Studio: respect explicit --numa and reorder the NUMA total-RAM check

Follow-up review on the CPU-only / NUMA hardening:

- Respect an explicit user --numa policy. Auto-interleave wraps the process with
  a numactl argv prefix the user cannot override, so when the user sets --numa we
  now skip the wrapper (and --numa distribute) instead of forcing interleaving
  over their placement.
- Check the total-RAM condition before numactl availability in decide_interleave,
  so a model too large for total free RAM gets the smaller-quant / free-memory
  guidance instead of being told to install numactl, which could never help.

Unit tests cover both cases.

studio/backend/core/inference/llama_cpp.py

@@ -6048,10 +6048,14 @@ class LlamaCppBackend:
                         except Exception:
                             _numa_footprint = _resident
                     _numa = decide_interleave(_numa_footprint, cpu_only = _cpu_only)
-                    if _numa.interleave:
+                    if _numa.interleave and _extra_args_set_any_flag(extra_args, {"--numa"}):
+                        # User set an explicit --numa policy: respect it. The numactl wrap
+                        # is an argv prefix they can't override, so skip it (and --numa
+                        # distribute) rather than force interleaving over their choice.
+                        logger.info("NUMA: user --numa set; leaving auto-interleave off")
+                    elif _numa.interleave:
                         self._numa_prefix = list(_numa.prefix)
-                        if not _extra_args_set_any_flag(extra_args, {"--numa"}):
-                            cmd.extend(["--numa", "distribute"])
+                        cmd.extend(["--numa", "distribute"])
                         logger.info("NUMA: %s", _numa.reason)
                     elif _cpu_only and (
                         "numactl` is not installed" in _numa.reason

studio/backend/core/inference/numa.py

@@ -135,6 +135,16 @@ def decide_interleave(
             f"(~{largest} MiB); keeping local placement",
         )
 
+    # Impossible across all nodes regardless of numactl: surface the smaller-quant /
+    # free-memory path before the numactl hint, so a too-big model is not told to
+    # install numactl when interleaving could never make it fit.
+    if model_mib > total:
+        return InterleaveDecision(
+            False,
+            f"model ~{model_mib} MiB exceeds total free RAM across all nodes "
+            f"(~{total} MiB); interleave cannot help -- free memory or use a smaller quant",
+        )
+
     avail = numactl_available() if has_numactl is None else has_numactl
     if not avail:
         # Needed but unavailable: surface it; caller decides whether to block.
@@ -146,13 +156,6 @@ def decide_interleave(
             f"numactl`) or the model may fail to fit a single node.",
         )
 
-    if model_mib > total:
-        return InterleaveDecision(
-            False,
-            f"model ~{model_mib} MiB exceeds total free RAM across all nodes "
-            f"(~{total} MiB); interleave cannot help -- free memory or use a smaller quant",
-        )
-
     return InterleaveDecision(
         True,
         f"model ~{model_mib} MiB exceeds the largest NUMA node's free RAM "

studio/backend/tests/test_cpu_only_defaults.py

@@ -182,6 +182,14 @@ def test_numa_surfaces_total_ram_failure():
     assert '"interleave cannot help" in _numa.reason' in src
 
 
+def test_explicit_user_numa_skips_auto_interleave_prefix():
+    """An explicit user --numa must skip the numactl argv prefix (which user extra args
+    can't override), not just the --numa distribute flag (PR review fix)."""
+    src = _load_model_src()
+    assert 'if _numa.interleave and _extra_args_set_any_flag(extra_args, {"--numa"}):' in src
+    assert "leaving auto-interleave off" in src
+
+
 def test_extra_args_forces_cpu_offload_helper():
     """The zero-offload detector: -ngl 0 / --n-gpu-layers 0 / --gpu-layers 0 (last wins)."""
     from core.inference.llama_cpp import _extra_args_forces_cpu_offload as f

studio/backend/tests/test_numa_interleave.py

@@ -87,6 +87,15 @@ def test_numactl_missing_surfaces_actionable_warning():
     assert "numactl` is not installed" in d.reason
 
 
+def test_too_big_and_numactl_missing_prefers_total_ram_message():
+    # Impossible across all nodes AND no numactl: the total-RAM guidance must win, so the
+    # user is not told to install numactl when interleaving could never help (PR review fix).
+    d = decide_interleave(800 * _GiB, cpu_only = True, topology = _USER_TOPO, has_numactl = False)
+    assert d.interleave is False
+    assert "exceeds total free RAM" in d.reason
+    assert "numactl` is not installed" not in d.reason
+
+
 def test_unknown_model_size_does_not_force_interleave():
     for size in (None, 0, -1):
         d = decide_interleave(size, cpu_only = True, topology = _USER_TOPO, has_numactl = True)