From 6915c3091121df190e7968f0012f4b0bb187c0a7 Mon Sep 17 00:00:00 2001
From: claudio <claudio@hubris.local>
Date: Wed, 13 May 2026 23:08:10 +0200
Subject: [PATCH] build(backend): cache torch wheel + pip downloads across
 rebuilds

Splits torch+torchvision into its own RUN layer so edits to
requirements.txt don't invalidate the ~200MB CPU-only torch download.
Adds buildkit cache mounts on both pip-install layers so even when a
layer is invalidated (or buildkit evicts it) the wheel is reused from
the on-disk pip cache instead of refetching from download.pytorch.org.

Triggered by two consecutive deploy failures where pytorch.org timed
out mid-download (2026-05-13 ~21:41 and possibly ~22:47).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 backend/Dockerfile | 23 +++++++++++++++++------
 1 file changed, 17 insertions(+), 6 deletions(-)

diff --git a/backend/Dockerfile b/backend/Dockerfile
index 2a67133..b238c75 100644
--- a/backend/Dockerfile
+++ b/backend/Dockerfile
@@ -1,3 +1,4 @@
+# syntax=docker/dockerfile:1.7
 FROM python:3.12-slim
 
 # Install system dependencies
@@ -22,13 +23,23 @@ RUN apt-get update && apt-get install -y \
 
 WORKDIR /app
 
-# Copy requirements first for better caching
+# Install PyTorch CPU-only FIRST, in its own layer, so open-clip-torch
+# doesn't pull the full CUDA build (~7 GB). CPU inference is all we need
+# — the heavy lifting happens through ONNX Runtime.
+#
+# Two cache wins here:
+#   1. Its own RUN layer means edits to requirements.txt don't force a
+#      re-pull of the ~200MB torch wheel.
+#   2. The buildkit cache mount keeps pip's download cache on disk
+#      across builds even when the layer itself is invalidated, so a
+#      torch-version bump or a builder cache eviction still reuses the
+#      wheel from local cache instead of re-fetching from pytorch.org.
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
+
 COPY requirements.txt .
-# Install PyTorch CPU-only FIRST so open-clip-torch doesn't pull the full
-# CUDA build (~7 GB).  CPU inference is all we need — the heavy lifting
-# happens through ONNX Runtime.
-RUN pip install --no-cache-dir torch torchvision --index-url https://download.pytorch.org/whl/cpu \
-    && pip install --no-cache-dir -r requirements.txt
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install -r requirements.txt
 
 # Copy application code
 COPY . .