From 79ddfa2ef49a8cb0a83c0f9da57c8edc4277279a Mon Sep 17 00:00:00 2001
From: danielhanchen <unslothai@gmail.com>
Date: Wed, 29 Jul 2026 05:43:59 +0000
Subject: [PATCH] Reject booleans for the numeric override fields

Pydantic parses non-strictly and bool subclasses int, so a payload with
max_seq_length true was stored as 1, a one-token context, and gpu_ids [true] as
[1], an unintended GPU pin. _bounded_int already rejects bools for exactly that
reason, but never saw one: coercion happens at the route boundary first, which
left that guard unreachable through this path. A mode=before validator rejects
only booleans, including inside the gpu_ids list, so every other lax conversion
still runs and tensor_parallel, remove and fill_absent_fields keep working.
---
 studio/backend/routes/settings.py             | 25 ++++++++++-
 .../backend/tests/test_openai_auto_switch.py  | 43 +++++++++++++++++++
 2 files changed, 67 insertions(+), 1 deletion(-)

diff --git a/studio/backend/routes/settings.py b/studio/backend/routes/settings.py
index 67699d4e51..d4641f7981 100644
--- a/studio/backend/routes/settings.py
+++ b/studio/backend/routes/settings.py
@@ -2,7 +2,7 @@
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import re
-from typing import Literal, Optional
+from typing import Any, Literal, Optional
 from urllib.parse import unquote, urlsplit
 
 from fastapi import APIRouter, Depends, HTTPException
@@ -209,6 +209,29 @@ class ModelOverridePayload(BaseModel):
             raise ValueError(f"Chat template exceeds the {MAX_CHAT_TEMPLATE_BYTES}-byte limit.")
         return value
 
+    @field_validator(
+        "max_seq_length",
+        "custom_context_length",
+        "spec_draft_n_max",
+        "n_parallel",
+        "gpu_layers",
+        "n_cpu_moe",
+        "gpu_ids",
+        mode = "before",
+    )
+    @classmethod
+    def _no_booleans(cls, value: Any) -> Any:
+        # bool subclasses int and pydantic parses non-strictly, so `true` arrives
+        # as 1 and `false` as 0: a payload could pin GPU 1 or set a one-token
+        # context. _bounded_int in the normalizer rejects bools for exactly that
+        # reason, but never sees one, because coercion happens here first. Reject
+        # only bools, so every other lax conversion the field relies on still runs.
+        if isinstance(value, bool):
+            raise ValueError("Expected a number, got a boolean.")
+        if isinstance(value, list) and any(isinstance(item, bool) for item in value):
+            raise ValueError("Expected numbers, got a boolean.")
+        return value
+
 
 class ModelOverridesResponse(BaseModel):
     overrides: dict[str, dict]
diff --git a/studio/backend/tests/test_openai_auto_switch.py b/studio/backend/tests/test_openai_auto_switch.py
index 883ee47e3b..5fe77ef4e6 100644
--- a/studio/backend/tests/test_openai_auto_switch.py
+++ b/studio/backend/tests/test_openai_auto_switch.py
@@ -5627,3 +5627,46 @@ def test_a_fill_does_not_replay_a_stored_flag_through_validation(monkeypatch):
             "tester",
         )
     assert excinfo.value.status_code == 400
+
+
+def test_override_payload_rejects_booleans_for_numeric_fields():
+    """bool subclasses int and pydantic parses non-strictly, so `true` would
+    arrive as 1: `max_seq_length: true` becomes a one-token context and
+    `gpu_ids: [true]` pins GPU 1. _bounded_int rejects bools for exactly that
+    reason, but never sees one, because coercion happens at the route boundary
+    first. Reject them there so that guard is reachable through this path."""
+    import pytest
+    from pydantic import ValidationError
+    from routes.settings import ModelOverridePayload
+
+    for field, value in (
+        ("max_seq_length", True),
+        ("custom_context_length", True),
+        ("spec_draft_n_max", True),
+        ("n_parallel", True),
+        ("gpu_layers", False),
+        ("n_cpu_moe", True),
+        ("gpu_ids", [True]),
+        ("gpu_ids", [0, False, 2]),
+    ):
+        with pytest.raises(ValidationError):
+            ModelOverridePayload(model_id = "unsloth/x-GGUF:Q4_K_M", **{field: value})
+
+    # Only bools: every real value the picker sends still validates, and the
+    # fields that ARE booleans keep working.
+    ok = ModelOverridePayload(
+        model_id = "unsloth/x-GGUF:Q4_K_M",
+        max_seq_length = 4096,
+        gpu_layers = -1,
+        n_cpu_moe = 0,
+        gpu_ids = [0, 1],
+        tensor_parallel = True,
+        remove = True,
+        fill_absent_fields = True,
+    )
+    assert ok.max_seq_length == 4096
+    assert ok.gpu_ids == [0, 1]
+    assert ok.gpu_layers == -1
+    assert ok.tensor_parallel is True
+    assert ok.remove is True
+    assert ok.fill_absent_fields is True