Merge pull request #22511 from arash77/docs/huggingface-data-table

doc: add huggingface shared data table schema

Author: mvdbeek

doc/source/admin/data_tables.md

@@ -103,3 +103,128 @@ When a new tool is installed that uses a data table a new entry is added to
 subdirectory in `tool_data_path` (in a subdirectory that has the name of the
 toolshed). By default this is `tool-data/toolshed.g2.bx.psu.edu/`. Note that
 these directories will also contain tool data table config files, but they are unused.
+
+## The `huggingface` shared data table
+
+Galaxy tools that consume pre-downloaded Hugging Face models share a single
+data table named `huggingface`.  Using one shared table means admins maintain
+one `.loc` file and all tools benefit from every registered model entry.
+
+### Declaring the table
+
+Add the following block to `tool_data_table_conf.xml`:
+
+```xml
+<!-- Hugging Face models -->
+<table name="huggingface" comment_char="#" allow_duplicate_entries="False">
+    <columns>value, name, pipeline_tag, domain, free_tag, version, path</columns>
+    <file path="/opt/galaxy/tool-data/huggingface.loc" />
+</table>
+```
+
+Each tool ships a `tool-data/huggingface.loc.sample` that uses the same
+7-column layout.
+
+### Column reference
+
+| # | Column | Purpose |
+|---|--------|---------|
+| 0 | `value` | Unique row ID across the whole table |
+| 1 | `name` | Human-readable label shown in the Galaxy select widget |
+| 2 | `pipeline_tag` | Model role — see controlled vocabulary below |
+| 3 | `domain` | Coarse data domain — see controlled vocabulary below |
+| 4 | `free_tag` | Optional narrowing tag; fallback filter when `pipeline_tag`/`domain` alone are not specific enough |
+| 5 | `version` | Model version string |
+| 6 | `path` | Path to the model data, a directory or a specific file, depending on the model structure |
+
+**`value` (column 0)**
+
+Must be globally unique across every row in `huggingface.loc`, regardless of
+which tool added it.  Use the Hugging Face model ID (`<owner>/<model-name>`)
+directly — it is stable and unambiguous.  If the same model is registered at
+more than one version, append the version:
+
+```
+black-forest-labs/FLUX.1-dev
+sentence-transformers/all-MiniLM-L6-v2
+openai/whisper-large-v3_3.0
+```
+
+**`pipeline_tag` (column 2)**
+
+Use the official [Hugging Face pipeline tag](https://huggingface.co/models).
+Common values:
+
+| Value | When to use |
+|-------|-------------|
+| `text-to-image` | Image generation models |
+| `automatic-speech-recognition` | ASR / transcription models |
+| `feature-extraction` | Sentence / document embedding models |
+| `tabular-classification` | Tabular ML classifiers |
+| `tabular-regression` | Tabular ML regressors |
+| `text-generation` | Causal / instruction-tuned LLMs |
+
+Do not invent synonyms for existing Hugging Face tags.
+
+**`domain` (column 3)**
+
+A broad category for the data type the model works with:
+`image` · `text` · `audio` · `tabular` · `sequence` · `video` · `multimodal`
+
+**`free_tag` (column 4)**
+
+An optional short identifier used as a fallback narrowing filter when
+`pipeline_tag` and `domain` alone are not specific enough.  Because a model
+can be consumed by multiple tools, `free_tag` must not encode a specific tool
+name.  Choose a short, lowercase, descriptive value and document it alongside
+the tool that introduces it.
+
+**`version` (column 5)**
+
+The model version string.  A tool declares in its XML which version(s) it
+accepts, allowing multiple versions of the same model to coexist.  Where
+possible, rows are only added, never removed or edited.
+
+**`path` (column 6)**
+
+The path to the model data on the production server (maintained by admins).
+Can be a directory (when the tool reads the whole Hugging Face cache layout)
+or a specific file (e.g. a `.ckpt` checkpoint).
+
+### XML filter convention
+
+Filter primarily by `pipeline_tag` (column 2) and/or `domain` (column 3) so
+only relevant model types are shown to the user.  Add a `version` or
+`free_tag` filter only when you need to narrow the selection further:
+
+```xml
+<param name="model" type="select" label="Model">
+    <options from_data_table="huggingface">
+        <filter type="static_value" column="2" value="<pipeline_tag>"/>
+        <filter type="static_value" column="3" value="<domain>"/>
+        <!-- optional: narrow further by version or free_tag -->
+        <!-- <filter type="static_value" column="5" value="<version>"/> -->
+        <!-- <filter type="static_value" column="4" value="<free_tag>"/> -->
+    </options>
+</param>
+```
+
+Example from the Flux tool (filters by `free_tag` to restrict to Flux-specific model variants):
+
+```xml
+<options from_data_table="huggingface">
+    <filter type="static_value" column="4" value="flux"/>
+    <filter type="static_value" column="5" value="1"/>
+</options>
+```
+
+### Example `.loc` entry
+
+Each row is TAB-separated (7 columns):
+
+```
+# Columns: value <TAB> name <TAB> pipeline_tag <TAB> domain <TAB> free_tag <TAB> version <TAB> path
+#
+# Flux
+black-forest-labs/FLUX.1-dev	FLUX.1 [dev]	text-to-image	image	flux	1	/data/hf_models
+```