Merge pull request #1841 from transformerlab/fix/local-model-vllm

Use tlab registry models with interactive tasks

Author: deep1401

api/localprovider_pyproject.toml

@@ -32,7 +32,7 @@ dependencies = [
   "soundfile==0.13.1",
   "tensorboardX==2.6.2.2",
   "timm==1.0.15",
-  "transformerlab==0.1.26",
+  "transformerlab==0.1.27",
   "transformerlab-inference==0.2.52",
   "transformers==4.57.1",
   "wandb==0.23.1",

api/pyproject.toml

@@ -24,7 +24,7 @@ dependencies = [
   "python-dotenv==1.1.0",
   "python-multipart==0.0.20",
   "sqlalchemy[asyncio]==2.0.40",
-  "transformerlab==0.1.26",
+  "transformerlab==0.1.27",
   "transformerlab-inference==0.2.52",
   "uvicorn==0.35.0",
   "watchfiles==1.0.5",

api/transformerlab/galleries/channels/stable/latest/interactive-gallery.json

@@ -104,7 +104,7 @@
       {
         "field_name": "Model Name",
         "env_var": "MODEL_NAME",
-        "field_type": "str",
+        "field_type": "model",
         "required": true,
         "placeholder": "e.g. meta-llama/Llama-2-7b-chat-hf",
         "help_text": "HuggingFace model identifier"

api/transformerlab/galleries/channels/stable/latest/manifest.json

@@ -1,7 +1,7 @@
 {
-  "bundle_version": "20260421193056",
+  "bundle_version": "20260424160054",
   "channel": "stable",
-  "released_at": "2026-04-21T19:30:56.442725+00:00",
+  "released_at": "2026-04-24T16:00:54.223569+00:00",
   "files": {
     "task-gallery.json": 25,
     "interactive-gallery.json": 9,

api/transformerlab/galleries/examples/interactive-vllm/run.py

@@ -9,6 +9,8 @@
 import time
 import urllib.request
 import urllib.error
+from lab import lab
+
 
 LOG_FILES = [
     pathlib.Path("/tmp/vllm.log"),
@@ -121,6 +123,12 @@ def main() -> None:
     tp_size = os.environ.get("TP_SIZE", "1")
     env = os.environ.copy()
 
+    # If a TLab registry model was selected, download it and use the local path.
+    if os.environ.get("MODEL_REGISTRY_TLAB") == "1":
+        print(f"Downloading registry model '{model_name}'...", flush=True)
+        model_name = lab.download_registry_model(model_name)
+        print(f"Model downloaded to: {model_name}", flush=True)
+
     _touch_logs()
 
     vllm_python = os.path.expanduser("~/vllm-venv/bin/python")

api/transformerlab/galleries/src/interactive/interactive.json

@@ -104,7 +104,7 @@
       {
         "field_name": "Model Name",
         "env_var": "MODEL_NAME",
-        "field_type": "str",
+        "field_type": "model",
         "required": true,
         "placeholder": "e.g. meta-llama/Llama-2-7b-chat-hf",
         "help_text": "HuggingFace model identifier"

api/transformerlab/routers/model.py

@@ -38,6 +38,45 @@ async def model_local_list():
     return models
 
 
+@router.get("/model/registry_versions")
+@cached(key="models:registry_versions", ttl="1h", tags=["models", "models:registry_versions", "asset_versions"])
+async def model_registry_versions():
+    """List model registry versions as selectable model IDs."""
+    from transformerlab.services import asset_version_service
+
+    groups = await asset_version_service.list_groups("model")
+    versions: list[dict] = []
+
+    for group in groups:
+        group_id = group.get("group_id")
+        group_name = group.get("group_name") or group_id
+        if not group_id:
+            continue
+
+        group_versions = await asset_version_service.list_versions("model", group_id)
+        for version in group_versions:
+            model_id = version.get("asset_id")
+            version_label = version.get("version_label")
+            if not model_id or not version_label:
+                continue
+
+            tag = version.get("tag")
+            tag_suffix = f" ({tag})" if tag else ""
+            versions.append(
+                {
+                    "model_id": model_id,
+                    "name": f"{group_name}/{version_label}{tag_suffix}",
+                    "group_id": group_id,
+                    "group_name": group_name,
+                    "version_label": version_label,
+                    "tag": tag,
+                    "created_at": version.get("created_at"),
+                }
+            )
+
+    return versions
+
+
 @router.get("/model/create")
 async def model_local_create(id: str, name: str, json_data={}):
     # Use filesystem instead of database

lab-sdk/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "transformerlab"
-version = "0.1.26"
+version = "0.1.27"
 description = "Python SDK for Transformer Lab"
 readme = "README.md"
 requires-python = ">=3.10"

lab-sdk/src/lab/lab_facade.py

@@ -648,6 +648,18 @@ def finish(
         # Important: update cached_jobs only when all completion fields are already written.
         _run_async(self._job.update_status(JobStatus.COMPLETE))  # type: ignore[union-attr]
 
+    def download_registry_model(self, model_id: str) -> str:
+        """Download a TLab registry model to ~/tmp/<model_id> and return the local path."""
+        return _run_async(self.async_download_registry_model(model_id))
+
+    async def async_download_registry_model(self, model_id: str) -> str:
+        """Download model from TLab registry storage to ~/tmp/<model_id> and return the local path."""
+        model = await ModelService.get(model_id)
+        remote_dir = await model.get_dir()
+        local_path = os.path.expanduser(f"~/tmp/{model_id}")
+        await storage.copy_dir(remote_dir, local_path)
+        return local_path
+
     def save_artifact(
         self,
         source_path: Union[str, Any],

lab-sdk/tests/test_lab_facade.py

@@ -1244,3 +1244,43 @@ async def test_copy_file_mounts_without_lab_init(tmp_path, monkeypatch):
     assert os.path.isfile(dest)
     with open(dest) as f:
         assert f.read() == "hi"
+
+
+def test_download_registry_model(tmp_path, monkeypatch):
+    _fresh(monkeypatch)
+    home = tmp_path / ".tfl_home"
+    ws = tmp_path / ".tfl_ws"
+    home.mkdir()
+    ws.mkdir()
+    monkeypatch.setenv("TFL_HOME_DIR", str(home))
+    monkeypatch.setenv("TFL_WORKSPACE_DIR", str(ws))
+
+    import asyncio
+    import os
+    import lab.storage as storage_module
+    from lab.lab_facade import Lab
+    from lab.model import Model
+
+    model_id = "test_registry_model"
+    asyncio.run(Model.create(model_id))
+
+    lab = Lab()
+    lab.init(experiment_id="test_exp")
+
+    copy_calls = []
+
+    async def fake_copy_dir(src: str, dest: str) -> None:
+        copy_calls.append((src, dest))
+
+    monkeypatch.setattr(storage_module, "copy_dir", fake_copy_dir)
+
+    result = lab.download_registry_model(model_id)
+
+    expected_dest = os.path.expanduser(f"~/tmp/{model_id}")
+    assert result == expected_dest
+    assert len(copy_calls) == 1
+
+    _m = asyncio.run(Model.get(model_id))
+    expected_src = asyncio.run(_m.get_dir())
+    assert copy_calls[0][0] == expected_src
+    assert copy_calls[0][1] == expected_dest