From 6123b70b5c28215c6beb3cd227e72b6761fb0b5b Mon Sep 17 00:00:00 2001
From: dxrpl <gpudlowskidxr@gmail.com>
Date: Thu, 15 Oct 2026 01:15:08 +0200
Subject: [PATCH] fix(hwfit): keep non-LLM HF collection models out of the LLM
 catalog (#6731)

Tracked HF collections also contain time-series forecasters, image
encoders and image/video generators. They were imported as "general"
LLMs and offered for vllm serve. Skip their pipeline tags; embedding,
reranker and RL-tagged LLM repos stay.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 services/hwfit/hf_discovery.py                | 16 ++++++++++
 ...test_hwfit_hf_discovery_pipeline_filter.py | 31 +++++++++++++++++++
 2 files changed, 47 insertions(+)
 create mode 100644 tests/test_hwfit_hf_discovery_pipeline_filter.py

diff --git a/services/hwfit/hf_discovery.py b/services/hwfit/hf_discovery.py
index 3bea44931..143245b95 100644
--- a/services/hwfit/hf_discovery.py
+++ b/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
diff --git a/tests/test_hwfit_hf_discovery_pipeline_filter.py b/tests/test_hwfit_hf_discovery_pipeline_filter.py
new file mode 100644
index 000000000..f1b17468e
--- /dev/null
+++ b/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