Drop GPU-only flags from the CPU-only (-ngl 0) serve command (#1433)

A CPU-only llama.cpp serve config still emitted --flash-attn on and exported
GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 (independent toggles, often left on by an Auto
profile), so the command mixed "zero GPU layers" with CUDA/flash-attn and failed
to start (issue #1291). Gate both on a _cpuOnly check (ngl == 0). GPU serving is
unchanged — the gate only affects the ngl=0 path.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

tests/test_cookbook_cpu_only_serve.py

@@ -0,0 +1,30 @@
+"""Regression guard for issue #1291 — CPU-only serve still emitted GPU-only flags.
+
+The llama.cpp serve command builder (static/js/cookbook.js) added
+`--flash-attn on` and exported `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` from
+independent toggles, so a CPU-only config (`-ngl 0`, often with flash-attn left
+on by an Auto profile) produced a command that mixes "zero GPU layers" with
+CUDA/flash-attn and fails to start. The builder now drops those GPU-only flags
+when ngl == 0, per the maintainer's guidance.
+
+cookbook.js pulls in browser globals so it can't run under node; guard the fix
+at the source level: a `_cpuOnly` gate exists and is applied to flash-attn and
+the CUDA unified-memory env.
+"""
+import re
+from pathlib import Path
+
+SRC = Path(__file__).resolve().parent.parent / "static/js/cookbook.js"
+
+
+def test_cpu_only_drops_gpu_only_flags():
+    text = SRC.read_text(encoding="utf-8")
+    # A CPU-only flag derived from ngl == 0.
+    assert re.search(r"_cpuOnly\s*=\s*String\(f\.ngl\)\.trim\(\)\s*===\s*'0'", text), \
+        "expected a _cpuOnly gate derived from ngl==0"
+    # flash-attn must be suppressed for CPU-only.
+    assert re.search(r"if\s*\(\s*f\.flash_attn\s*&&\s*!_cpuOnly\s*\)", text), \
+        "flash-attn must be gated on !_cpuOnly"
+    # The CUDA unified-memory env must be suppressed for CPU-only too.
+    assert "f.unified_mem && !_cpuOnly" in text, \
+        "GGML_CUDA_ENABLE_UNIFIED_MEMORY must be gated on !_cpuOnly"