Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: stop hiding every model pulled with ollama pull (#10222)

Nilay04bce46a611e
diff --git a/studio/backend/hub/services/models/ollama.py b/studio/backend/hub/services/models/ollama.py
index 7edccb20fdc..648c549b24f 100644
--- a/studio/backend/hub/services/models/ollama.py
+++ b/studio/backend/hub/services/models/ollama.py
@@ -53,6 +53,36 @@ _OLLAMA_LOADABLE_LAYER_MEDIA_TYPES = frozenset(
}
)
+# Modelfile metadata nearly every pulled model carries. llama.cpp uses the GGUF's own template
+# and defaults instead, so none of it reaches the load and none of it should hide the row.
+# `image.adapter` is deliberately absent: it changes the weights.
+_OLLAMA_METADATA_LAYER_MEDIA_TYPES = frozenset(
+ {
+ "application/vnd.ollama.image.template",
+ "application/vnd.ollama.image.params",
+ "application/vnd.ollama.image.system",
+ "application/vnd.ollama.image.messages",
+ "application/vnd.ollama.image.prompt",
+ }
+)
+
+# Layers the load drops without changing what it returns. `image.draft` only feeds
+# speculative decoding, which is output-equivalent by construction, so losing it costs
+# speed and nothing else; `image.embed` has been deprecated since ollama 0.1.2 and
+# ollama's own load ignores it too.
+_OLLAMA_IGNORABLE_LAYER_MEDIA_TYPES = frozenset(
+ {
+ "application/vnd.ollama.image.draft",
+ "application/vnd.ollama.image.embed",
+ }
+)
+
+_OLLAMA_ADMITTED_LAYER_MEDIA_TYPES = (
+ _OLLAMA_LOADABLE_LAYER_MEDIA_TYPES
+ | _OLLAMA_METADATA_LAYER_MEDIA_TYPES
+ | _OLLAMA_IGNORABLE_LAYER_MEDIA_TYPES
+)
+
_OLLAMA_MATERIALIZE_LOCKS: dict[str, threading.Lock] = {}
_OLLAMA_MATERIALIZE_LOCKS_GUARD = threading.Lock()
@@ -79,7 +109,7 @@ def is_ollama_manifest_ref(ref: str) -> bool:
def _unsupported_ollama_layer_media_types(layers: list[object]) -> tuple[str, ...]:
- """Layer types whose behavior the direct llama.cpp load cannot preserve."""
+ """Layer types that would make the row describe something the load cannot deliver."""
unsupported: set[str] = set()
for layer in layers:
if not isinstance(layer, dict):
@@ -88,7 +118,7 @@ def _unsupported_ollama_layer_media_types(layers: list[object]) -> tuple[str, ..
media_type = layer.get("mediaType")
if not isinstance(media_type, str) or not media_type:
unsupported.add("<missing mediaType>")
- elif media_type not in _OLLAMA_LOADABLE_LAYER_MEDIA_TYPES:
+ elif media_type not in _OLLAMA_ADMITTED_LAYER_MEDIA_TYPES:
unsupported.add(media_type)
return tuple(sorted(unsupported))
diff --git a/studio/backend/tests/test_ollama_manifest_load_resolution.py b/studio/backend/tests/test_ollama_manifest_load_resolution.py
index cf87ee724d8..a46e0b030e6 100644
--- a/studio/backend/tests/test_ollama_manifest_load_resolution.py
+++ b/studio/backend/tests/test_ollama_manifest_load_resolution.py
@@ -540,22 +540,34 @@ def test_ollama_intent_loads_the_link_but_keeps_the_manifest_identity(tmp_path,
assert active_intent.gguf_path == resolved
-_UNSUPPORTED_RUNTIME_LAYERS = (
+# Modelfile metadata nearly every pulled model carries.
+_METADATA_LAYERS = (
"application/vnd.ollama.image.params",
"application/vnd.ollama.image.template",
"application/vnd.ollama.image.system",
"application/vnd.ollama.image.messages",
- "application/vnd.ollama.image.adapter",
"application/vnd.ollama.image.prompt",
+)
+
+_UNSUPPORTED_RUNTIME_LAYERS = (
+ "application/vnd.ollama.image.adapter",
"application/vnd.ollama.image.future-runtime",
)
+# Carried by a real manifest but dropped by the load without changing its output.
+_IGNORABLE_LAYERS = (
+ "application/vnd.ollama.image.draft",
+ "application/vnd.ollama.image.embed",
+)
+
def _rich_manifest_ref(tmp_path: Path, monkeypatch) -> tuple[Path, str]:
from hub.services.models import ollama
root = tmp_path / "ollama-rich"
- tag_file = _write_ollama_store(root, extra_layers = _UNSUPPORTED_RUNTIME_LAYERS)
+ tag_file = _write_ollama_store(
+ root, extra_layers = _METADATA_LAYERS + _UNSUPPORTED_RUNTIME_LAYERS
+ )
monkeypatch.setattr(ollama, "ollama_model_dirs", lambda: [root])
ref = f"ollama-manifest:{quote(str(tag_file), safe = '')}"
return root, ref
@@ -568,6 +580,65 @@ def test_rich_manifest_is_withheld_from_inventory(tmp_path, monkeypatch):
assert ollama.scan_ollama_dir(root) == []
+def test_a_normally_pulled_model_is_listed(tmp_path, monkeypatch):
+ from hub.services.models import ollama
+
+ root = tmp_path / "ollama-pulled"
+ _write_ollama_store(root, extra_layers = _METADATA_LAYERS)
+ monkeypatch.setattr(ollama, "ollama_model_dirs", lambda: [root])
+
+ rows = ollama.scan_ollama_dir(root)
+ assert len(rows) == 1
+ assert ollama.is_ollama_manifest_ref(rows[0].load_id)
+
+
+def test_a_normally_pulled_model_resolves_for_a_load(tmp_path, monkeypatch):
+ from hub.services.models import ollama
+ from models.inference import LoadRequest
+ from routes.inference import _resolve_model_identifier_for_request
+
+ root = tmp_path / "ollama-pulled-load"
+ tag_file = _write_ollama_store(root, extra_layers = _METADATA_LAYERS)
+ monkeypatch.setattr(ollama, "ollama_model_dirs", lambda: [root])
+ ref = f"ollama-manifest:{quote(str(tag_file), safe = '')}"
+
+ resolved, _link_name, _is_dir = _resolve_model_identifier_for_request(
+ LoadRequest(model_path = ref), operation = "load-model"
+ )
+ assert resolved.endswith(".gguf")
+
+
[email protected]("ignorable", _IGNORABLE_LAYERS)
+def test_an_ignorable_layer_does_not_hide_the_primary_model(tmp_path, monkeypatch, ignorable):
+ # Neither layer stops llama.cpp loading the model layer beside it, so neither is a
+ # reason to withhold the row: a draft model only accelerates speculative decoding,
+ # and ollama has ignored the embed layer since 0.1.2.
+ from hub.services.models import ollama
+
+ root = tmp_path / f"ollama-{ignorable.rpartition('.')[2]}"
+ _write_ollama_store(root, extra_layers = _METADATA_LAYERS + (ignorable,))
+ monkeypatch.setattr(ollama, "ollama_model_dirs", lambda: [root])
+
+ rows = ollama.scan_ollama_dir(root)
+ assert len(rows) == 1
+ assert ollama.is_ollama_manifest_ref(rows[0].load_id)
+
+
[email protected]("unsupported", _UNSUPPORTED_RUNTIME_LAYERS)
+def test_one_unsupported_layer_still_withholds_beside_the_metadata(
+ tmp_path, monkeypatch, unsupported
+):
+ # Pins each unsupported type on its own. The withholding test above carries both at
+ # once, so admitting just one of them back into the loadable set would still pass it.
+ from hub.services.models import ollama
+
+ root = tmp_path / f"ollama-{unsupported.rpartition('.')[2]}"
+ _write_ollama_store(root, extra_layers = _METADATA_LAYERS + (unsupported,))
+ monkeypatch.setattr(ollama, "ollama_model_dirs", lambda: [root])
+
+ assert ollama.scan_ollama_dir(root) == []
+
+
def test_rich_manifest_ref_is_rejected_without_creating_links(tmp_path, monkeypatch):
from models.inference import LoadRequest
from routes.inference import _resolve_model_identifier_for_request