feat: add Ray GRPO runtime image (CUDA 12.8, Python 3.12)

images/runtime/ray/cuda/2.53.0-py312-cu128-grpo/Dockerfile

@@ -0,0 +1,79 @@
+# Ray + GRPO runtime image for KubeRay on OpenShift AI
+#
+# Base: quay.io/modh/ray:2.53.0-py312-cu128
+#   Ray 2.53.0, Python 3.12, CUDA 12.8, CUDA dev headers, numpy<2
+#
+# Adds: vLLM 0.12.0 (rollout engine), verl 0.7.1 (FSDP training orchestrator),
+#       training-hub, flash-attn, and supporting packages for LoRA-GRPO training.
+#
+# Used as the container image for both head and worker pods in a KubeRay
+# RayCluster. verl launches via `python -m verl.trainer.main_ppo` on the
+# head node and distributes FSDP training + vLLM rollouts across workers.
+#
+# Build:
+#   podman build -t quay.io/<org>/ray-grpo:2.53.0-py312-cu128 .
+#
+# Install order matters:
+#   1. vllm first — pins torch 2.9.0, numba 0.61.2, flashinfer, triton
+#   2. flash-attn — pre-built wheel matched to torch 2.9 ABI
+#   3. verl + ML libs
+#   4. training-hub[grpo] (numba constraint relaxed to >=0.61.2 upstream)
+
+FROM quay.io/modh/ray:2.53.0-py312-cu128
+
+LABEL name="ray-grpo-cu128" \
+      summary="Ray + vLLM + verl GRPO runtime for KubeRay on OpenShift AI" \
+      description="Extends the ODH Ray 2.53.0 CUDA 12.8 image with vLLM 0.12.0, \
+verl 0.7.1, flash-attn 2.8.3, and training-hub for LoRA-GRPO training." \
+      io.k8s.display-name="Ray GRPO Runtime (CUDA 12.8)" \
+      io.k8s.description="Ray 2.53.0 + vLLM 0.12.0 + verl 0.7.1 GRPO training image" \
+      authoritative-source-url="https://github.com/opendatahub-io/distributed-workloads"
+
+USER 0
+WORKDIR /opt/app-root/bin
+
+# ── 1. vLLM + pinned companions ─────────────────────────────────────
+# vllm 0.12.0 requires torch==2.9.0 and numba==0.61.2.
+# verl 0.7.1 requires vllm>=0.8.5 and its code needs run_headless
+# (first available in vllm ~0.10). vllm 0.12.0 is the max verl allows.
+RUN pip install --no-cache-dir \
+    "vllm==0.12.0" \
+    "torch==2.9.0" \
+    "numpy<2.0.0" \
+    "numba==0.61.2" \
+    "llvmlite==0.44.*"
+
+# ── 2. flash-attn (pre-built wheel) ──────────────────────────────────
+# torch 2.9.0 uses cxx11_abi=True (the cu126+ ABI switch from PyTorch PR #142064).
+# Use the pre-built wheel from flash-attention GitHub releases matched
+# to torch 2.9, CUDA 12, cp312, cxx11abiTRUE.
+RUN pip install --no-cache-dir \
+    "https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu12torch2.9cxx11abiTRUE-cp312-cp312-linux_x86_64.whl"
+
+# ── 3. verl + ML framework libs ─────────────────────────────────────
+RUN pip install --no-cache-dir \
+    "verl>=0.7" \
+    "peft>=0.15" \
+    "transformers>=4.57.6,<5.0" \
+    "datasets>=4.0.0"
+
+# ── 4. training-hub[grpo] ────────────────────────────────────────────
+# GRPO code lives on the lora-grpo branch (not yet merged to main).
+# instructlab-training & mini_trainer from main to pick up the relaxed
+# numba>=0.61.2 constraint (merged but not yet released to PyPI).
+# Once all released, replace with: pip install "training-hub[grpo]"
+RUN pip install --no-cache-dir \
+    "instructlab-training @ git+https://github.com/instructlab/training.git@main" \
+    "rhai-innovation-mini-trainer @ git+https://github.com/Red-Hat-AI-Innovation-Team/mini_trainer.git@main" \
+    "training-hub[grpo] @ git+https://github.com/Red-Hat-AI-Innovation-Team/training_hub.git@lora-grpo"
+
+# ── 5. Ray / NCCL / vLLM environment defaults ───────────────────────
+ENV NCCL_DEBUG=WARN \
+    TOKENIZERS_PARALLELISM=true \
+    VLLM_ALLOW_RUNTIME_LORA_UPDATING=true \
+    VLLM_LOGGING_LEVEL=WARN \
+    VLLM_USE_V1=1 \
+    RAY_DEDUP_LOGS=0
+
+USER 1001
+WORKDIR /opt/app-root/src

images/runtime/ray/cuda/2.53.0-py312-cu128-grpo/README.md

@@ -0,0 +1,29 @@
+# Ray GRPO Runtime — CUDA 12.8, Python 3.12
+
+Extends `quay.io/modh/ray:2.53.0-py312-cu128` with the ML stack needed for **LoRA-GRPO** (Group Relative Policy Optimization with LoRA) training using the [training-hub](https://github.com/Red-Hat-AI-Innovation-Team/training_hub) library, orchestrated by [verl](https://github.com/volcengine/verl) on KubeRay.
+
+## Key packages
+
+| Package | Version | Role |
+|---------|---------|------|
+| vLLM | 0.12.0 | LLM rollout engine |
+| PyTorch | 2.9.0 | Deep learning framework |
+| verl | 0.7.1 | GRPO training orchestrator (FSDP + vLLM) |
+| flash-attn | 2.8.3 | Memory-efficient attention |
+| training-hub | latest | High-level GRPO training API |
+| peft | ≥0.15 | LoRA adapters |
+| transformers | ≥4.57.6 | HuggingFace model loading |
+
+## Build
+
+```bash
+podman build -t quay.io/<org>/ray-grpo:2.53.0-py312-cu128 .
+```
+
+## Usage
+
+Use this image as the `rayImage` for both head and worker pods in a KubeRay `RayCluster` CR. verl launches training via `python -m verl.trainer.main_ppo` on the head node and distributes FSDP training + vLLM rollouts across workers.
+
+## Notes
+
+- **Git branch installs**: `instructlab-training` and `rhai-innovation-mini-trainer` are currently installed from their `main` branches to pick up the relaxed `numba>=0.61.2` constraint (merged but not yet released to PyPI). `training-hub[grpo]` is installed from the `lora-grpo` branch (GRPO code not yet merged to `main`). Once all packages are released, Step 4 simplifies to `pip install "training-hub[grpo]"`.