offgrid-ai 0.9.6 → 0.10.1

package/README.md

@@ -2,7 +2,7 @@
 
 # offgrid-ai
 
-**Helper CLI for running local AI models on Mac with llama.cpp, ollama, and oMLX.**
+**Helper CLI for running local AI models on Mac with llama-server, mlx-vlm, and oMLX.**
 
 [![node](https://img.shields.io/badge/node-20%2B-3c873a)](package.json)
 [![platform](https://img.shields.io/badge/platform-macOS%20%7C%20Linux-blue)]()
@@ -12,19 +12,19 @@
 
 ## What is offgrid-ai?
 
-offgrid-ai is a command-line tool that lets you run AI models locally. Running local models with llama.cpp, ollama, or oMLX have a steep learning curve compared to cloud-based models, so offgrid-ai is designed to abstract away the complexity, while still providing a powerful and flexible way to run local models.
+offgrid-ai is a command-line tool that lets you run AI models locally. Running local models with llama-server, mlx-vlm, or oMLX have a steep learning curve compared to cloud-based models, so offgrid-ai is designed to abstract away the complexity, while still providing a powerful and flexible way to run local models.
 
 This is the recommended workflow:
 
-1. Download models from **LM Studio**, **Ollama**, or **oMLX**
+1. Download models from **LM Studio** or **oMLX**
 2. Do minimal configuration using the `offgrid-ai` command
 3. Run the model with `offgrid-ai` with Pi in interactive mode
 
 ## Core Features
-- Auto-detects available models from LM Studio, Ollama, and oMLX
+- Auto-detects available models from LM Studio, oMLX, and HuggingFace
 - Auto-detects MTP (multi-token prediction) or QAT (quantization aware training) models, and applies the correct flags for llama.cpp
-- Auto-applies the optimal flags for the model type in llama.cpp
-- Start / stop llama.cpp server automatically for chat sessions
+- Auto-applies the optimal flags for the model type (llama.cpp server flags, mlx-vlm APC/thinking/context flags)
+- Start / stop local servers automatically for chat sessions (llama-server and mlx-vlm)
 
 ## Quick start
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "offgrid-ai",
-  "version": "0.9.6",
+  "version": "0.10.1",
   "description": "Privacy-first CLI for running local LLMs — discover, configure, run, benchmark",
   "author": "Eeshan Srivastava (https://eeshans.com)",
   "type": "module",
@@ -12,6 +12,8 @@
     "src/*.mjs",
     "src/commands/*.mjs",
     "src/benchmark/*.mjs",
+    "resources/*.py",
+    "resources/recommendations.json",
     "install.sh"
   ],
   "publishConfig": {
@@ -32,7 +34,7 @@
     "start": "node bin/offgrid-ai.mjs",
     "test": "node --test test/*.mjs",
     "test:integration": "OFFGRID_INTEGRATION=1 node --test test/integration/*.mjs",
-    "lint": "eslint src/*.mjs src/commands/*.mjs src/benchmark/*.mjs bin/*.mjs",
+    "lint": "eslint src/*.mjs src/commands/*.mjs src/benchmark/*.mjs scripts/*.mjs bin/*.mjs",
     "check:privacy": "node scripts/privacy-gate.mjs",
     "release:check": "bash scripts/release-check.sh",
     "release:check:fast": "bash scripts/release-check.sh --skip-install --skip-manual",
@@ -46,7 +48,6 @@
   "keywords": [
     "local-llm",
     "llama-cpp",
-    "ollama",
     "cli",
     "privacy",
     "llm",

package/resources/hf-download.py

@@ -0,0 +1,79 @@
+#!/usr/bin/env python3
+"""
+Download a HuggingFace model into the standard HF cache.
+
+Usage:
+  python3 hf-download.py --repo mlx-community/gemma-4-e2b-it-4bit
+  python3 hf-download.py --repo unsloth/gemma-4-E2B-it-GGUF --file gemma-4-E2B-it-Q4_K_S.gguf
+
+Streams NDJSON progress events to stdout.
+"""
+import argparse
+import json
+import os
+import sys
+
+
+def emit(event):
+    print(json.dumps(event), flush=True)
+
+
+def progress_callback(relative_path, downloaded, total):
+    emit({
+        "type": "progress",
+        "file": relative_path,
+        "downloadedBytes": downloaded,
+        "totalBytes": total,
+    })
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Download a HuggingFace model into the standard cache.")
+    parser.add_argument("--repo", required=True, help="HuggingFace repo ID (e.g. mlx-community/gemma-4-e2b-it-4bit)")
+    parser.add_argument("--file", help="Specific filename to download (for GGUF). Omit to download the full repo (MLX).")
+    parser.add_argument("--cache-dir", help="HF hub cache directory (where models--org--name/... live). Defaults to $HF_HUB_CACHE or $HF_HOME/hub or ~/.cache/huggingface/hub.")
+    args = parser.parse_args()
+
+    try:
+        from huggingface_hub import hf_hub_download, snapshot_download
+    except ImportError as e:
+        emit({"type": "error", "message": f"huggingface_hub is not installed: {e}"})
+        sys.exit(1)
+
+    cache_dir = args.cache_dir or os.environ.get("HF_HUB_CACHE") or os.path.join(
+        os.environ.get("HF_HOME") or os.path.join(os.path.expanduser("~"), ".cache", "huggingface"),
+        "hub",
+    )
+
+    try:
+        if args.file:
+            local_path = hf_hub_download(
+                repo_id=args.repo,
+                filename=args.file,
+                cache_dir=cache_dir,
+                resume_download=True,
+            )
+            emit({
+                "type": "complete",
+                "localDir": os.path.dirname(local_path),
+                "localPath": local_path,
+                "format": "gguf",
+            })
+        else:
+            local_dir = snapshot_download(
+                repo_id=args.repo,
+                cache_dir=cache_dir,
+                resume_download=True,
+            )
+            emit({
+                "type": "complete",
+                "localDir": local_dir,
+                "format": "mlx",
+            })
+    except Exception as e:
+        emit({"type": "error", "message": str(e)})
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()

package/resources/mlxvlm-server-wrapper.py

@@ -0,0 +1,112 @@
+#!/usr/bin/env python3
+"""
+mlx-vlm server wrapper with strict=False model loading + APC merge fix.
+
+Two monkey-patches are applied before the server starts:
+
+1. strict=False model loading — needed for architectures with shared-KV weight
+   schemes (e.g. Gemma 4). Most models (Qwen, Llama, Mistral, Phi) load fine
+   with strict=True — strict=False is a no-op for them.
+
+2. BatchRotatingKVCache.merge() shape-mismatch fix — upstream mlx-lm bug
+   (ml-explore/mlx-lm PR #1116, Blaizzy/mlx-vlm Issue #923). The merge() method
+   crashes with `ValueError: [broadcast_shapes] Shapes (1,1,28,256) and
+   (1,1,512,256) cannot be broadcast` when APC merges exact-cache entries with
+   different fill levels. This affects all sliding-window attention models
+   (Gemma 4, Mistral, Mixtral). The fix uses explicit slicing instead of
+   negative indexing to guarantee exactly `l` elements are extracted.
+
+   This patch can be removed once mlx-lm fixes merge() upstream (not fixed in
+   0.31.2 or 0.31.3 — the merge() method is identical in both).
+
+Benchmark finding: mlx-vlm clears Metal cache after every request (GitHub Issue
+#999) unless APC_ENABLED=1 is set. The env var is set by the Electron app at
+spawn time, not in this wrapper.
+
+Usage:
+  python3 mlxvlm-server-wrapper.py --model <path> --host 127.0.0.1 --port <port>
+"""
+import sys
+
+# ── Patch 1: strict=False model loading ──────────────────────────────────────
+
+import mlx_vlm.utils as _utils
+_orig_load_model = _utils.load_model
+
+def _patched_load_model(model_path, lazy=False, strict=True, **kwargs):
+    return _orig_load_model(model_path, lazy=lazy, strict=False, **kwargs)
+
+_utils.load_model = _patched_load_model
+
+# ── Patch 2: BatchRotatingKVCache.merge() shape-mismatch fix ──────────────────
+#
+# Upstream bug: _temporal_order() can return a buffer whose seq dimension differs
+# from c.size(). The negative slice [..., -l:, :] then produces a mismatched shape,
+# crashing with ValueError: [broadcast_shapes].
+#
+# Fix: use explicit slicing to extract exactly `l` elements, right-aligning within
+# the target slice when the buffer is shorter than `l` (left-padded by zeros from
+# the pre-allocated target tensor).
+
+import mlx.core as mx
+from mlx_lm.models import cache as _lm_cache
+
+_orig_merge = _lm_cache.BatchRotatingKVCache.merge
+
+@classmethod
+def _patched_merge(cls, caches):
+    if not all(c.max_size == caches[0].max_size for c in caches):
+        raise ValueError(
+            "BatchRotatingKVCache can only merge caches with the same maximum size"
+        )
+
+    offsets = [c.offset for c in caches]
+    lengths = [c.size() for c in caches]
+    max_length = max(lengths)
+
+    if max_length == 0:
+        return cls(caches[0].max_size, [0] * len(caches))
+
+    padding = [max_length - l for l in lengths]
+    B = len(caches)
+    H = max(c.keys.shape[1] for c in caches if c.keys is not None)
+    Dk = max(c.keys.shape[3] for c in caches if c.keys is not None)
+    Dv = max(c.values.shape[3] for c in caches if c.values is not None)
+    dt = next(iter(c.keys.dtype for c in caches if c.keys is not None))
+
+    keys = mx.zeros((B, H, max_length, Dk), dtype=dt)
+    values = mx.zeros((B, H, max_length, Dv), dtype=dt)
+    for i, (p, l, c) in enumerate(zip(padding, lengths, caches)):
+        if c.keys is None:
+            continue
+        ordered_k = c._temporal_order(c.keys)
+        ordered_v = c._temporal_order(c.values)
+        seq_len = ordered_k.shape[2]
+        if seq_len >= l:
+            # Normal case: extract the last `l` tokens.
+            start = seq_len - l
+            keys[i : i + 1, :, p : p + l] = ordered_k[..., start : start + l, :]
+            values[i : i + 1, :, p : p + l] = ordered_v[..., start : start + l, :]
+        else:
+            # Buffer shorter than l: right-align within the slice (left-padded
+            # by zeros from the pre-allocated target tensor).
+            gap = l - seq_len
+            keys[i : i + 1, :, p + gap : p + l] = ordered_k
+            values[i : i + 1, :, p + gap : p + l] = ordered_v
+
+    cache = cls(caches[0].max_size, padding)
+    cache.keys = keys
+    cache.values = values
+    cache.offset = mx.array(offsets)
+    cache._idx = keys.shape[2]
+    cache._offset = keys.shape[2]
+
+    return cache
+
+_lm_cache.BatchRotatingKVCache.merge = _patched_merge
+
+# ── Run the server ────────────────────────────────────────────────────────────
+# main() parses sys.argv for --model, --host, --port, etc.
+from mlx_vlm.server import main
+main()
+

package/resources/recommendations.json

@@ -0,0 +1,60 @@
+{
+  "models": [
+    {
+      "id": "gemma-4-e2b",
+      "label": "Gemma 4 E2B",
+      "minRamGb": 8,
+      "gguf": "unsloth/gemma-4-E2B-it-GGUF/gemma-4-E2B-it-Q4_K_S.gguf",
+      "mlx": "mlx-community/gemma-4-e2b-it-4bit"
+    },
+    {
+      "id": "qwen-3.5-9b",
+      "label": "Qwen 3.5 9B",
+      "minRamGb": 16,
+      "gguf": "unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-UD-Q4_K_S.gguf",
+      "mlx": "lmstudio-community/Qwen3.5-9B-MLX-4bit"
+    },
+    {
+      "id": "gemma-4-12b-qat",
+      "label": "Gemma 4 12B",
+      "minRamGb": 24,
+      "gguf": "unsloth/gemma-4-12B-it-qat-GGUF/gemma-4-12B-it-qat-UD-Q4_K_XL.gguf",
+      "mlx": "mlx-community/gemma-4-12B-it-qat-4bit"
+    },
+    {
+      "id": "gemma-4-26b",
+      "label": "Gemma 4 26B",
+      "minRamGb": 32,
+      "gguf": "unsloth/gemma-4-26B-A4B-it-qat-GGUF/gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf",
+      "mlx": "mlx-community/gemma-4-26b-a4b-4bit"
+    },
+    {
+      "id": "qwen-3.6-35b-compact",
+      "label": "Qwen 3.6 35B",
+      "minRamGb": 32,
+      "gguf": "unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_S.gguf",
+      "mlx": "mlx-community/Qwen3.6-35B-A3B-4bit"
+    },
+    {
+      "id": "qwen-3.6-35b",
+      "label": "Qwen 3.6 35B",
+      "minRamGb": 48,
+      "gguf": "unsloth/Qwen3.6-35B-A3B-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf",
+      "mlx": ""
+    },
+    {
+      "id": "gemma-4-31b",
+      "label": "Gemma 4 31B",
+      "minRamGb": 64,
+      "gguf": "unsloth/gemma-4-31B-it-qat-GGUF/gemma-4-31B-it-qat-UD-Q4_K_XL.gguf",
+      "mlx": "mlx-community/gemma-4-31b-4bit"
+    },
+    {
+      "id": "qwen-3.6-27b",
+      "label": "Qwen 3.6 27B",
+      "minRamGb": 64,
+      "gguf": "unsloth/Qwen3.6-27B-MTP-GGUF/Qwen3.6-27B-Q4_K_M.gguf",
+      "mlx": "mlx-community/Qwen3.6-27B-4bit"
+    }
+  ]
+}

package/src/backend-installers.mjs

@@ -15,21 +15,6 @@ export const BACKEND_INSTALLERS = {
     failure: "Download it manually from https://lmstudio.ai",
     allFailure: "✗ LM Studio installation failed. Download from https://lmstudio.ai",
   },
-  ollama: {
-    label: "Ollama",
-    choiceLabel: "Ollama",
-    hint: "brew install ollama — models download on demand",
-    commands: [["brew", ["install", "ollama"], "Ollama"]],
-    success(model) {
-      console.log(pc.green("✓ Ollama installed"));
-      console.log(pc.yellow("\nStart Ollama and pull a model:"));
-      console.log(pc.bold(`  ollama serve &   ollama pull ${model.ollama}`));
-      console.log(pc.dim(`Recommended for your machine: ${model.label}`));
-      console.log(pc.dim("Then run offgrid-ai again to pick and run a model."));
-    },
-    failure: "Install it manually from https://ollama.com",
-    allFailure: "✗ Ollama installation failed. Install manually from https://ollama.com",
-  },
   omlx: {
     label: "oMLX",
     choiceLabel: "oMLX",
@@ -52,6 +37,6 @@ export const BACKEND_INSTALLERS = {
 
 export const BACKEND_INSTALL_CHOICES = [
   ...Object.entries(BACKEND_INSTALLERS).map(([value, installer]) => ({ value, label: installer.choiceLabel, hint: installer.hint })),
-  { value: "all", label: "Install all three", hint: "LM Studio + Ollama + oMLX" },
+  { value: "all", label: "Install both", hint: "LM Studio + oMLX" },
   { value: "skip", label: "Skip for now", hint: "I'll set up models myself" },
 ];

package/src/backends.mjs

@@ -1,13 +1,14 @@
 import { findLlamaServer } from "./config.mjs";
 import { scanGgufModels } from "./scan.mjs";
 import { parseModelName } from "./model-name.mjs";
+import { scanMlxModels } from "./mlx-discovery.mjs";
+import { DEFAULT_PORT as MLX_VLM_PORT } from "./mlx-flags.mjs";
 
 // ── Backend definitions ────────────────────────────────────────────────────
 
 export const LOCAL_HOST = "127.0.0.1";
 export const LLAMA_CPP_PORT = 8080;
 export const LLAMA_CPP_MTP_PORT = 8081;
-export const OLLAMA_PORT = 11434;
 export const OMLX_PORT = 8000;
 
 export function baseUrlFor({ host = LOCAL_HOST, port, path = "/v1" }) {
@@ -41,18 +42,6 @@ export const BACKENDS = {
     needsCommandFile: true,
     scanModels: async () => (await scanGgufModels()).models,
   },
-  "ollama": {
-    id: "ollama",
-    label: "Ollama",
-    type: "managed-server",
-    providerId: "ollama",
-    defaultHost: "localhost",
-    defaultPort: OLLAMA_PORT,
-    defaultBaseUrl: baseUrlFor({ host: "localhost", port: OLLAMA_PORT }),
-    apiBaseUrl: baseUrlFor({ host: "localhost", port: OLLAMA_PORT, path: "" }),
-    needsCommandFile: false,
-    scanModels: () => scanOllamaModels(),
-  },
   "omlx": {
     id: "omlx",
     label: "oMLX",
@@ -65,6 +54,17 @@ export const BACKENDS = {
     needsCommandFile: false,
     scanModels: () => scanOmlxModels(),
   },
+  "mlx-vlm": {
+    id: "mlx-vlm",
+    label: "mlx-vlm",
+    type: "local-server",
+    providerId: "mlx-vlm",
+    defaultHost: LOCAL_HOST,
+    defaultPort: MLX_VLM_PORT,
+    defaultBaseUrl: baseUrlFor({ port: MLX_VLM_PORT }),
+    needsCommandFile: true,
+    scanModels: async () => scanMlxModels(),
+  },
 };
 
 export function backendFor(backendId) {
@@ -75,6 +75,7 @@ export function backendFor(backendId) {
 
 export async function backendBinaryFor(backendId) {
   const backend = BACKENDS[backendId ?? "llama-cpp"];
+  if (backend.id === "mlx-vlm") return "python3"; // mlx-vlm spawns via python3 + the strict=False wrapper
   if (backend.type === "managed-server") return null;
   const discovered = await findLlamaServer();
   return discovered; // null means "not found — trigger onboarding"
@@ -85,29 +86,6 @@ export function defaultFlagsForBackend(backendId) {
   return { host: backend.defaultHost ?? LOCAL_HOST, port: backend.defaultPort };
 }
 
-// ── Ollama model discovery ──────────────────────────────────────────────
-
-async function scanOllamaModels() {
-  const response = await fetch(`${BACKENDS.ollama.apiBaseUrl}/api/tags`, { signal: AbortSignal.timeout(3000) });
-  if (!response.ok) {
-    throw new Error(`Ollama /api/tags returned ${response.status} ${response.statusText}`);
-  }
-  const body = await response.json();
-  if (!Array.isArray(body?.models)) return [];
-  return body.models
-    .filter((model) => isLocalOllamaModel(model))
-    .map((model) => ({
-      id: model.name,
-      label: parseModelName(model.name, "ollama").display,
-      aliasSuggestion: model.name,
-      sizeBytes: model.size ?? 0,
-      quant: model.details?.quantization_level,
-      family: model.details?.family,
-      backend: "ollama",
-      source: "ollama",
-    })).sort((a, b) => a.label.localeCompare(b.label));
-}
-
 // ── oMLX model discovery ───────────────────────────────────────────────
 
 async function scanOmlxModels() {
@@ -123,7 +101,8 @@ async function scanOmlxModels() {
       id: model.id,
       label: parseModelName(model.id, "omlx").display,
       aliasSuggestion: model.id,
-      sizeBytes: 0,
+      sizeBytes: model.size ?? 0,
+      contextLength: model.max_model_len ?? null,
       quant: null,
       family: null,
       backend: "omlx",
@@ -133,13 +112,6 @@ async function scanOmlxModels() {
 
 // ── Labels ──────────────────────────────────────────────────────────────
 
-function isLocalOllamaModel(model) {
-  const name = String(model?.name ?? "");
-  if (/:cloud(?:$|\b)/i.test(name)) return false;
-  if (!Number.isFinite(model?.size) || model.size <= 0) return false;
-  return true;
-}
-
 function isChatOmlxModel(model) {
   if (typeof model?.id !== "string" || !model.id.trim()) return false;
   const type = String(model.type ?? model.model_type ?? "").toLowerCase();
@@ -148,4 +120,5 @@ function isChatOmlxModel(model) {
   return true;
 }
 
-// (ollamaLabel and omlxLabel removed — parseModelName in model-name.mjs is the single path)
+// (ollamaLabel and omlxLabel removed — parseModelName in model-name.mjs is the single path)
+// (Ollama backend removed — offgrid-ai now uses llama-server + mlx-vlm + oMLX)

package/src/benchmark/finalize.mjs

@@ -1,99 +1,12 @@
-// ── Unload model from server memory after benchmark ────────────────────────────
+// ── Benchmark finalization (metadata + summary rendering) ───────────────────
+// unloadModelFromServer has been moved to src/process.mjs (it's the managed-server
+// counterpart to stopProfile, used by both the benchmark flow and the Pi chat flow).
 
-import { backendFor } from "../backends.mjs";
-import { apiRootUrl, serverModelIds } from "../process.mjs";
 import { existsSync } from "node:fs";
 import { readFile, writeFile } from "node:fs/promises";
 import { join } from "node:path";
 import { pc, renderRows, renderSection } from "../ui.mjs";
 
-export async function unloadModelFromServer(profile) {
-  const backend = backendFor(profile.backend);
-
-  if (backend.id === "ollama") {
-    const apiBaseUrl = apiRootUrl(profile.baseUrl || backend.apiBaseUrl || "");
-
-    try {
-      await fetch(`${apiBaseUrl}/api/generate`, {
-        method: "POST",
-        headers: { "Content-Type": "application/json" },
-        body: JSON.stringify({ model: profile.modelAlias, prompt: "", stream: false, keep_alive: 0 }),
-        signal: AbortSignal.timeout(10000),
-      });
-      return { unloaded: true, backend: backend.id };
-    } catch (err) {
-      return { unloaded: false, backend: backend.id, error: err.message };
-    }
-  }
-
-  if (backend.id === "llama-cpp" || backend.id === "llama-cpp-mtp") {
-    // llama.cpp unloads when the server process exits; no HTTP unload API exists.
-    // If offgrid-ai started the server, stopProfile already handled it.
-    return { unloaded: false, backend: backend.id, reason: "stop server to unload" };
-  }
-
-  if (backend.id === "omlx") {
-    return await unloadOmlxModel(profile);
-  }
-
-  return { unloaded: false, backend: backend.id, reason: "unsupported backend" };
-}
-
-async function unloadOmlxModel(profile) {
-  const baseUrl = profile.baseUrl?.replace(/\/v1\/?$/u, "") || "";
-  const adminUrl = `${baseUrl}/admin/api/models`;
-  const modelId = profile.modelAlias || profile.omlxModel || profile.id;
-
-  try {
-    const ids = await serverModelIds(profile.baseUrl);
-    const match = ids.find((id) => id.toLowerCase() === modelId.toLowerCase());
-    const targetId = match ?? modelId;
-
-    const response = await fetch(`${adminUrl}/${encodeURIComponent(targetId)}/unload`, {
-      method: "POST",
-      headers: { "Content-Type": "application/json" },
-      signal: AbortSignal.timeout(30000),
-    });
-
-    if (response.ok) {
-      return { unloaded: true, backend: "omlx", modelId: targetId };
-    }
-
-    const detail = await responseErrorDetail(response);
-
-    if (response.status === 400 && /not loaded/i.test(detail)) {
-      return { unloaded: true, backend: "omlx", modelId: targetId, reason: "model was not loaded" };
-    }
-
-    if (response.status === 401 || response.status === 403) {
-      return {
-        unloaded: false,
-        backend: "omlx",
-        modelId: targetId,
-        error: "oMLX admin authentication required. Enable skip_api_key_verification in oMLX settings, or unload manually from the admin panel.",
-      };
-    }
-
-    return { unloaded: false, backend: "omlx", modelId: targetId, error: `HTTP ${response.status}: ${detail}` };
-  } catch (err) {
-    if (err?.name === "AbortError" || err?.name === "TimeoutError") {
-      return { unloaded: false, backend: "omlx", modelId, error: "Unload request timed out. The model may still be unloading in the background." };
-    }
-    return { unloaded: false, backend: "omlx", modelId, error: err.message };
-  }
-}
-
-async function responseErrorDetail(response) {
-  const text = await response.text().catch(() => "");
-  if (!text) return "";
-  try {
-    const body = JSON.parse(text);
-    return body?.detail ?? body?.message ?? text;
-  } catch {
-    return text;
-  }
-}
-
 export async function finalizeBenchmarkRun(runDirectory, runResult, speedMetrics, speedMetricsError = null) {
   const metadataPath = join(runDirectory, "metadata.json");
   const metadata = JSON.parse(await readFile(metadataPath, "utf8"));

package/src/benchmark/flow.mjs

@@ -4,7 +4,7 @@ import { join } from "node:path";
 import { ensureDirs } from "../config.mjs";
 import { backendFor } from "../backends.mjs";
 import { hasPi, hasPiModel, syncPiConfig } from "../harness-pi.mjs";
-import { serverReady, startServer, waitForReady, stopProfile, modelAvailableOnServer } from "../process.mjs";
+import { serverReady, startServer, waitForReady, stopProfile, modelAvailableOnServer, unloadModelFromServer } from "../process.mjs";
 import { loadProfiles } from "../profiles.mjs";
 import { pc, createPrompt } from "../ui.mjs";
 import { linkBenchmarkRepo } from "./repo.mjs";
@@ -12,12 +12,11 @@ import { loadBenchmarks } from "./shared.mjs";
 import { prepareBenchmarkRun } from "./prepare.mjs";
 import { runBenchmarkInPi } from "./pi-runner.mjs";
 import { queryServerMetrics } from "./metrics.mjs";
-import { unloadModelFromServer } from "./finalize.mjs";
 import { finalizeBenchmarkRun, renderBenchmarkSummary } from "./finalize.mjs";
 
 function benchmarkModelSource(profile) {
   if (!profile) return "cloud";
-  return profile.providerId === "llama-cpp-mtp" ? "llama-cpp-mtp" : profile.backend === "ollama" ? "ollama" : profile.backend === "omlx" ? "omlx" : "llama-cpp";
+  return profile.providerId === "llama-cpp-mtp" ? "llama-cpp-mtp" : profile.backend === "omlx" ? "omlx" : "llama-cpp";
 }
 
 async function chooseBenchmarkAction(prompt, canRun) {
@@ -29,7 +28,7 @@ async function chooseBenchmarkAction(prompt, canRun) {
 }
 
 function managedModelId(profile) {
-  return profile.omlxModel ?? profile.ollamaModel ?? profile.modelAlias ?? profile.label;
+  return profile.omlxModel ?? profile.modelAlias ?? profile.label;
 }
 
 async function ensureManagedModelAvailableForBenchmark(profile, backend) {

package/src/benchmark/metrics.mjs

@@ -1,7 +1,6 @@
 // ── Backend-aware server speed metrics ───────────────────────────────────────
 
 import { backendFor } from "../backends.mjs";
-import { apiRootUrl } from "../process.mjs";
 
 const BENCH_SPEED_PROMPT = "Write a one-sentence summary of machine learning.";
 const SPEED_QUERY_TIMEOUT_MS = 120_000;
@@ -16,9 +15,6 @@ export async function queryServerMetrics(profile) {
   if (backend.id === "omlx") {
     return await queryOmlxMetrics(profile);
   }
-  if (backend.id === "ollama") {
-    return await queryOllamaMetrics(profile);
-  }
 
   throw new Error(`Unsupported backend for benchmark speed metrics: ${backend.id}`);
 }
@@ -115,43 +111,3 @@ async function queryOmlxMetrics(profile) {
     metricSource: "oMLX /v1/chat/completions streaming include_usage",
   };
 }
-
-async function queryOllamaMetrics(profile) {
-  const body = {
-    model: profile.modelAlias,
-    prompt: BENCH_SPEED_PROMPT,
-    stream: false,
-    options: { num_predict: SPEED_QUERY_MAX_TOKENS },
-  };
-
-  const apiBaseUrl = apiRootUrl(profile.baseUrl || backendFor(profile.backend).apiBaseUrl || "");
-
-  const response = await fetch(`${apiBaseUrl}/api/generate`, {
-    method: "POST",
-    headers: { "Content-Type": "application/json" },
-    body: JSON.stringify(body),
-    signal: AbortSignal.timeout(SPEED_QUERY_TIMEOUT_MS),
-  });
-
-  if (!response.ok) {
-    throw new Error(`Ollama speed query failed: ${response.status} ${response.statusText}`);
-  }
-
-  const data = await response.json();
-  const promptEvalNs = data.prompt_eval_duration ?? 0;
-  const evalNs = data.eval_duration ?? 0;
-  const loadNs = data.load_duration ?? 0;
-
-  const promptEvalCount = data.prompt_eval_count ?? 0;
-  const evalCount = data.eval_count ?? 0;
-
-  return {
-    prefillTokensPerSecond: promptEvalNs > 0 ? (promptEvalCount / (promptEvalNs / 1e9)) : null,
-    generationTokensPerSecond: evalNs > 0 ? (evalCount / (evalNs / 1e9)) : null,
-    ttftMs: promptEvalNs / 1e6,
-    modelLoadMs: loadNs / 1e6,
-    speculativeDecodeAcceptance: null,
-    kvCacheTokens: null,
-    metricSource: "Ollama /api/generate",
-  };
-}

package/src/benchmark/prepare.mjs

@@ -55,7 +55,7 @@ export async function prepareBenchmarkRun({ repoPath, benchmark, kind, modelId,
     kind,
     runId,
     benchmark: { id: benchmark.id, title: benchmark.title, description: benchmark.description, prompt: benchmark.prompt },
-    model: { id: modelId, slug: modelSlug, displayName: parseModelName(modelId, modelSource === "ollama" ? "ollama" : modelSource === "omlx" ? "omlx" : "local-gguf").display },
+    model: { id: modelId, slug: modelSlug, displayName: parseModelName(modelId, modelSource === "omlx" ? "omlx" : "local-gguf").display },
     status: "prepared",
     createdAt: now.toISOString(),
     updatedAt: now.toISOString(),