's GGUF export pipeline)
+#         if [ "$BUILD_OK" = true ]; then
+#             run_quiet "build llama-quantize" cmake --build "$LLAMA_CPP_DIR/build" --config Release --target llama-quantize -j"$NCPU" || true
+#             # Symlink to llama.cpp root — check_llama_cpp() looks for the binary there
+#             QUANTIZE_BIN="$LLAMA_CPP_DIR/build/bin/llama-quantize"
+#             if [ -f "$QUANTIZE_BIN" ]; then
+#                 ln -sf build/bin/llama-quantize "$LLAMA_CPP_DIR/llama-quantize"
+#             fi
+#         fi
+#
+#         if [ "$BUILD_OK" = true ]; then
+#             if [ -f "$LLAMA_SERVER_BIN" ]; then
+#                 echo "✅ llama-server built at $LLAMA_SERVER_BIN"
+#             else
+#                 echo "⚠️  llama-server binary not found after build — GGUF inference won't be available"
+#             fi
+#             if [ -f "$LLAMA_CPP_DIR/llama-quantize" ]; then
+#                 echo "✅ llama-quantize available for GGUF export"
+#             fi
+#         else
+#             echo "⚠️  llama-server build failed — GGUF inference won't be available, but everything else works"
+#         fi
+#     fi
+# }
 
 echo ""
 if [ "$IS_COLAB" = true ]; then
@@ -514,9 +413,6 @@ if [ "$IS_COLAB" = true ]; then
     echo "╠══════════════════════════════════════╣"
     echo "║ Unsloth Studio is ready to start     ║"
     echo "║ in your Colab notebook!              ║"
-    echo "║                                      ║"
-    echo "║ from colab import start              ║"
-    echo "║ start()                              ║"
     echo "╚══════════════════════════════════════╝"
 else
     echo "╔══════════════════════════════════════╗"
@@ -524,6 +420,6 @@ else
     echo "╠══════════════════════════════════════╣"
     echo "║ Launch with:                         ║"
     echo "║                                      ║"
-    echo "║ unsloth studio -H 0.0.0.0 -p 8888    ║"
+    echo "║ unsloth studio -H 0.0.0.0 -p 8000    ║"
     echo "╚══════════════════════════════════════╝"
 fi