fix: added

README.md

@@ -10,4 +10,4 @@ pinned: false
 
 Image captioning API using `microsoft/Florence-2-base` with a Python FastAPI backend. Open `/docs` for Swagger UI.
 
-Speed tuning env vars: `DEFAULT_MAX_TOKENS` (default `64`), `MAX_IMAGE_SIDE` (default `896`), `MAX_MAX_TOKENS` (default `256`), `MODEL_ID` (default `microsoft/Florence-2-base`), `MODEL_REVISION` (optional commit SHA to pin remote model code).
+Speed tuning env vars: `DEFAULT_MAX_TOKENS` (default `64`), `MAX_IMAGE_SIDE` (default `896`), `MAX_MAX_TOKENS` (default `256`), `MODEL_ID` (default `microsoft/Florence-2-base`), `MODEL_REVISION` (pin to a commit SHA, e.g. `5ca5edf5bd017b9919c05d08aebef5e4c7ac3bac`).

docker-compose.yml

@@ -9,6 +9,5 @@ services:
       - MAX_IMAGE_SIDE=896
       - MAX_MAX_TOKENS=256
       - MODEL_ID=microsoft/Florence-2-base
-      # Optional: pin to a specific commit SHA from huggingface.co/microsoft/Florence-2-base
-      # - MODEL_REVISION=<commit_sha>
+      - MODEL_REVISION=5ca5edf5bd017b9919c05d08aebef5e4c7ac3bac
     restart: unless-stopped

requirements.txt

@@ -1,6 +1,6 @@
 fastapi==0.116.1
 uvicorn[standard]==0.35.0
-transformers==4.55.4
+transformers==4.49.0
 torch==2.8.0
 pillow==11.3.0
 python-multipart==0.0.20

src/model.py

@@ -69,6 +69,7 @@ def load_model() -> tuple[Any, Any]:
         _model = AutoModelForCausalLM.from_pretrained(
             MODEL_ID,
             torch_dtype=_dtype,
+            attn_implementation="eager",
             **pretrained_kwargs,
         ).to(_device)
         _model.eval()