TheTom · ljubomirj · Apr 26, 2026
diff --git a/README.md b/README.md
@@ -31,8 +31,13 @@ source activate.sh         # sets DYLD_LIBRARY_PATH (generated by install.sh)
 ### 2. Run
 
 ```bash
+# Baseline model
 vllm-swift download mlx-community/Qwen3-4B-4bit
 vllm-swift serve ~/models/Qwen3-4B-4bit --max-model-len 4096  # increase as needed, max 40960
+
+# Qwen3.6 ConfigI (requires gated_delta kernels in mlx.metallib)
+hf download thetom-ai/Qwen3.6-27B-ConfigI-MLX
+vllm-swift serve thetom-ai/Qwen3.6-27B-ConfigI-MLX --max-model-len 4096
 ```
 
 > Homebrew users don't need `activate.sh` — `vllm-swift serve` handles everything.
@@ -243,7 +248,13 @@ git clone https://github.com/TheTom/vllm-swift.git
 cd vllm-swift
 ./scripts/install.sh       # builds Swift, installs plugin, creates activate.sh
 source activate.sh         # sets DYLD_LIBRARY_PATH
+
+# Baseline model
 vllm serve ~/models/Qwen3-4B-4bit --max-model-len 4096
+
+# Qwen3.6 ConfigI (requires gated_delta kernels in mlx.metallib)
+hf download thetom-ai/Qwen3.6-27B-ConfigI-MLX
+vllm serve thetom-ai/Qwen3.6-27B-ConfigI-MLX --max-model-len 4096
 ```
 
 ### Manual (full control)
@@ -292,6 +303,13 @@ cp swift/.build/arm64-apple-macosx/release/mlx.metallib \
    $(dirname $(echo $DYLD_LIBRARY_PATH | cut -d: -f1))/
 ```
 
+If `./scripts/install.sh` warns that `gated_delta` kernels are missing, your Xcode Metal compiler/toolchain may be incomplete or outdated:
+```bash
+xcodebuild -downloadComponent MetalToolchain
+./scripts/install.sh
+strings swift/.build/arm64-apple-macosx/release/mlx.metallib | grep gated_delta
+```
+
 ### Download a model
 
 ```bash

diff --git a/chat.html b/chat.html
@@ -0,0 +1,156 @@
+<!doctype html>
+<html lang="en">
+<head>
+  <meta charset="utf-8" />
+  <meta name="viewport" content="width=device-width, initial-scale=1" />
+  <title>vLLM Chat</title>
+  <style>
+    :root {
+      --bg: #f8fafc;
+      --card: #ffffff;
+      --border: #dbe3ec;
+      --text: #101419;
+      --muted: #5a6775;
+      --accent: #0c66d6;
+    }
+    html, body { margin: 0; padding: 0; background: var(--bg); color: var(--text); }
+    body { font-family: ui-sans-serif, system-ui, -apple-system, Segoe UI, Roboto, sans-serif; }
+    .app { max-width: 980px; margin: 16px auto; padding: 0 12px; }
+    .panel { background: var(--card); border: 1px solid var(--border); border-radius: 10px; padding: 10px; }
+    .row { display: flex; gap: 8px; margin: 8px 0; }
+    .grow { flex: 1; }
+    label { font-size: 12px; color: var(--muted); display: block; margin-bottom: 4px; }
+    input, textarea, button {
+      font: inherit;
+      border: 1px solid var(--border);
+      border-radius: 8px;
+      padding: 8px 10px;
+      box-sizing: border-box;
+    }
+    input, textarea { width: 100%; background: #fff; color: var(--text); }
+    textarea { min-height: 90px; resize: vertical; }
+    button { background: #fff; cursor: pointer; }
+    button.primary { background: var(--accent); color: #fff; border-color: var(--accent); }
+    #log {
+      height: 58vh;
+      overflow: auto;
+      white-space: pre-wrap;
+      line-height: 1.35;
+      padding: 10px;
+      border: 1px solid var(--border);
+      border-radius: 8px;
+      background: #fff;
+    }
+    .hint { font-size: 12px; color: var(--muted); margin: 6px 0 0; }
+  </style>
+</head>
+<body>
+  <div class="app">
+    <div class="panel">
+      <div class="row">
+        <div class="grow">
+          <label for="baseUrl">Base URL</label>
+          <input id="baseUrl" value="http://127.0.0.1:8000" />
+        </div>
+        <div class="grow">
+          <label for="model">Model</label>
+          <input id="model" value="thetom-ai/Qwen3.6-27B-ConfigI-MLX" />
+        </div>
+        <div class="grow">
+          <label for="apiKey">API key (optional)</label>
+          <input id="apiKey" placeholder="sk-..." />
+        </div>
+      </div>
+
+      <div id="log"></div>
+
+      <div class="row">
+        <div class="grow">
+          <label for="prompt">Message</label>
+          <textarea id="prompt" placeholder="Type message, then click Send."></textarea>
+          <div class="hint">Shortcut: Ctrl+Enter (or Cmd+Enter)</div>
+        </div>
+      </div>
+
+      <div class="row">
+        <button id="clearBtn">Clear</button>
+        <button id="sendBtn" class="primary">Send</button>
+      </div>
+    </div>
+  </div>
+
+  <script>
+    const logEl = document.getElementById("log");
+    const promptEl = document.getElementById("prompt");
+    const sendBtn = document.getElementById("sendBtn");
+    const clearBtn = document.getElementById("clearBtn");
+
+    let messages = [{ role: "system", content: "You are a helpful assistant." }];
+
+    function append(role, text) {
+      logEl.textContent += role + ": " + text + "\n\n";
+      logEl.scrollTop = logEl.scrollHeight;
+    }
+
+    async function send() {
+      const text = promptEl.value.trim();
+      if (!text) return;
+
+      const baseUrl = document.getElementById("baseUrl").value.replace(/\/+$/, "");
+      const model = document.getElementById("model").value.trim();
+      const apiKey = document.getElementById("apiKey").value.trim();
+      if (!model) return;
+
+      promptEl.value = "";
+      append("user", text);
+      messages.push({ role: "user", content: text });
+
+      sendBtn.disabled = true;
+      sendBtn.textContent = "Sending...";
+      try {
+        const headers = { "Content-Type": "application/json" };
+        if (apiKey) headers.Authorization = "Bearer " + apiKey;
+
+        const res = await fetch(baseUrl + "/v1/chat/completions", {
+          method: "POST",
+          headers,
+          body: JSON.stringify({
+            model,
+            messages,
+            temperature: 0.7
+          })
+        });
+        const data = await res.json();
+
+        if (!res.ok) {
+          append("error", JSON.stringify(data, null, 2));
+          return;
+        }
+
+        const out = data && data.choices && data.choices[0] &&
+          data.choices[0].message && data.choices[0].message.content
+          ? data.choices[0].message.content
+          : JSON.stringify(data);
+
+        append("assistant", out);
+        messages.push({ role: "assistant", content: out });
+      } catch (err) {
+        append("error", String(err));
+      } finally {
+        sendBtn.disabled = false;
+        sendBtn.textContent = "Send";
+      }
+    }
+
+    clearBtn.addEventListener("click", () => {
+      messages = [{ role: "system", content: "You are a helpful assistant." }];
+      logEl.textContent = "";
+      promptEl.focus();
+    });
+    sendBtn.addEventListener("click", send);
+    promptEl.addEventListener("keydown", (e) => {
+      if ((e.ctrlKey || e.metaKey) && e.key === "Enter") send();
+    });
+  </script>
+</body>
+</html>
diff --git a/scripts/install.sh b/scripts/install.sh
@@ -23,6 +23,21 @@ SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
 PROJECT_DIR="$(dirname "$SCRIPT_DIR")"
 SWIFT_DIR="$PROJECT_DIR/swift"
 BUILD_DIR="$SWIFT_DIR/.build/arm64-apple-macosx/$CONFIG"
+CHECKOUT_MLXLM_DIR="$SWIFT_DIR/.build/checkouts/mlx-swift-lm"
+CHECKOUT_METALLIB_SCRIPT="$CHECKOUT_MLXLM_DIR/scripts/build-metallib.sh"
+CHECKOUT_METALLIB_PATH="$CHECKOUT_MLXLM_DIR/.build/arm64-apple-macosx/$CONFIG/mlx.metallib"
+
+_metallib_has_gdn_kernels() {
+    local metallib="$1"
+    [ -f "$metallib" ] || return 1
+    # Use grep directly on binary data (-a) to avoid pipefail/SIGPIPE false negatives
+    # from `strings | grep -q` pipelines.
+    LC_ALL=C grep -aq "gated_delta_step_fused_" "$metallib"
+}
+
+_metal_compiler_available() {
+    xcrun metal -v >/dev/null 2>&1
+}
 
 echo "=== vllm-swift installer ==="
 echo "Config: $CONFIG"
@@ -57,7 +72,15 @@ echo ""
 # Build Swift bridge
 echo "Building Swift bridge ($CONFIG)..."
 cd "$SWIFT_DIR"
-swift build -c "$CONFIG" 2>&1 | tail -3
+SWIFT_BUILD_LOG="$(mktemp)"
+if ! swift build -c "$CONFIG" >"$SWIFT_BUILD_LOG" 2>&1; then
+    echo "ERROR: Swift build failed. Last 80 lines:"
+    tail -80 "$SWIFT_BUILD_LOG"
+    rm -f "$SWIFT_BUILD_LOG"
+    exit 1
+fi
+tail -3 "$SWIFT_BUILD_LOG"
+rm -f "$SWIFT_BUILD_LOG"
 
 DYLIB="$BUILD_DIR/libVLLMBridge.dylib"
 if [ ! -f "$DYLIB" ]; then
@@ -70,20 +93,49 @@ echo ""
 # Find and copy MLX metallib
 echo "Setting up MLX metallib..."
 MLX_METALLIB=""
+MLX_METALLIB_FALLBACK=""
+
+if [ -f "$CHECKOUT_METALLIB_SCRIPT" ]; then
+    echo "  Attempting metallib build from mlx-swift-lm checkout..."
+    if _metal_compiler_available; then
+        if bash "$CHECKOUT_METALLIB_SCRIPT" "$CONFIG"; then
+            if [ -f "$CHECKOUT_METALLIB_PATH" ]; then
+                cp "$CHECKOUT_METALLIB_PATH" "$BUILD_DIR/mlx.metallib"
+                echo "  Built and copied checkout metallib: $CHECKOUT_METALLIB_PATH"
+            fi
+        else
+            echo "  WARNING: Checkout metallib build failed. Will fall back to existing metallib candidates."
+        fi
+    else
+        echo "  WARNING: Metal compiler is not runnable in this Xcode setup."
+        echo "           Install/update Metal Toolchain, then rerun install:"
+        echo "           xcodebuild -downloadComponent MetalToolchain"
+    fi
+fi
 
 # Check common locations for the metallib
 for candidate in \
     "$BUILD_DIR/mlx.metallib" \
+    "$CHECKOUT_METALLIB_PATH" \
     "$SWIFT_DIR/.build/artifacts/mlx-swift/mlxc.artifactbundle/"*"/mlx.metallib" \
     "$(python3 -c 'import mlx; import os; print(os.path.join(os.path.dirname(mlx.__file__), "lib", "mlx.metallib"))' 2>/dev/null || echo '')" \
     "$HOME/Library/Developer/Xcode/DerivedData/"*"/Build/Products/"*"/mlx.metallib"
 do
     if [ -n "$candidate" ] && [ -f "$candidate" ]; then
-        MLX_METALLIB="$candidate"
-        break
+        if _metallib_has_gdn_kernels "$candidate"; then
+            MLX_METALLIB="$candidate"
+            break
+        fi
+        if [ -z "$MLX_METALLIB_FALLBACK" ]; then
+            MLX_METALLIB_FALLBACK="$candidate"
+        fi
     fi
 done
 
+if [ -z "$MLX_METALLIB" ] && [ -n "$MLX_METALLIB_FALLBACK" ]; then
+    MLX_METALLIB="$MLX_METALLIB_FALLBACK"
+fi
+
 if [ -n "$MLX_METALLIB" ]; then
     if [ "$MLX_METALLIB" != "$BUILD_DIR/mlx.metallib" ]; then
         cp "$MLX_METALLIB" "$BUILD_DIR/mlx.metallib"
@@ -107,6 +159,26 @@ except: pass
         echo "  To fix: pip install mlx && python3 -c 'import mlx.core; mlx.core.eval(mlx.core.array([1]))'  "
     fi
 fi
+
+if [ -f "$BUILD_DIR/mlx.metallib" ]; then
+    if _metallib_has_gdn_kernels "$BUILD_DIR/mlx.metallib"; then
+        echo "  Verified: gated_delta kernels present in mlx.metallib"
+    else
+        echo "  ERROR: gated_delta kernels NOT found in $BUILD_DIR/mlx.metallib"
+        echo "  Models like Qwen3.6-27B-ConfigI-MLX will fail at runtime."
+        if ! _metal_compiler_available; then
+            echo "  Metal compiler is unavailable. Install/update Metal Toolchain:"
+            echo "    xcodebuild -downloadComponent MetalToolchain"
+        fi
+        echo "  Verify with: strings $BUILD_DIR/mlx.metallib | grep gated_delta"
+        if [ "${VLLM_SWIFT_ALLOW_STOCK_METALLIB:-0}" != "1" ]; then
+            echo "  Failing install because required GDN kernels are missing."
+            echo "  Override (not recommended): VLLM_SWIFT_ALLOW_STOCK_METALLIB=1 ./scripts/install.sh"
+            exit 1
+        fi
+        echo "  WARNING: continuing because VLLM_SWIFT_ALLOW_STOCK_METALLIB=1"
+    fi
+fi
 echo ""
 
 # Find Python 3.10-3.13 (vLLM doesn't support 3.14+)
@@ -185,5 +257,10 @@ echo ""
 echo "Quick start:"
 echo "  cd $PROJECT_DIR"
 echo "  source activate.sh"
+echo "  # Baseline model"
 echo "  vllm serve ~/models/Qwen3-4B-4bit --max-model-len 4096"
 echo ""
+echo "  # ConfigI model (requires gated_delta kernels)"
+echo "  hf download thetom-ai/Qwen3.6-27B-ConfigI-MLX"
+echo "  vllm serve thetom-ai/Qwen3.6-27B-ConfigI-MLX --max-model-len 4096"
+echo ""