codemaivanngu/nemo-rl-v070-b200-runtime

NeMo RL v0.7.0 runtime for B200

Runtime for the Llama-3.1-8B-Instruct student, with a separate Qwen3-235B-A22B teacher node. Each project node has 8 B200 GPUs. This package contains no model weights, dataset, credentials, or OPD implementation. The future OPD source overlay remains a separate change from this dependency runtime.

Candidate until qualification.json reports pass. A Modal receipt verifies the tests and userspaces named in that receipt. It does not qualify the company's driver, fabric, scheduler, full Llama-8B training, or Qwen-235B serving.

Pinned inputs

Component	Input
NeMo RL	v0.7.0, 81aa43dda4765b0429cf31dab44441e4e4383911
Python	3.13.13
uv builder	0.11.18
Torch	2.11.0, cu130
vLLM	0.20.0, upstream Linux x86_64 wheel
FlashInfer	0.6.8.post1
Dependency resolution	upstream uv.lock; each environment has a package inventory
Base	CUDA 13.0.2 devel, Ubuntu 22.04, digest in spec.json

Three environments are shipped: base, automodel, and vllm. Upstream declares the two backend extras incompatible, so combining them in one venv is unsupported. MCore, old FSDP, SGLang, Gym and quantized ModelOpt workers are outside this runtime's scope. Selecting an unavailable tier fails before launch. Use policy.dtensor_cfg.enabled=true and policy.dtensor_cfg._v2=true, matching the upstream v0.7.0 X-Token recipe and the AutoModel worker environment.

Portable installation on Ubuntu 22.04 or 24.04

Use the SHA-256 from releases/v070-r1/portable.json and a fixed HF commit revision, then invoke the same helper on either OS:

export PROXY='<company proxy URL>'
# Download ensure_runtime.sh from this HF dataset at a fixed commit, through PROXY,
# and verify the helper's SHA-256 from the delivery receipt before running it.
bash ensure_runtime.sh '<fixed HF URL>/releases/v070-r1/runtime.tar.gz' '<archive SHA-256>'
source "$HOME/.local/share/nemo-rl-v070-b200/activate.sh"
python "$NEMO_RL_PREBUILT_ENV_ROOT/check_runtime.py" --imports --gpu

The optional third helper argument overrides the install prefix. The default is $HOME/.local/share/nemo-rl-v070-b200, or NRL_RUNTIME_PREFIX when set. Use the same absolute prefix on every node in one Ray cluster. The helper requires existing Bash, Python 3, curl, tar, gzip and sha256sum. It verifies the archive, refuses to replace a different installation and rechecks an existing matching installation. Failed staging is retained for diagnosis. It installs no packages on the company host and accepts remote downloads only from Hugging Face through the explicit proxy.

Host prerequisites: Linux x86_64, Ubuntu 22.04/24.04, a CUDA-13-compatible NVIDIA driver (R580 or later, subject to actual GPU checks), scheduler-allocated B200s, and a working host C/C++ compiler for runtime JIT. CUDA headers, nvcc, libdevice and Python package libraries are shipped. No host glibc or NVIDIA driver is replaced. JIT caches live outside the dependency prefix under NRL_CACHE_ROOT. CUDA_VISIBLE_DEVICES is preserved. Choose a node-local filesystem for the prefix; dependency metadata scans can be slow on a network-mounted home directory. Override NRL_RUNTIME_PREFIX when needed.

activate.sh defaults to offline model access. Stage weights and datasets separately through the company proxy, pin revisions, and then point recipes to their local paths. Do not place weights or checkpoints in the runtime prefix.

Container

releases/v070-r1/runtime.docker.tar.gz loads directly into an existing Docker installation. It includes the pinned CUDA base layers:

docker load -i runtime.docker.tar.gz
docker run --rm --gpus all --ipc=host \
  -v /path/to/models:/models:ro -v /path/to/cache:/cache \
  -e NRL_CACHE_ROOT=/cache nemo-rl-v070-b200:r1 \
  python /opt/nrl-runtime/check_runtime.py --imports --gpu

Run only inside the GPU allocation. The host needs its existing NVIDIA Container Toolkit. The image uses Ubuntu22 userspace on either supported Ubuntu host. Dockerfile provides an equivalent recipe using the portable archive. Container archive integrity, image assembly and execution are separate qualification gates. An equivalent complete runtime.oci.tar is also provided for OCI/Apptainer/Skopeo workflows. Both archives carry the same image config and filesystem layers.

Source overlays and offline Ray workers

The shipped source adds NEMO_RL_PREBUILT_ENV_ROOT. It maps supported backend interpreters directly to envs/<tier>/bin/python, avoiding worker-side uv sync. Without the variable, upstream uv behavior remains unchanged. Do not combine this variable with NEMO_RL_PY_EXECUTABLES_SYSTEM=1.

For later source edits, mount or checkout an explicit source commit separately:

export NEMO_RL_SOURCE_ROOT=/work/source-at-pinned-commit
source /path/to/runtime/activate.sh
python "$NEMO_RL_SOURCE_ROOT/examples/run_grpo.py" --config /path/to/config.yaml

That source checkout must include the prebuilt interpreter integration. Source overlays do not require rebuilding dependencies unless the dependency contract changes. Record source commit, runtime image/archive digest, config, data/model revisions and run ID for every run.

Remote build and verification

On an existing Modal client, select the budget-checked profile explicitly:

MODAL_PROFILE=no1ceboy modal run runtime/b200/modal_build.py --stage build
MODAL_PROFILE=no1ceboy modal run runtime/b200/modal_build.py --stage package
MODAL_PROFILE=no1ceboy modal run runtime/b200/modal_build.py --stage container
# Recheck an existing export without recreating the large archives:
MODAL_PROFILE=no1ceboy modal run runtime/b200/modal_build.py --stage container-check
MODAL_PROFILE=no1ceboy modal run runtime/b200/modal_qualify.py --stage single
MODAL_PROFILE=no1ceboy modal run runtime/b200/modal_qualify.py --stage nccl8
MODAL_PROFILE=no1ceboy modal run runtime/b200/modal_build.py --stage qualification

Builds use a dedicated v2 Volume, because uv requires filesystem hardlinks. Container export uses non-preemptible CPU with a 90-minute timeout. Modal bills this mode at 3x its standard CPU and memory rates. HF publication uses the same CPU mode with a 60-minute timeout and enables online access only in its own subprocess. The nccl8 stage checks each OS sequentially; nccl8-u24 runs only Ubuntu24 when Ubuntu22 already has a valid receipt for the same portable SHA-256. Packaging streams the payload once to the remote runner's local scratch disk; relocation, symlink checks and compression run there to avoid repeated scans of the network Volume. Failed scratch staging is retained for diagnosis. The OCI execution check similarly stages checksum-verified image blobs locally before unpacking. Its chroot supplies real null/zero/random/urandom device nodes, as a container runtime normally does; they are not embedded into the image. For an unqualified candidate, modal_repack.py can repair packaging files from its frozen restored Modal image without compiling native dependencies again:

NRL_REPACK_IMAGE_ID=<restored-candidate-image-id> MODAL_PROFILE=no1ceboy \
  modal run runtime/b200/modal_repack.py

The tool checks the original archive marker, runs an offline CPU preflight, keeps the previous archive and refuses to overwrite a qualified release. The resulting SHA-256 must pass the complete qualification sequence again.

No application-error GPU retries are configured. Modal may restart Functions after infrastructure preemption. Models are synthetic tiny random Llama checkpoints for kernel/generation tests; no Llama8B/Qwen235B weights are downloaded. Receipts distinguish import, relocation, backward/optimizer, vLLM generation and logprobs, Ray backend dispatch, and 8-rank NCCL tests. Benchmark quality and OPD correctness are not measured by these runtime checks. Single-GPU/import qualification blocks network egress and starts with empty JIT and model caches. Image preparation is keyed by the portable archive's SHA-256. CPU receipts defer driver-linked DeepEP/DeepGEMM imports; the single-GPU gates must import them successfully before final qualification can pass.

CPU regression checks use only Python's standard library:

python3 -m unittest runtime.b200.test_prebuilt_env runtime.b200.test_relocate runtime.b200.test_ensure_runtime runtime.b200.test_launchers runtime.b200.test_export_oci -v

Upstream dependencies retain their own licenses. CUDA/container terms and bundled third-party notices are included with the exported runtime.