fix(cookbook): treat local Windows as Windows for serve commands (odysseus-dev#3975)

zoomdbz · samooth · commit b67f644d252f · 2026-07-13T03:03:58.000+02:00
* fix(cookbook): prefer native llama-server on local Windows

* fix(cookbook): harden local llama-server launch commands

* fix(cookbook): build serve commands for selected target
diff --git a/routes/cookbook_helpers.py b/routes/cookbook_helpers.py
@@ -574,7 +574,7 @@ def _bash_squote(v: str) -> str:
 # Allow-list of binaries permitted as the leading token of `req.cmd` for /api/model/serve.
 # Anything else is rejected before the cmd is interpolated into a tmux/PowerShell wrapper.
 _SERVE_CMD_ALLOWLIST = {
-    "vllm", "llama-server", "llama_server", "llama.cpp", "ollama",
+    "vllm", "llama-server", "llama-server.exe", "llama_server", "llama.cpp", "ollama",
     "python", "python3",
     "sglang", "lmdeploy",
     "node", "npx",
diff --git a/routes/cookbook_routes.py b/routes/cookbook_routes.py
@@ -74,6 +74,9 @@ def _mask_secret(value: str) -> str:
             return "stored"
         return f"{value[:4]}...{value[-4:]}"
 
+    def _client_host_platform() -> str:
+        return "windows" if IS_WINDOWS else ""
+
     def _decrypt_secret(value: str | None) -> str:
         if not value:
             return ""
@@ -246,11 +249,15 @@ def _state_for_client(state):
         """Return cookbook state without raw secrets for browser clients."""
         _strip_task_secrets(state)
         env = state.get("env") if isinstance(state, dict) else None
+        if isinstance(state, dict) and not isinstance(env, dict):
+            env = {}
+            state["env"] = env
         if isinstance(env, dict):
             token = _decrypt_secret(env.get("hfToken"))
             env.pop("hfToken", None)
             env["hfTokenConfigured"] = bool(token)
             env["hfTokenMasked"] = _mask_secret(token)
+            env["hostPlatform"] = _client_host_platform()
         return state
 
     def _state_for_storage(state, on_disk=None):
@@ -269,6 +276,7 @@ def _state_for_storage(state, on_disk=None):
                 env.pop("hfToken", None)
             env.pop("hfTokenMasked", None)
             env.pop("hfTokenConfigured", None)
+            env.pop("hostPlatform", None)
         return state
 
     def _load_stored_hf_token() -> str:
@@ -1486,6 +1494,10 @@ async def model_serve(request: Request, req: ServeRequest):
             # shell resolves the bundled python3/hf, mirroring the download flow.
             if not remote:
                 runner_lines.append(_local_tooling_path_export(sys.executable))
+                if local_windows:
+                    # Detached Git Bash runs do not always inherit recently edited
+                    # user PATH entries from the already-running Odysseus process.
+                    runner_lines.append('export PATH="$HOME/bin:$HOME/llama.cpp/build-cuda/bin/Release:$HOME/llama.cpp/build/bin/Release:$HOME/llama.cpp/build/bin/Debug:$HOME/llama.cpp/build/bin:$PATH"')
             runner_lines.append("export FLASHINFER_DISABLE_VERSION_CHECK=1")
             if req.hf_token:
                 runner_lines.append(f"export HF_TOKEN='{_bash_squote(req.hf_token)}'")
@@ -1500,7 +1512,8 @@ async def model_serve(request: Request, req: ServeRequest):
             runner_lines.append(_HF_TOKEN_STATUS_SNIPPET)
             handled_ollama_serve = False
             # Auto-install inference engine if missing
-            if "llama_cpp" in req.cmd or "llama-server" in req.cmd:
+            local_windows_llama_cmd = local_windows and ("llama_cpp" in req.cmd or "llama-server" in req.cmd)
+            if ("llama_cpp" in req.cmd or "llama-server" in req.cmd) and not local_windows_llama_cmd:
                 # Prefer the NATIVE llama-server binary — its minja templating
                 # renders modern GGUF chat templates that the Python bindings'
                 # Jinja2 rejects (do_tojson ensure_ascii). Build it once from
@@ -2413,8 +2426,8 @@ async def get_cookbook_state(request: Request):
                     logger.warning(f"state orphan sweep failed (non-fatal): {_sweep_e!r}")
                 return _state_for_client(state)
             except Exception:
-                return {}
-        return {}
+                return _state_for_client({})
+        return _state_for_client({})
 
     @router.post("/api/cookbook/state")
     async def save_cookbook_state(request: Request):
diff --git a/static/js/cookbook.js b/static/js/cookbook.js
@@ -74,7 +74,7 @@ function _platformIcon(platform) {
   return '';
 }
 
-export let _envState = { env: 'none', envPath: '', hfToken: '', hfTokenConfigured: false, hfTokenMasked: '', gpus: '', remoteHost: '', servers: [], modelPaths: [], platform: '', defaultServer: '' };
+export let _envState = { env: 'none', envPath: '', hfToken: '', hfTokenConfigured: false, hfTokenMasked: '', gpus: '', remoteHost: '', servers: [], modelPaths: [], platform: '', hostPlatform: '', defaultServer: '' };
 let _lastCacheHostVal = null;
 let _cookbookOpeningSpinners = [];
 export function _lastCacheHost() { return _lastCacheHostVal; }
@@ -211,8 +211,13 @@ function _getPort(hostOrTask) {
 
 /** Get platform for a given host (or task object). Returns 'windows', 'termux', 'linux', or '' */
 export function _getPlatform(hostOrTask) {
-  if (!hostOrTask) return _envState.platform || '';
-  if (typeof hostOrTask === 'object') return hostOrTask.platform || _getPlatform(hostOrTask.remoteServerKey || hostOrTask.remoteHost);
+  if (hostOrTask === 'local') return _envState.hostPlatform || '';
+  if (!hostOrTask) return _envState.remoteHost ? (_envState.platform || '') : (_envState.hostPlatform || '');
+  if (typeof hostOrTask === 'object') {
+    const taskHost = hostOrTask.remoteServerKey || hostOrTask.remoteHost || '';
+    if (!taskHost || taskHost === 'local') return _envState.hostPlatform || '';
+    return hostOrTask.platform || _getPlatform(taskHost);
+  }
   const selected = hostOrTask === _envState.remoteHost ? _selectedServer() : null;
   const srv = selected || _serverByVal(hostOrTask);
   return srv?.platform || '';
@@ -636,7 +641,12 @@ export function _buildServeCmd(f, modelName, backend) {
     // GPU list — read from gpus (button strip); fall back to gpu_id for
     // backward-compat with older saved presets that pre-date the removal.
     const gpuId = (f.gpus || f.gpu_id || '').toString().trim();
-    const py = _isWindows() ? 'python' : 'python3';
+    const _targetHost = Object.prototype.hasOwnProperty.call(f, 'host')
+      ? String(f.host || '').trim()
+      : String(_envState.remoteHost || '').trim();
+    const _isWin = _targetHost ? _isWindows(_targetHost) : _isWindows('local');
+    const _localWindows = _isWin && !_targetHost;
+    const py = _isWin ? 'python' : 'python3';
     // CPU-only serve (-ngl 0): drop the GPU-only flags, otherwise the command
     // mixes "zero GPU layers" with CUDA unified-memory + flash-attn and fails to
     // start (issue #1291). Only affects the ngl=0 path; GPU serving is unchanged.
@@ -658,19 +668,19 @@ export function _buildServeCmd(f, modelName, backend) {
     // with misleading prefixes.
     const _sb = String(_hwfitCache?.system?.backend || '').toLowerCase();
     const _hwfitHost = String(_hwfitCache?._scannedHost || '');
-    const _curHost = String(_envState.remoteHost || '');
+    const _curHost = _targetHost;
     const _isCudaTarget = (_sb === 'cuda') && (_hwfitHost === _curHost);
     const lcPrefix = (() => {
       let p = '';
-      if (f.unified_mem && !_cpuOnly && !_isWindows() && _isCudaTarget) p += `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 `;
-      // No GPU env var in CPU mode — `-ngl 0` already disables offload
+      if (f.unified_mem && !_cpuOnly && (!_isWin || _localWindows) && _isCudaTarget) p += `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 `;
+      // No GPU env var in CPU mode - `-ngl 0` already disables offload
       // so CUDA_VISIBLE_DEVICES / HIP_VISIBLE_DEVICES would be misleading
       // clutter ("why is CUDA pinned for a CPU run?").
-      if (!_isWindows() && !_cpuOnly) p += _gpuEnvPrefix(gpuId);
+      if ((!_isWin || _localWindows) && !_cpuOnly) p += _gpuEnvPrefix(gpuId);
       return p;
     })();
-    if (f.unified_mem && !_cpuOnly && _isWindows() && _isCudaTarget) cmd += `$env:GGML_CUDA_ENABLE_UNIFIED_MEMORY="1"; `;
-    if (_isWindows() && !_cpuOnly) cmd += _gpuEnvPrefix(gpuId, true);
+    if (f.unified_mem && !_cpuOnly && _isWin && !_localWindows && _isCudaTarget) cmd += `$env:GGML_CUDA_ENABLE_UNIFIED_MEMORY="1"; `;
+    if (_isWin && !_localWindows && !_cpuOnly) cmd += _gpuEnvPrefix(gpuId, true);
     const needsGgufPrelude = /^\$\(\{\s*find\s/.test(String(ggufPath || ''));
     const modelArg = needsGgufPrelude ? '"$MODEL_FILE"' : `"${ggufPath}"`;
     // Prefer native llama-server. The backend bootstrap resolves/builds the
@@ -742,11 +752,16 @@ export function _buildServeCmd(f, modelName, backend) {
       // llama-cpp-python takes the projector via --clip_model_path.
       _lcpExtra += ` --clip_model_path "${f._mmproj_path}"`;
     }
-    if (_isWindows()) {
-      const _lcpServer = `${lcPrefix}${py} -m llama_cpp.server --model ${modelArg} --host 0.0.0.0 --port ${f.port || '8080'} --n_gpu_layers ${f.ngl || '99'} --n_ctx ${f.ctx || '8192'}${_lcpExtra}`;
+    const _lcServer = `${lcPrefix}llama-server --model ${modelArg} --host 0.0.0.0 --port ${f.port || '8080'} -ngl ${f.ngl || '99'} -c ${f.ctx || '8192'}${_lcExtra}`;
+    const _lcpServer = `${lcPrefix}${py} -m llama_cpp.server --model ${modelArg} --host 0.0.0.0 --port ${f.port || '8080'} --n_gpu_layers ${f.ngl || '99'} --n_ctx ${f.ctx || '8192'}${_lcpExtra}`;
+    if (_localWindows) {
+      // Local Windows serve is launched through Git Bash, so use the native
+      // llama-server shape and let PATH resolve the CUDA Release wrapper.
+      cmd += _lcServer;
+    } else if (_isWin) {
       cmd += _lcpServer;
     } else {
-      cmd += `${lcPrefix}llama-server --model ${modelArg} --host 0.0.0.0 --port ${f.port || '8080'} -ngl ${f.ngl || '99'} -c ${f.ctx || '8192'}${_lcExtra}`;
+      cmd += _lcServer;
     }
     if (needsGgufPrelude) {
       cmd = `MODEL_FILE=${ggufPath} && { [ -n "$MODEL_FILE" ] && [ -f "$MODEL_FILE" ]; } || { echo "ERROR: No GGUF found on this host"; exit 1; } && ${cmd}`;
@@ -2615,13 +2630,14 @@ function _renderRecipes() {
     const isLocal = !s.host || s.host.toLowerCase() === 'local';
     if (isLocal) {
       s.host = '';
+      s.platform = _envState.hostPlatform || '';
       if (_localSeen) return false;
       _localSeen = true;
     }
     return true;
   });
   if (!_localSeen) {
-    _es.servers.unshift({ host: '', env: _es.env || 'none', envPath: _es.envPath || '', modelDir: '~/.cache/huggingface/hub' });
+    _es.servers.unshift({ host: '', env: _es.env || 'none', envPath: _es.envPath || '', modelDir: '~/.cache/huggingface/hub', platform: _envState.hostPlatform || '' });
   }
   if (_es.remoteHost && !_es.servers.some(s => s.host === _es.remoteHost)) {
     _es.servers.push({ host: _es.remoteHost, env: _es.env || 'none', envPath: _es.envPath || '', modelDir: '~/.cache/huggingface/hub' });
diff --git a/static/js/cookbookRunning.js b/static/js/cookbookRunning.js
@@ -801,6 +801,7 @@ function _stripStateSecrets(state) {
   const safe = { ...state };
   if (safe.env && typeof safe.env === 'object') {
     const { hfToken, ...env } = safe.env;
+    delete env.hostPlatform;
     safe.env = env;
   }
   if (Array.isArray(safe.tasks)) safe.tasks = safe.tasks.map(_redactTaskForStorage);
@@ -1694,7 +1695,7 @@ export async function _launchServeTask(shortName, repo, cmd, fields, hostOverrid
     || _envState.servers.find(s => s.host === _host) || {};
   const _serverMetaKey = _targetKey || (_hsrv && _serverKey ? _serverKey(_hsrv) : '') || (_host || 'local');
   const _serverMetaName = targetMeta?.serverName || _hsrv.name || (_host ? _host : 'Local');
-  const _hplatform = _host ? (_hsrv.platform || '') : (_envState.platform || '');
+  const _hplatform = _host ? (_hsrv.platform || '') : (_envState.hostPlatform || '');
   const _replaceTaskId = fields?._replaceTaskId || '';
   if (_replaceTaskId) {
     try {
@@ -1709,7 +1710,6 @@ export async function _launchServeTask(shortName, repo, cmd, fields, hostOverrid
       }
     } catch {}
   }
-
   // Replace any serve already targeting this same host:port — you can't run two
   // servers on one port, so re-serving (or retrying) should stop & remove the
   // old one instead of leaving a dead duplicate behind. (The retry buttons
diff --git a/static/js/cookbookServe.js b/static/js/cookbookServe.js
@@ -559,7 +559,7 @@ function _selectedServeTarget(panel) {
     env: server?.env || '',
     port: host ? (server?.port || _getPort(host) || '') : '',
     venv,
-    platform: server?.platform || _envState.platform || '',
+    platform: host ? (server?.platform || '') : (_envState.hostPlatform || ''),
     label,
   };
 }
@@ -690,6 +690,12 @@ function _selectedGgufSizeGb(model, relPath) {
   return bytes / (1024 ** 3);
 }
 
+function _projectorGgufFiles(model) {
+  return _ggufFilesForModel(model)
+    .filter(f => (f.role || '') === 'projector' || /(^|\/)mmproj[^/]*\.gguf$/i.test(f.rel_path || f.name || ''))
+    .sort((a, b) => String(a.rel_path || a.name || '').localeCompare(String(b.rel_path || b.name || '')));
+}
+
 function _ggufFileLabel(file) {
   const base = (file.name || file.rel_path || '').split('/').pop();
   const size = _formatGgufSize(file.size_bytes);
@@ -1231,6 +1237,8 @@ function _rerenderCachedModels() {
           : `This model's download isn't complete yet (${esc(m.size || 'partial')}). The serve will start but is likely to crash on a missing shard. Wait for the download to finish, or relaunch after it's done.`;
         panelHtml += `<div class="hwfit-serve-warn" style="margin:0 0 8px;padding:6px 10px;border-radius:5px;font-size:11px;background:color-mix(in srgb, var(--color-warning, #f0ad4e) 14%, transparent);border:1px solid color-mix(in srgb, var(--color-warning, #f0ad4e) 40%, transparent);color:var(--color-warning, #f0ad4e);display:flex;gap:6px;align-items:flex-start;line-height:1.4;"><span aria-hidden="true">⚠</span><span>${_warnText}</span></div>`;
       }
+      panelHtml += `<div class="hwfit-serve-preset-row">${_slotsHtml}</div>`;
+      panelHtml += `<div class="hwfit-serve-vision-warn" style="display:none;margin:0 0 8px;padding:6px 10px;border-radius:5px;font-size:11px;background:color-mix(in srgb, var(--color-warning, #f0ad4e) 14%, transparent);border:1px solid color-mix(in srgb, var(--color-warning, #f0ad4e) 40%, transparent);color:var(--color-warning, #f0ad4e);gap:6px;align-items:flex-start;line-height:1.4;"><span aria-hidden="true">⚠</span><span>Vision is enabled, but no mmproj GGUF projector was found in the cached model scan. Download an mmproj-*.gguf for this model, then refresh the cached model list before launching.</span></div>`;
       // Row 1: Engine + Server + Env
       panelHtml += `<div class="hwfit-serve-row">`;
       const backendOpts = _backendChoices.map(([v,l]) => `<option value="${v}"${defaultBackend===v?' selected':''}>${l}</option>`).join('');
@@ -1556,6 +1564,11 @@ function _rerenderCachedModels() {
           if (el.type === 'checkbox') f[el.dataset.field] = el.checked;
           else f[el.dataset.field] = el.value;
         });
+        const buildTarget = _selectedServeTarget(panel);
+        f.host = buildTarget.host || '';
+        f.platform = buildTarget.platform || '';
+        const hostField = panel.querySelector('[data-field="host"]');
+        if (hostField) hostField.value = f.host;
         const backend = f.backend || 'vllm';
         const serveModel = (f.model_path || '').trim() || (m.is_local_dir && m.path ? `${m.path}/${repo}` : repo);
         if (backend === 'llamacpp') {
@@ -1575,11 +1588,11 @@ function _rerenderCachedModels() {
             : m.is_local_dir && m.path
             ? `$({ find ${_ldir} -name '*-00001-of-*.gguf' 2>/dev/null | sort; find ${_ldir} -name '*.gguf' 2>/dev/null | sort; } | head -1)`
             : `$({ find ${dir} -name '*-00001-of-*.gguf' 2>/dev/null | sort; find ${dir} -name '*.gguf' 2>/dev/null | sort; } | head -1)`;
-          // Vision: auto-find the mmproj (CLIP/projector) file in the same dir.
-          // Resolved at runtime so the toggle just works if an mmproj-*.gguf is
-          // present (downloaded alongside the model). Empty if none → cmd omits it.
-          const _vsearchdir = (m.is_local_dir && m.path) ? _ldir : dir;
-          f._mmproj_path = `$(find ${_vsearchdir} -iname 'mmproj*.gguf' 2>/dev/null | sort | head -1)`;
+          // Vision: use the scanned projector (CLIP/mmproj) file when present.
+          // Keeping this as a printf path avoids generating a command substitution
+          // that the backend serve-command validator must reject as unsafe.
+          const selectedProjector = _projectorGgufFiles(m)[0];
+          f._mmproj_path = selectedProjector ? _selectedGgufExpr(m, repo, selectedProjector.rel_path) : '';
         }
         if (f.reasoning_parser) {
           const _rpEl2 = panel.querySelector('[data-field="reasoning_parser"]');
@@ -1595,6 +1608,10 @@ function _rerenderCachedModels() {
         }
         let cmd = _buildServeCmd(f, serveModel, backend);
         if (f.extra && f.extra.trim()) cmd += ' ' + f.extra.trim();
+        const missingVisionProjector = backend === 'llamacpp' && !!f.vision && !f._mmproj_path;
+        panel._visionMissingProjector = missingVisionProjector;
+        const _visionWarn = panel.querySelector('.hwfit-serve-vision-warn');
+        if (_visionWarn) _visionWarn.style.display = missingVisionProjector ? 'flex' : 'none';
         const _ce2 = panel.querySelector('.hwfit-serve-cmd'); _ce2.value = _formatServeCmdPreview(cmd); _ce2.style.height = 'auto'; _ce2.style.height = _ce2.scrollHeight + 'px';
         panel._cmd = cmd;
         panel._host = f.host || '';
@@ -2981,12 +2998,16 @@ function _rerenderCachedModels() {
         });
         serveState.backend = serveState.backend || (_detectBackend(m).backend) || 'vllm';
         const launchTarget = _selectedServeTarget(panel);
+        if (serveState.backend === 'llamacpp' && serveState.vision && !/(?:^|\s)(?:--mmproj|--clip_model_path)\b/.test(launchCmd)) {
+          _restoreLaunchBtn();
+          uiModule.showToast('Vision is checked, but no mmproj projector is in the launch command. Refresh cached models after downloading mmproj, or add --mmproj manually.', 8000);
+          return;
+        }
         if (serveState.backend === 'diffusers' && _remoteWindowsDiffusersUnsupported(launchTarget)) {
           _restoreLaunchBtn();
           uiModule.showToast(t('cookbookServe.diffusersWindowsUnsupported'), 9000);
           return;
         }
-
         // Pre-launch: check our own task list for a serve already running
         // on this host. Offer to stop+launch as the default action — the
         // SSH-based port probe below is more thorough but it can miss
diff --git a/tests/test_cookbook_cpu_only_serve.py b/tests/test_cookbook_cpu_only_serve.py
diff --git a/tests/test_cookbook_helpers.py b/tests/test_cookbook_helpers.py
