feat(downloads): allow HF_MODEL_ID call-arg (defauls to MODEL_ID)

Sometimes you might want the unique model_id you use to differ
from the HuggingFace MODEL_ID (user/repo), e.g. maybe the model
is constantly updated and you want to create new unique local
models from each update.

TODO, separate MODEL_REVISION and MODEL_PRECISION.  Allow
HF_MODEL_ID env variable for builds.

Author: gadicc

api/app.py

@@ -151,6 +151,7 @@ def inference(all_inputs: dict) -> dict:
     normalized_model_id = model_id
 
     if RUNTIME_DOWNLOADS:
+        hf_model_id = call_inputs.get("HF_MODEL_ID", None)
         model_precision = call_inputs.get("MODEL_PRECISION", None)
         checkpoint_url = call_inputs.get("CHECKPOINT_URL", None)
         checkpoint_config_url = call_inputs.get("CHECKPOINT_CONFIG_URL", None)
@@ -173,6 +174,7 @@ def inference(all_inputs: dict) -> dict:
                     model_revision=model_precision,
                     checkpoint_url=checkpoint_url,
                     checkpoint_config_url=checkpoint_config_url,
+                    hf_model_id=hf_model_id,
                 )
                 # downloaded_models.update({normalized_model_id: True})
             clearPipelines()

api/download.py

@@ -43,17 +43,19 @@ def download_model(
     model_revision=None,
     checkpoint_url=None,
     checkpoint_config_url=None,
+    hf_model_id=None,
 ):
     print(
         "download_model",
         {
             "model_url": model_url,
             "model_id": model_id,
             "model_revision": model_revision,
+            "hf_model_id": hf_model_id,
         },
     )
-    id = model_id or MODEL_ID
     url = model_url or MODEL_URL
+    hf_model_id = hf_model_id or model_id
     revision = model_revision or revision_from_precision()
     normalized_model_id = id
 
@@ -100,11 +102,11 @@ def download_model(
                 # this conveniently logs all the timings (and doesn't happen often)
                 print("download")
                 send("download", "start", {})
-                model = loadModel(model_id, False, precision=model_revision)  # download
+                model = loadModel(hf_model_id, False, precision=model_revision)  # download
                 send("download", "done", {})
 
             print("load")
-            model = loadModel(model_id, True, precision=model_revision)  # load
+            model = loadModel(hf_model_id, True, precision=model_revision)  # load
             # dir = "models--" + model_id.replace("/", "--") + "--dda"
             dir = os.path.join(MODELS_DIR, normalized_model_id)
             model.save_pretrained(dir, safe_serialization=True)