Merge pull request #6732 from dxrpl/fix/hwfit-skip-non-llm-collection-models

fix(hwfit): keep non-LLM HF collection models out of the LLM catalog (#6731)

services/hwfit/hf_discovery.py

@@ -202,10 +202,26 @@ def _infer_use_case(repo_id, pipeline_tag):
     return "general"
 
 
+# HF collections mix in models no Cookbook LLM backend can serve (vision
+# encoders, time-series forecasters, image/video generators). Image models
+# have their own registry, so keep these out of the LLM catalog.
+_NON_LLM_PIPELINE_TAGS = frozenset({
+    "time-series-forecasting",
+    "zero-shot-image-classification",
+    "image-classification",
+    "image-feature-extraction",
+    "text-to-image",
+    "image-to-video",
+    "text-to-video",
+})
+
+
 def _entry_from_collection_item(collection, item, source):
     repo_id = item.get("id") or ""
     if item.get("type") != "model" or not repo_id:
         return None
+    if (item.get("pipeline_tag") or "").strip().lower() in _NON_LLM_PIPELINE_TAGS:
+        return None
     repo_prefix = source.get("repo_prefix")
     if repo_prefix and not repo_id.startswith(repo_prefix):
         return None

tests/test_hwfit_hf_discovery_pipeline_filter.py

@@ -0,0 +1,31 @@
+import pytest
+
+from services.hwfit.hf_discovery import _entry_from_collection_item
+
+_SOURCE = {"owner": "google", "provider": "Google"}
+_COLLECTION = {"title": "collection"}
+
+
+@pytest.mark.parametrize("repo_id,pipeline_tag", [
+    ("google/timesfm-3.0-pytorch", "time-series-forecasting"),
+    ("google/siglip2-base-patch16-224", "zero-shot-image-classification"),
+    ("moonshotai/MoonViT-SO-400M", "image-feature-extraction"),
+    ("zai-org/CogView4-6B", "text-to-image"),
+    ("zai-org/CogVideoX-5b-I2V", "image-to-video"),
+    ("zai-org/CogVideoX1.5-5B", "text-to-video"),
+])
+def test_non_llm_collection_models_are_skipped(repo_id, pipeline_tag):
+    item = {"id": repo_id, "type": "model", "pipeline_tag": pipeline_tag, "numParameters": 1_000_000_000}
+    assert _entry_from_collection_item(_COLLECTION, item, _SOURCE) is None
+
+
+@pytest.mark.parametrize("repo_id,pipeline_tag", [
+    ("google/gemma-4-12B-it", "image-text-to-text"),
+    ("Qwen/Qwen3.6-27B", "text-generation"),
+    ("Qwen/Qwen3-Embedding-4B", "feature-extraction"),
+    ("NousResearch/DeepHermes-Egregore-v2-RLAIF-8b-Atropos", "reinforcement-learning"),
+    ("Qwen/Qwen3-Next-80B-A3B-Instruct", ""),
+])
+def test_servable_collection_models_are_kept(repo_id, pipeline_tag):
+    item = {"id": repo_id, "type": "model", "pipeline_tag": pipeline_tag, "numParameters": 4_000_000_000}
+    assert _entry_from_collection_item(_COLLECTION, item, _SOURCE) is not None