public CUDA inference base image (image-estimate)

Dockerfile

@@ -0,0 +1,57 @@
+# syntax=docker/dockerfile:1.6
+# PUBLIC base image for the cheap CPU ESTIMATE job (HF Jobs) — slim, no GPU stack.
+#
+# Same contract as demo/image/Dockerfile: ONLY open-source, generic dependencies — NO proprietary
+# code, and no names that reveal which models/techniques the pipeline uses. The Job's bootstrap
+# (/opt/run-job.sh) pulls the private wheel + demo modules at startup and runs run_job.py.
+#
+# This variant is CPU-ONLY and slim: it omits CUDA/torch and the GPU inference servers (the estimate
+# runs zero signals and starts no model server — run_job.py returns at the ESTIMATE_ONLY branch
+# before anything loads), and it omits an optional dependency group that `import dataset_reviewer`
+# no longer pulls eagerly. Both shrink the image and its pull time, which is the bulk of the
+# estimate's wall-clock. A lightweight tokenizer loads via the Rust `tokenizers` backend, so no
+# torch is needed.
+#
+# Build: JOB_IMAGE_DIR=image-estimate JOB_IMAGE_BASE_PROVIDES=torch \
+#        JOB_IMAGE_SPACE=<owner>/cpu-estimate-base make demo-deploy-job
+# (run-job.sh + the dataset_reviewer stub are shared from demo/image/; the omitted dep group is
+#  applied automatically for this variant — see demo/deploy_job.py.)
+
+FROM python:3.12-slim
+
+ENV DEBIAN_FRONTEND=noninteractive \
+    PIP_NO_CACHE_DIR=0 \
+    PYTHONUNBUFFERED=1 \
+    HF_HUB_ENABLE_HF_TRANSFER=1 \
+    PIP_BREAK_SYSTEM_PACKAGES=1
+
+# git/curl/ca-certificates for the HF pulls at bootstrap. python3.12 + pip are in the base.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        git curl ca-certificates \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+# Install the generic DEPENDENCIES only — the generated pyproject omits torch
+# (JOB_IMAGE_BASE_PROVIDES=torch) since the estimate needs no torch/CUDA. A stub package
+# (empty dataset_reviewer/__init__.py) lets `pip install .` resolve the deps WITHOUT shipping
+# proprietary source into this public image. At runtime the Job replaces the stub with the
+# real wheel; run-job.sh skips the technique-revealing runtime deps for ESTIMATE_ONLY (never
+# imported on this path).
+COPY pyproject.toml /app/pyproject.toml
+COPY dataset_reviewer /app/dataset_reviewer
+RUN --mount=type=cache,target=/root/.cache/pip pip install /app
+
+# Explicit pins for the bootstrap + demo modules (transitive above, pinned so the bootstrap
+# never depends on resolution order). telethon is the Telegram client the Job imports — not a
+# dataset_reviewer dep.
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install "huggingface_hub[hf_transfer]" python-dotenv "telethon==1.43.2"
+
+COPY run-job.sh /opt/run-job.sh
+RUN chmod +x /opt/run-job.sh
+
+# Default command keeps the Space itself idle+RUNNING (so the image publishes cleanly);
+# HF Jobs override this with `bash /opt/run-job.sh`.
+EXPOSE 7860
+CMD ["python3", "-m", "http.server", "7860"]

README.md

@@ -1,10 +1,18 @@
 ---
-title: Cpu Estimate Base
-emoji: 👁
-colorFrom: indigo
-colorTo: purple
+title: CUDA LLM Inference Base
+emoji: 🧱
+colorFrom: gray
+colorTo: indigo
 sdk: docker
+app_port: 7860
 pinned: false
+short_description: CUDA inference base image for GPU jobs.
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Public base image: CUDA + torch + a GPU inference server + generic ML/runtime
+dependencies, for running GPU inference jobs on HF Jobs. It contains **no
+proprietary code**; application code and any remaining dependencies are installed
+at runtime from a private repo.
+
+This Space exists only to build + publish the image; it serves a trivial idle
+endpoint.

dataset_reviewer/__init__.py

@@ -0,0 +1,4 @@
+"""Stub package — present only so `pip install .` resolves dataset-reviewer's
+dependencies into this public base image WITHOUT shipping any proprietary source.
+The real dataset_reviewer wheel is force-reinstalled over this stub at Job runtime
+(see /opt/run-job.sh)."""

pyproject.toml

@@ -0,0 +1,36 @@
+[project]
+name = "dataset-reviewer"
+version = "0.1.0"
+requires-python = ">=3.12"
+dependencies = [
+    "datasets>=4.5.0",
+    "datatrove[io]>=0.9.0",
+    "duckdb>=1.4.4",
+    "httpx",
+    "jinja2>=3.1.6",
+    "openai>=1.0",
+    "markdownify",
+    "pyarrow>=23.0.1",
+    "pydantic>=2.12.5",
+    "markdown-it-py>=4.0.0",
+    "rich>=14.3.3",
+    "loguru",
+    "python-dotenv",
+    "huggingface-hub",
+    "pandas",
+    "requests",
+    "markupsafe",
+    "fsspec",
+    "transformers>=4.40",
+    "tiktoken>=0.12.0",
+    "tenacity>=9.1.4",
+    "pillow>=12.2.0",
+    "nh3>=0.2",
+]
+
+[build-system]
+requires = ["setuptools>=61"]
+build-backend = "setuptools.build_meta"
+
+[tool.setuptools.packages.find]
+include = ["dataset_reviewer*"]

run-job.sh

@@ -0,0 +1,40 @@
+#!/usr/bin/env bash
+# HF Job bootstrap: pull the private dataset_reviewer wheel + demo modules +
+# runtime requirements from JOB_CODE_REPO (authenticated by HF_TOKEN), install the
+# wheel over the baked stub, install the technique-revealing deps kept out of the
+# public image, then run the headless pipeline. Each phase is timed so its cost is
+# visible in the Job log. Keeps proprietary code + model/technique names OUT of the
+# public base image.
+set -euo pipefail
+
+: "${JOB_CODE_REPO:?JOB_CODE_REPO is required}"
+
+t=$SECONDS
+echo "[bootstrap] downloading job code from ${JOB_CODE_REPO}"
+python3 - <<'PY'
+import os
+from huggingface_hub import snapshot_download
+snapshot_download(repo_id=os.environ["JOB_CODE_REPO"], repo_type="model", local_dir="/code")
+PY
+echo "[bootstrap] job code downloaded in $((SECONDS - t))s"
+
+t=$SECONDS
+echo "[bootstrap] installing dataset_reviewer wheel over the baked stub"
+pip install --no-deps --force-reinstall --no-index /code/wheels/*.whl
+echo "[bootstrap] wheel installed in $((SECONDS - t))s"
+
+# The estimate job runs no signals (run_job.py returns at the ESTIMATE_ONLY branch before any
+# signal loads), so skip the technique-revealing runtime deps it never imports — they dominate the
+# bootstrap, and `import dataset_reviewer` succeeds without them (its heavy backends import lazily).
+if [ -f /code/requirements-runtime.txt ] && [ "${ESTIMATE_ONLY:-}" != "1" ]; then
+  t=$SECONDS
+  echo "[bootstrap] installing runtime ML dependencies"
+  pip install -r /code/requirements-runtime.txt
+  echo "[bootstrap] runtime deps installed in $((SECONDS - t))s"
+elif [ "${ESTIMATE_ONLY:-}" = "1" ]; then
+  echo "[bootstrap] estimate-only run — skipping runtime ML dependencies"
+fi
+
+echo "[bootstrap] launching run_job.py"
+cd /code
+exec python3 -u run_job.py