feat: remove large models from model meta commands

Controlled by `AI_HORDE_MODEL_META_LARGE_MODELS`. Any non-null value will allow very large models, such as cascade and flux, to be returned from model meta commands.

horde_sdk/__init__.py

@@ -51,6 +51,12 @@ def _dev_env_var_warnings() -> None:  # pragma: no cover
         if len(dev_key) != 10 and len(dev_key) != 22:
             raise ValueError("AI_HORDE_DEV_APIKEY must be the anon key or 22 characters long.")
 
+    AI_HORDE_MODEL_META_LARGE_MODELS = os.getenv("AI_HORDE_MODEL_META_LARGE_MODELS")
+    if AI_HORDE_MODEL_META_LARGE_MODELS:
+        logger.debug(
+            f"AI_HORDE_MODEL_META_LARGE_MODELS is {AI_HORDE_MODEL_META_LARGE_MODELS}.",
+        )
+
 
 _dev_env_var_warnings()
 

horde_sdk/ai_horde_worker/model_meta.py

@@ -1,6 +1,7 @@
+import os
 import re
 
-from horde_model_reference.meta_consts import MODEL_REFERENCE_CATEGORY
+from horde_model_reference.meta_consts import MODEL_REFERENCE_CATEGORY, STABLE_DIFFUSION_BASELINE_CATEGORY
 from horde_model_reference.model_reference_manager import ModelReferenceManager
 from horde_model_reference.model_reference_records import StableDiffusion_ModelRecord
 from loguru import logger
@@ -128,7 +129,7 @@ def resolve_meta_instructions(
                 return_list.extend(self.resolve_all_nsfw_model_names())
 
         # If no valid meta instruction were found, return None
-        return set(return_list)
+        return self.remove_large_models(set(return_list))
 
     @staticmethod
     def meta_instruction_regex_match(instruction: str, target_string: str) -> re.Match[str] | None:
@@ -140,9 +141,22 @@ def meta_instruction_regex_match(instruction: str, target_string: str) -> re.Mat
 
         Returns:
             A Match object if the target string matches the regex pattern, otherwise None.
+
         """
         return re.match(instruction, target_string, re.IGNORECASE)
 
+    def remove_large_models(self, models: set[str]) -> set[str]:
+        """Remove large models from the input set of models."""
+        AI_HORDE_MODEL_META_LARGE_MODELS = os.getenv("AI_HORDE_MODEL_META_LARGE_MODELS")
+        if not AI_HORDE_MODEL_META_LARGE_MODELS:
+            cascade_models = self.resolve_all_models_of_baseline(STABLE_DIFFUSION_BASELINE_CATEGORY.stable_cascade)
+            flux_models = self.resolve_all_models_of_baseline(STABLE_DIFFUSION_BASELINE_CATEGORY.flux_1)
+
+            logger.debug(f"Removing cascade models: {cascade_models}")
+            logger.debug(f"Removing flux models: {flux_models}")
+            models = models - cascade_models - flux_models
+        return models
+
     def resolve_all_model_names(self) -> set[str]:
         """Get the names of all models defined in the model reference.
 
@@ -153,11 +167,13 @@ def resolve_all_model_names(self) -> set[str]:
 
         sd_model_references = all_model_references[MODEL_REFERENCE_CATEGORY.stable_diffusion]
 
-        if sd_model_references:
-            return set(sd_model_references.root.keys())
+        all_models = set(sd_model_references.root.keys()) if sd_model_references is not None else set()
 
-        logger.error("No stable diffusion models found in model reference.")
-        return set()
+        all_models = self.remove_large_models(all_models)
+
+        if not all_models:
+            logger.error("No stable diffusion models found in model reference.")
+        return all_models
 
     def _resolve_sfw_nsfw_model_names(self, nsfw: bool) -> set[str]:
         """Get the names of all SFW or NSFW models defined in the model reference.

tests/ai_horde_worker/test_model_meta_api_calls.py

@@ -29,6 +29,16 @@ def test_image_model_load_resolver_all(image_model_load_resolver: ImageModelLoad
 
     assert len(all_model_names) > 0
 
+    import os
+
+    os.environ["AI_HORDE_MODEL_META_LARGE_MODELS"] = "true"
+
+    all_model_names_with_large = image_model_load_resolver.resolve_all_model_names()
+
+    del os.environ["AI_HORDE_MODEL_META_LARGE_MODELS"]
+
+    assert len(all_model_names_with_large) > len(all_model_names)
+
 
 def test_image_model_load_resolver_top_n(
     image_model_load_resolver: ImageModelLoadResolver,
@@ -179,6 +189,19 @@ def test_image_models_unique_results_only(
 
     assert len(resolved_model_names) >= (len(all_model_names) - 1)  # FIXME: -1 is to account for SDXL beta
 
+    import os
+
+    os.environ["AI_HORDE_MODEL_META_LARGE_MODELS"] = "true"
+
+    resolved_models_names_with_large = image_model_load_resolver.resolve_meta_instructions(
+        ["top 1000", "bottom 1000"],
+        AIHordeAPIManualClient(),
+    )
+
+    del os.environ["AI_HORDE_MODEL_META_LARGE_MODELS"]
+
+    assert len(resolved_models_names_with_large) >= len(resolved_model_names)
+
 
 def test_resolve_all_models_of_baseline(
     image_model_load_resolver: ImageModelLoadResolver,