Pin dependencies to a verified stack; align container to Python 3.12

.gitignore

@@ -9,3 +9,6 @@ runs/*
 *.pt
 __pycache__/
 __pycache__/*
+
+# Local dependency-verification artifact (CPU freeze); not for the repo
+.verified-freeze-cpu.txt

Dockerfile

@@ -1,14 +1,16 @@
 # no-code-deeplearning-prod/Dockerfile
 
-# Use NVIDIA's official CUDA base image
-FROM nvidia/cuda:12.8.1-cudnn-runtime-ubuntu22.04
+# NVIDIA CUDA base. Ubuntu 24.04 ships Python 3.12, which the pinned stack in
+# requirements.txt requires (numpy 2.4 / pandas 3.0 need Python >= 3.11).
+FROM nvidia/cuda:12.8.1-cudnn-runtime-ubuntu24.04
 
 ENV DEBIAN_FRONTEND=noninteractive
 
-# Install Python 3.10 and pip
+# Python 3.12 + the OpenCV/rendering shared libs pulled in by opencv/albumentations.
 RUN apt-get update && \
     apt-get install -y \
-    python3.10 \
+    python3.12 \
+    python3.12-venv \
     python3-pip \
     libgl1 \
     libglib2.0-0 \
@@ -18,16 +20,21 @@ RUN apt-get update && \
     libfontconfig1 \
     && rm -rf /var/lib/apt/lists/*
 
-RUN ln -sf /usr/bin/python3.10 /usr/bin/python && \
-    ln -sf /usr/bin/pip3 /usr/bin/pip
+# Use an isolated virtualenv. On Ubuntu 24.04 the system Python is
+# externally-managed (PEP 668), so installing into it needs a venv (or
+# --break-system-packages); a venv is cleaner and keeps the image reproducible.
+RUN python3.12 -m venv /opt/venv
+ENV PATH="/opt/venv/bin:$PATH"
 
 WORKDIR /app
 
 # Copy requirements first for layer caching
 COPY requirements.txt .
 
-# Install Python dependencies (with CUDA torch)
-RUN pip install --no-cache-dir -r requirements.txt
+# Install the pinned dependencies. On this CUDA base image, PyPI serves the
+# CUDA build of the pinned torch version.
+RUN pip install --no-cache-dir --upgrade pip && \
+    pip install --no-cache-dir -r requirements.txt
 
 # Copy the entire DL service codebase
-COPY . .
+COPY . .

requirements.txt

@@ -1,40 +1,54 @@
+# Pinned to a verified-working stack.
+#
+# Every version below was confirmed by running the full train -> evaluate ->
+# save -> inference cycle for all three tasks and both model families per task
+# (image classification, object detection via DETR and YOLO, semantic
+# segmentation via SegFormer and SMP U-Net) on Python 3.12 / CPU torch.
+#
+# Torch is pinned by version only (no +cpu local tag): on the CUDA base image in
+# the Dockerfile, PyPI serves the matching CUDA build of the same version.
+#
+# To refresh: bump a pin, re-run the smoke tests for all tasks, then commit.
+
 # --- Core ML/DL ---
-torch
-torchvision
-torchaudio
-accelerate
+torch==2.13.0
+torchvision==0.28.0
+torchaudio==2.11.0
+accelerate==1.14.0
 
 # --- Hugging Face ---
-transformers
-datasets
-evaluate
-huggingface_hub
+transformers==5.13.1
+datasets==5.0.0
+evaluate==0.4.6
+huggingface_hub==1.23.0
 
 # --- Task-Specific ---
-albumentations
-torchmetrics
-timm
-pycocotools
-ultralytics
-faster-coco-eval
-segmentation-models-pytorch
+albumentations==2.0.8
+torchmetrics==1.9.0
+timm==1.0.28
+pycocotools==2.0.11
+ultralytics==8.4.92
+faster-coco-eval==1.7.2
+segmentation-models-pytorch==0.5.0
 
-# --- API & UI ---
-fastapi
-uvicorn[standard]
-python-multipart
-gradio
+# --- API ---
+fastapi==0.139.0
+uvicorn[standard]==0.51.0
+python-multipart==0.0.32
 
 # --- Utilities ---
-numpy
-pandas
-scipy
-Pillow
-wandb
-psutil
-scikit-learn
+numpy==2.4.4
+pandas==3.0.3
+scipy==1.18.0
+Pillow==12.2.0
+psutil==7.2.2
+scikit-learn==1.9.0
 
 # --- Job Persistence & Queuing ---
-sqlalchemy
-celery
-redis
+sqlalchemy==2.0.51
+celery==5.6.3
+redis==8.0.1
+
+# Removed (2026-07): gradio and wandb — neither is imported anywhere in the
+# codebase (training sets WANDB_DISABLED=true), and gradio pulls a large,
+# unused dependency tree. Re-add with a pin if a demo UI / W&B logging is built.