public base image (hf-image-vllm)

Dockerfile

@@ -1,18 +1,18 @@
 # syntax=docker/dockerfile:1.6
 # PUBLIC base image for GPU inference jobs (HF Jobs) — vLLM serving variant.
 #
-# 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 + the remaining runtime dependencies at startup and runs run_job.py.
+# Same contract as deploy/hf-image-llama-cpp/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 +
+# job modules + the remaining runtime dependencies at startup and runs run_job.py.
 #
 # Built FROM vLLM's official OpenAI-server image rather than the llama.cpp
 # CUDA-runtime base: vLLM's Triton JIT needs a CUDA build chain + a matched torch
 # (gcc/ninja/nvcc are absent from the runtime base). torch therefore comes from the
-# base and is omitted from the generated pyproject (JOB_IMAGE_BASE_PROVIDES=torch).
+# base and is omitted from the generated pyproject (HF_JOB_IMAGE_BASE_PROVIDES=torch).
 #
-# Build: JOB_IMAGE_DIR=image-vllm JOB_IMAGE_BASE_PROVIDES=torch make demo-deploy-job
-# (run-job.sh + the dataset_reviewer stub are shared from demo/image/.)
+# Build: HF_JOB_IMAGE_DIR=hf-image-vllm HF_JOB_IMAGE_BASE_PROVIDES=torch make deploy-hf-job
+# (run-job.sh + the dataset_reviewer stub are shared from deploy/hf-image-common/.)
 
 # Pinned (not :latest) so the serving CLI contract can't shift under a live demo run:
 # vLLM 0.12 renamed the structured-outputs backend flag, and this repo's serve argv +
@@ -22,7 +22,7 @@ FROM vllm/vllm-openai:v0.19.1
 
 # The base image sets ENTRYPOINT to the OpenAI server; reset it so HF Jobs'
 # `bash /opt/run-job.sh` (and the idle Space CMD) run as plain commands. The Job
-# starts its own server as a co-located subprocess (demo/job/llm_server.py).
+# starts its own server as a co-located subprocess (deploy/job/llm_server.py).
 ENTRYPOINT []
 
 ENV DEBIAN_FRONTEND=noninteractive \
@@ -49,11 +49,10 @@ 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 uses to DM a submitter — 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"
+# Non-library job deps, single-sourced in job_deps.JOB_EXTRA_DEPS — deploy_hf_job.py
+# stages the requirements file into this build context.
+COPY requirements-extras.txt /app/requirements-extras.txt
+RUN --mount=type=cache,target=/root/.cache/pip pip install -r /app/requirements-extras.txt
 
 COPY run-job.sh /opt/run-job.sh
 RUN chmod +x /opt/run-job.sh

pyproject.toml

@@ -4,7 +4,7 @@ version = "0.1.0"
 requires-python = ">=3.12"
 dependencies = [
     "datasets>=5.0.0",
-    "datatrove[io]>=0.9.0",
+    "datatrove[io] @ git+https://github.com/huggingface/datatrove.git@a035d36e8a13d22ce04e84b84f30b4ec45f9c48e",
     "duckdb>=1.4.4",
     "ddgs",
     "httpx",
@@ -26,7 +26,7 @@ dependencies = [
     "langgraph",
     "loguru",
     "python-dotenv",
-    "huggingface-hub",
+    "huggingface-hub>=1.13,<1.17",
     "pandas",
     "pytz",
     "requests",

requirements-extras.txt

@@ -0,0 +1,2 @@
+huggingface_hub[hf_transfer]
+telethon==1.43.2

run-job.sh

@@ -1,20 +1,23 @@
 #!/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.
+# HF Job bootstrap: pull the private dataset_reviewer wheel + job modules +
+# runtime requirements from HF_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
 
-: "${JOB_CODE_REPO:?JOB_CODE_REPO is required}"
+# JOB_CODE_REPO is the pre-rename env name; launchers older than the rename send only it.
+: "${HF_JOB_CODE_REPO:=${JOB_CODE_REPO:-}}"
+: "${HF_JOB_CODE_REPO:?HF_JOB_CODE_REPO is required}"
+export HF_JOB_CODE_REPO
 
 t=$SECONDS
-echo "[bootstrap] downloading job code from ${JOB_CODE_REPO}"
+echo "[bootstrap] downloading job code from ${HF_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")
+snapshot_download(repo_id=os.environ["HF_JOB_CODE_REPO"], repo_type="model", local_dir="/code")
 PY
 echo "[bootstrap] job code downloaded in $((SECONDS - t))s"