From 340584d5c0afd22101d25ab8266dd1f0d4aa4f4b Mon Sep 17 00:00:00 2001
From: Datta Nimmaturi <venkatadattasainimmaturi@gmail.com>
Date: Thu, 4 Jun 2026 13:57:44 +0000
Subject: [PATCH] Disable auto MTP on Windows and macOS

---
 studio/backend/core/inference/llama_cpp.py    | 21 +++++++--
 .../tests/test_llama_cpp_mtp_detection.py     | 45 +++++++++++++++++++
 2 files changed, 63 insertions(+), 3 deletions(-)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 0f23549138..834671388e 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -492,6 +492,10 @@ def _extra_args_set_spec_type(extra_args: Optional[Iterable[str]]) -> bool:
     return False
 
 
+def _auto_mtp_supported_platform() -> bool:
+    return sys.platform not in ("win32", "darwin")
+
+
 def _build_ngram_mod_flags(
     caps: Optional[dict],
     n_match: int = 24,
@@ -2919,6 +2923,7 @@ class LlamaCppBackend:
                             _mtp_effective in ("mtp", "mtp+ngram")
                             or (
                                 _mtp_effective == "auto"
+                                and _auto_mtp_supported_platform()
                                 and (
                                     bool(self._nextn_predict_layers)
                                     or _is_mtp_model_name(model_identifier, model_path)
@@ -3635,7 +3640,8 @@ class LlamaCppBackend:
           4B+ GPU/CPU: spec on is a net win (1.08x-1.46x).
         Auto falls back to ngram-mod (zero-VRAM, near-zero idle cost on
         diverse content); forced MTP variants engage anyway and just log
-        a warning per the user's choice.
+        a warning per the user's choice. Windows and macOS skip auto-MTP
+        until the llama.cpp MTP path no longer pegs CPU during generation.
         """
         flags: List[str] = []
         # Reset; emit branches re-set on the resolved emission.
@@ -3779,9 +3785,18 @@ class LlamaCppBackend:
 
         # effective_mode == "auto": today's promotion path. llama.cpp
         # #22673: MTP is compatible with mmproj, so there's no vision gate.
-        if is_mtp_model and not _mtp_too_small:
-            # GPU: MTP-only. CPU/Mac: chain ngram-mod + MTP.
+        if (
+            is_mtp_model
+            and not _mtp_too_small
+            and _auto_mtp_supported_platform()
+        ):
+            # Supported platforms: GPU = MTP-only; CPU = chain ngram-mod + MTP.
             _emit_mtp(chain_ngram = not gpus)
+        elif is_mtp_model and not _mtp_too_small:
+            logger.info(
+                "MTP GGUF detected; auto-disabling MTP on Windows/macOS. "
+                "Use the Studio Speculative Decoding dropdown to force it."
+            )
         elif is_mtp_model and _mtp_too_small:
             # Sub-3B fallback: drop the MTP draft head, keep ngram-mod
             # when the binary supports it.
diff --git a/studio/backend/tests/test_llama_cpp_mtp_detection.py b/studio/backend/tests/test_llama_cpp_mtp_detection.py
index 4a8276adc0..73bc555018 100644
--- a/studio/backend/tests/test_llama_cpp_mtp_detection.py
+++ b/studio/backend/tests/test_llama_cpp_mtp_detection.py
@@ -1128,6 +1128,51 @@ def test_build_speculative_flags_matrix(
         assert "--spec-ngram-mod-n-match" not in parsed
 
 
+@pytest.mark.parametrize("platform_name", ["win32", "darwin"])
+def test_build_speculative_flags_auto_skips_mtp_on_windows_and_macos(
+    monkeypatch, platform_name
+):
+    monkeypatch.setattr(sys, "platform", platform_name)
+    backend = _resolver_backend(monkeypatch)
+
+    flags = backend._build_speculative_flags(
+        speculative_type = "auto",
+        spec_draft_n_max = None,
+        extra_args = None,
+        model_identifier = _MTP_MODEL,
+        model_path = None,
+        gpus = True,
+        binary = "/fake/llama-server",
+    )
+
+    parsed = _flags_dict(flags)
+    assert "--spec-type" not in parsed
+    assert "--spec-draft-n-max" not in parsed
+    assert backend.requested_spec_mode == "auto"
+    assert backend.speculative_type is None
+
+
+def test_build_speculative_flags_forced_mtp_still_works_on_windows(monkeypatch):
+    monkeypatch.setattr(sys, "platform", "win32")
+    backend = _resolver_backend(monkeypatch)
+
+    flags = backend._build_speculative_flags(
+        speculative_type = "mtp",
+        spec_draft_n_max = None,
+        extra_args = None,
+        model_identifier = _MTP_MODEL,
+        model_path = None,
+        gpus = True,
+        binary = "/fake/llama-server",
+    )
+
+    parsed = _flags_dict(flags)
+    assert parsed.get("--spec-type") == "draft-mtp"
+    assert parsed.get("--spec-draft-n-max") == "2"
+    assert backend.requested_spec_mode == "mtp"
+    assert backend.speculative_type == "draft-mtp"
+
+
 def test_build_speculative_flags_user_extra_args_owns_spec_type(monkeypatch):
     # User --spec-type in extra_args bypasses the dropdown entirely.
     backend = _resolver_backend(monkeypatch)