fix(llamacpp): resolve --model-dir to the top-level GGUF, not a nested draft

find_gguf globbed **/*.gguf recursively and returned the lexically-first match, so a
--model-dir pointing at a directory that also holds the draft/MTP GGUF in a subdirectory
resolved the model to the draft (e.g. an object-store FUSE prefix carrying both
Model.gguf and MTP/draft.gguf -- "MTP" sorts before the model name). Download mode never
hit this because model and draft are separate artifacts in separate mount dirs; fuse
delivery surfaces them under one prefix. Prefer GGUFs directly in the given directory and
only recurse when none are found there; sharded-shard and nested-only layouts are
unaffected.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Signed-off-by: Michael Hotan <mike@union.ai>

Author: mhotan

plugins/llamacpp/src/flyteplugins/llamacpp/_server.py

@@ -32,12 +32,18 @@
 def find_gguf(path: str) -> str:
     """Resolve the GGUF file to serve from a mounted file or directory.
 
-    For sharded models only the first shard is passed to llama-server (it discovers
+    GGUFs directly in `path` win over any in subdirectories: a `--model-dir` may point at a
+    mount that also holds the draft/MTP GGUF in a *subdirectory* (e.g. an object-store FUSE
+    prefix carrying both `Model.gguf` and `MTP/draft.gguf`), and a recursive match could
+    otherwise resolve the model to the draft. Only if no GGUF sits directly in `path` do we
+    recurse. For sharded models only the first shard is passed to llama-server (it discovers
     the rest itself), so `*-00001-of-*.gguf` wins over other matches.
     """
     if os.path.isfile(path):
         return path
-    matches = sorted(glob.glob(os.path.join(path, "**", "*.gguf"), recursive=True))
+    matches = sorted(glob.glob(os.path.join(path, "*.gguf")))
+    if not matches:
+        matches = sorted(glob.glob(os.path.join(path, "**", "*.gguf"), recursive=True))
     if not matches:
         raise FileNotFoundError(f"No .gguf files found under {path!r}")
     first_shards = [m for m in matches if "-00001-of-" in Path(m).name]

plugins/llamacpp/tests/test_server_shim.py

@@ -31,6 +31,15 @@ def test_find_gguf_nested(tmp_path):
     assert find_gguf(str(tmp_path)) == gguf
 
 
+def test_find_gguf_prefers_top_level_over_subdir_draft(tmp_path):
+    """A model-dir that also holds the draft/MTP GGUF in a subdirectory (the object-store
+    FUSE layout) must resolve to the top-level model, not the nested draft -- even though the
+    draft's path sorts first."""
+    model = _touch(tmp_path / "Qwen3-27B-Q4_K_M.gguf")
+    _touch(tmp_path / "MTP" / "mtp-draft-Q4_0.gguf")
+    assert find_gguf(str(tmp_path)) == model
+
+
 def test_find_gguf_prefers_first_shard(tmp_path):
     # "a-..." sorts before the shard files; the first shard must still win.
     _touch(tmp_path / "a-mmproj.gguf")