step 4: endpoint-ready server (Dockerfile)

Dockerfile

@@ -1,12 +1,18 @@
 # MolmoAct2-SO100_101 cloud-inference server as a HuggingFace **Docker** Space.
 #
-# WHY A DOCKER SPACE (not an Inference Endpoint): the managed Inference-Endpoint
-# container bakes in `huggingface_inference_toolkit`, whose bootstrap imports
-# `transformers.file_utils.is_tf_available` — REMOVED in the transformers >=4.57
-# that MolmoAct2's processor requires (it needs `transformers.video_utils`). The
-# two are mutually exclusive, so the toolkit path is a dead end. A Docker Space
-# runs OUR uvicorn directly — the toolkit never enters the picture, and we own
-# the exact transformers version.
+# WHY A DOCKER SPACE (not a managed Inference Endpoint): the managed container
+# bakes in `huggingface_inference_toolkit`, which HARD-PINS transformers==4.51.3.
+# That predates `transformers.video_utils.VideoInput` (added in 4.57.0), which
+# MolmoAct2's trust_remote_code processor imports — so the managed image simply
+# cannot satisfy the model. A Docker Space runs OUR uvicorn directly: no toolkit,
+# and we own the exact transformers version.
+#
+# CORRECTION (2026-07-21): this comment previously blamed the toolkit importing
+# `transformers.file_utils.is_tf_available`, said to be "REMOVED in >=4.57". That
+# is false — is_tf_available is present through 4.57.3 and only disappears in
+# v5.0.0. The version floor below is right; only the reason was wrong. Note the
+# managed path may still be reachable with a CUSTOM container image, which
+# bypasses the toolkit entirely — untested here.
 #
 # Base image = the SAME torch stack proven on HF Jobs (torch 2.5.1 + cu121):
 # pinning transformers on top of it does NOT upgrade torch, so torchvision /
@@ -39,6 +45,12 @@ COPY --chown=user:user rtc.py .
 RUN mkdir -p /home/user/.cache/huggingface && chown -R user:user /home/user /app
 USER user
 
-# HF Spaces expose the container on 7860 (declared as app_port in README.md).
+# PORT: Spaces route to 7860 (declared as app_port in README.md). Inference
+# Endpoints take the container port from the endpoint config — set the endpoint's
+# port to 7860 too, or override with a PORT env var here. A port mismatch is the
+# documented #1 cause of an endpoint stuck "initializing", hence the explicit
+# declaration + env override. (Also set the endpoint's health_route=/ready — it
+# 503s until the model is loaded; "/" and /health stay 200-while-loading for the
+# Space's readiness probe.)
 EXPOSE 7860
-CMD ["uvicorn", "molmoact2_server:app", "--host", "0.0.0.0", "--port", "7860"]
+CMD ["sh", "-c", "uvicorn molmoact2_server:app --host 0.0.0.0 --port ${PORT:-7860}"]