fix(studio): inherit llama_extra_args and honor --no-mmproj (#5902)

* fix(studio): inherit llama_extra_args and honor --no-mmproj

Reloading the same GGUF from the UI without gguf_variant no longer drops
CLI pass-through args like --no-mmproj. Skip mmproj download and launch
when --no-mmproj is present in llama_extra_args.

Co-authored-by: Cursor <cursoragent@cursor.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix(studio): tighten GGUF llama_extra_args variant inheritance guard

Reject inherited CLI args when the request changes gguf_variant or when
omitted variant resolves differently from the stored extra_args source.

Co-authored-by: Cursor <cursoragent@cursor.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Treat --no-mmproj-auto and --mmproj-auto with last-wins parsing for PR #5902

---------

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>

studio/backend/core/inference/llama_cpp.py

@@ -28,6 +28,7 @@ from typing import Callable, Generator, Iterable, List, Optional
 import httpx
 
 from core.inference.llama_server_args import (
+    extra_args_disable_mmproj,
     parse_cache_override,
     parse_ctx_override,
     resolve_cache_type_kv,
@@ -2929,8 +2930,8 @@ class LlamaCppBackend:
                         hf_variant = hf_variant,
                         hf_token = hf_token,
                     )
-                    # Auto-download mmproj for vision models
-                    if is_vision and not mmproj_path:
+                    # Auto-download mmproj for vision models unless opted out.
+                    if is_vision and not mmproj_path and not extra_args_disable_mmproj(extra_args):
                         mmproj_path = self._download_mmproj(
                             hf_repo = hf_repo,
                             hf_token = hf_token,
@@ -3191,10 +3192,12 @@ class LlamaCppBackend:
                     gpu_indices, use_fit = None, True
                     effective_ctx = requested_ctx  # fall back to original
 
-                launch_mmproj_path = self._resolve_launch_mmproj_path(
-                    model_path = model_path,
-                    mmproj_path = mmproj_path,
-                )
+                launch_mmproj_path = None
+                if not extra_args_disable_mmproj(extra_args):
+                    launch_mmproj_path = self._resolve_launch_mmproj_path(
+                        model_path = model_path,
+                        mmproj_path = mmproj_path,
+                    )
                 # Need both a resolved mmproj AND the config vision flag; a stray
                 # mmproj passing the family-name heuristic must not flip a non-VLM
                 # GGUF into vision mode.

studio/backend/core/inference/llama_server_args.py

@@ -260,6 +260,28 @@ def resolve_cache_type_kv(
     return override if override is not None else fallback_cache_type_kv
 
 
+_MMPROJ_DISABLE_FLAGS: frozenset[str] = frozenset({"--no-mmproj", "--no-mmproj-auto"})
+_MMPROJ_ENABLE_FLAGS: frozenset[str] = frozenset({"--mmproj-auto"})
+
+
+def extra_args_disable_mmproj(args: Optional[Iterable[str]]) -> bool:
+    """True when pass-through args opt out of vision mmproj loading.
+
+    llama-server parses --mmproj-auto / --no-mmproj / --no-mmproj-auto as one
+    boolean with last-wins semantics; mirror that here.
+    """
+    if not args:
+        return False
+    disabled = False
+    for raw in args:
+        flag = _flag_name(str(raw))
+        if flag in _MMPROJ_DISABLE_FLAGS:
+            disabled = True
+        elif flag in _MMPROJ_ENABLE_FLAGS:
+            disabled = False
+    return disabled
+
+
 def strip_shadowing_flags(
     args: Iterable[str],
     *,

studio/backend/routes/inference.py

@@ -1449,18 +1449,23 @@ async def load_model(
             # parse against a freshly-supplied first-class field.
             if request.llama_extra_args is None and llama_backend.extra_args:
                 source = llama_backend.extra_args_source
-                # Compare against the resolved variant, not the request field:
-                # callers commonly omit gguf_variant for local ``.gguf`` paths
-                # and HF auto-pick flows. ``config.gguf_variant`` is the variant
-                # load_model was actually invoked with (see HF / local branches
-                # below), so both sides key off the same string.
-                resolved_variant = config.gguf_variant
-                same_source = bool(
-                    source
-                    and source[0]
-                    and source[0].lower() == model_identifier.lower()
-                    and (source[1] or "").lower() == (resolved_variant or "").lower()
+                # Compare against the resolved variant, not the request
+                # field: callers commonly omit gguf_variant for local
+                # ``.gguf`` paths and HF auto-pick flows. ``config.gguf_
+                # variant`` is the variant load_model was actually
+                # invoked with (see the HF / local branches below), so
+                # both sides of the comparison key off the same string.
+                resolved_variant = (config.gguf_variant or "").lower()
+                request_variant = (request.gguf_variant or "").lower()
+                stored_variant = (source[1] or "").lower() if source else ""
+                same_model = bool(
+                    source and source[0] and source[0].lower() == model_identifier.lower()
                 )
+                if request.gguf_variant:
+                    variant_mismatch = request_variant != stored_variant
+                else:
+                    variant_mismatch = bool(stored_variant and resolved_variant != stored_variant)
+                same_source = same_model and not variant_mismatch
                 if not same_source:
                     logger.info(
                         "Not inheriting llama_extra_args: stored args came from %s, loading %s",

studio/backend/tests/test_llama_server_args.py

@@ -27,6 +27,7 @@ parse_cache_override = _lsa.parse_cache_override
 parse_ctx_override = _lsa.parse_ctx_override
 resolve_cache_type_kv = _lsa.resolve_cache_type_kv
 strip_shadowing_flags = _lsa.strip_shadowing_flags
+extra_args_disable_mmproj = _lsa.extra_args_disable_mmproj
 validate_extra_args = _lsa.validate_extra_args
 
 
@@ -509,6 +510,22 @@ def test_strip_shadowing_flags_defaults_strip_everything():
     assert out == []
 
 
+def test_extra_args_disable_mmproj_detects_flag():
+    assert extra_args_disable_mmproj(["--no-mmproj"]) is True
+    assert extra_args_disable_mmproj(["--threads", "12", "--no-mmproj"]) is True
+    assert extra_args_disable_mmproj(["--no-mmproj-auto"]) is True
+
+
+def test_extra_args_disable_mmproj_false_when_absent():
+    assert extra_args_disable_mmproj(None) is False
+    assert extra_args_disable_mmproj(["--threads", "12"]) is False
+
+
+def test_extra_args_disable_mmproj_last_wins():
+    assert extra_args_disable_mmproj(["--no-mmproj", "--mmproj-auto"]) is False
+    assert extra_args_disable_mmproj(["--mmproj-auto", "--no-mmproj-auto"]) is True
+
+
 def test_strip_shadowing_flags_drops_model_draft_with_spec():
     # --model-draft (and aliases) are Studio-managed since the separate
     # MTP drafter support: an inherited copy must not last-wins-override