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

* 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

Author: zoomdbz

routes/cookbook_helpers.py

@@ -561,7 +561,7 @@ 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",

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:
@@ -1528,6 +1536,10 @@ 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)}'")
@@ -1542,7 +1554,8 @@ 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
@@ -2445,8 +2458,8 @@ async def get_cookbook_state(request: Request):
             try:
                 return _state_for_client(json.loads(_cookbook_state_path.read_text(encoding="utf-8")))
             except Exception:
-                return {}
-        return {}
+                return _state_for_client({})
+        return _state_for_client({})
 
     @router.post("/api/cookbook/state")
     async def save_cookbook_state(request: Request):

static/js/cookbook.js

@@ -76,7 +76,7 @@ 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; }
@@ -213,8 +213,13 @@ 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 || '';
@@ -638,7 +643,12 @@ 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.
@@ -660,19 +670,19 @@ 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
@@ -744,11 +754,16 @@ 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}`;
@@ -2614,13 +2629,14 @@ 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' });

static/js/cookbookRunning.js

@@ -781,6 +781,7 @@ 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);
@@ -1673,7 +1674,7 @@ 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 {
@@ -1688,7 +1689,6 @@ 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