launch

# launch Run training, tests, or benchmarks for **Halo** inside the prebuilt Docker image. The host has **no usable Python** (PyTorch 2.11+cu130, DeepEP, Flash Attention live only inside the images), so anything that *executes* must run inside the container. The image is built with `uv` (`uv pip install` into the system interpreter — no venv, no Poetry), so `python`/`torchrun`/`accelerate`/ `pytest` are on `PATH` — call them directly, **no prefix**. ## 1. ALWAYS prefer a `make` target first The Docker incantation lives **once** in the repo `Makefile` — contributors, CI, and this skill all call these targets instead of re-typing `docker run`. Read the Makefile (`make help` or open the repo-root `Makefile`) and use a target if one fits. The relevant targets: ```bash # Training — CONFIG is required; METHOD defaults to sft; NPROC defaults to 8; # IMAGE defaults to halo:blackwell; EXTRA passes through extra CLI flags. # The recipe routes through `python -m src.cli launch METHOD CONFIG` (the halo CLI # picks python/torchrun/accelerate), so METHOD selects any scripts/training method. make train CONFIG=examples/sft/qwen3/qwen3-4b-ultrachat.yaml make train CONFIG=examples/sft/gptoss/gptoss-20b-multinode-ep.yaml \ NPROC=8 EXTRA="--expert_parallel_size=8" make train METHOD=offline-grpo CONFIG=examples/grpo/offline/gptoss/... \ NPROC=8 # non-SFT method via METHOD= make train CONFIG=examples/sft/qwen3/qwen3-4b-ultrachat.yaml \ IMAGE=halo:hopper # run on Hopper instead # Tests make seed-hf-cache # Hub configs + tokenizers the CPU tier reads (no weights) make test-cpu # pytest CPU tier (no GPU, runs in image) make test-gpu-core # core GPU tier (pre-merge, GPU changes; CI runs it on dispatch) make test-gpu-full # full GPU tier (heavy, many-GPU; run by hand) # Benchmarks make bench # EP/TP throughput benchmarks (NPROC GPUs) # Image builds (credential-free — every dep is public) make build-blackwell make build-hopper make build-vllm # vllm-server:0.26.0 make build-sglang # sglang-server:0.5.17 (NCCL matched to the training images) make build-all # all four # Rollout-server GPU tiers (the server must already serve on a GPU outside TRAINER_CUDA_DEVICES; # both check /health first and, without EFA=1, force the no-fabric NCCL_IB_DISABLE=1 NCCL_NET=Socket # both compose bases default to. EFA=1 passes --device=/dev/infiniband plus NCCL_NET=Libfabric # NCCL_NET_PLUGIN=ofi NCCL_IB_DISABLE=0 (and its own NCCL_SOCKET_IFNAME exclude list) to every # DOCKER_RUN instead — start the server with its # compose EFA overlay (-f docker-compose.{vllm,sglang}.efa.yml) then. The SGLang server needs # NCCL_CUMEM_ENABLE=1 on top, the compose default.) # Server and trainer must hold the same checkpoint, so each tier runs one pass per served model: # SERVER_TIER defaults to 'not moe' (dense Qwen/Qwen3-0.6B); the vLLM MoE half is # SERVER_TIER='moe and not gptoss' against Qwen/Qwen3-30B-A3B-Instruct-2507, then # SERVER_TIER='moe and gptoss' against unsloth/gpt-oss-20b-BF16; SGLang's (SERVER_TIER=moe) is gpt-oss only. # The vLLM server needs VLLM_REASONING_PARSER=qwen3 VLLM_USE_V2_MODEL_RUNNER=0: the benchmarks # send a per-effort CoT budget, which draws a 400 without a reasoning parser and another under # Model Runner V2. The health guard prints the exact compose line when it is missing. make test-gpu-vllm make test-gpu-sglang # Publish to Amazon ECR Public (anonymous pull; a moving tag plus an immutable -VERSION pin) make ecr-public-login make push-public-all # or push-public-{blackwell,hopper,vllm,sglang} # Host-only gates (pinned uvx ruff, no image; `make docs` is the link and anchor check) make lint make format make precommit make docs # In-image make install # check the lock installs into the image (throwaway container) make diagrams # regenerate agent-docs/assets from scripts/diagrams (CI byte-compares) make clean # prune wandb/, the ruff and pytest caches, $(HALO_SCRATCH) test scratch (leaves checkpoints/) ``` Notes that come straight from the Makefile recipes: - `make train` selects the method via `METHOD=` (default `sft`) and dispatches through the `halo` CLI (`python -m src.cli launch METHOD CONFIG --nproc NPROC -- EXTRA`), which resolves the right `scripts/training/` script and launcher. Method names come from `halo launch --list` (e.g. `sft`, `offline-grpo`, `classification`, `embedding`, `environmental-grpo`, and nested names like `preference/smpo`). For a raw one-off or an unusual flag layout the target doesn't fit, use the template in §2. - Every GPU target already bakes in `--gpus all --network host --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --shm-size=128g`, the `-e HF_HOME / HF_DATASETS_CACHE / TMPDIR / HALO_DATA_ROOT / PYTHONPATH / CUDA_DEVICE_MAX_CONNECTIONS=1` env, and the `$(CURDIR):/workspace`, `$(HALO_SCRATCH):$(HALO_SCRATCH)`, `$(AWS_DIR):/root/.aws` (`AWS_DIR ?= ~/.aws`) mounts. `--env-file .env` and the AWS mount are conditional on `ENV_FILE`/`AWS_DIR` (CI drops both to run creds-free). Don't re-specify those by hand when a target works. `DOCKER_RUN` does **not** add `--cap-add=SYS_PTRACE`, so py-spy attaches to a `make`-launched job only on a host whose `kernel.yama.ptrace_scope` is 0 — use §2 when you may need hang triage. - **`HALO_SCRATCH` (default `/mnt`) is the one override for this host's large volume** — the bind mount, the in-container `HF_HOME`/`HF_DATASETS_CACHE`/`TMPDIR`/`HALO_DATA_ROOT`, and `make clean`'s prune all derive from it. `/mnt` is **not** guaranteed large (on some hosts it shares the small root device), so verify with `findmnt`/`df -h` and pass `HALO_SCRATCH=/path/to/big/volume` when it isn't. `docker-compose.vllm.yml` and the devcontainer read the same variable. - `make` targets run in the **foreground** (`docker run --rm`, no `-d`). For a long detached job that survives the turn and logs to the scratch volume, use §2. ## 2. Raw detached `docker run` template (when `make` doesn't fit) This is the incantation from `CLAUDE.md`, parameterized. Use it for custom flags, disjoint-GPU runs, or long detached jobs. Fill in `<job>` (container + log name), `<nproc>`, `<script>`, `<config>`, and any extra flags. `$D` is the scratch volume, **resolved dynamically — never hardcode a path**: a directory's name does not prove its capacity. ```bash # Resolve the scratch volume dynamically (see Preflight in CLAUDE.md). D=$(findmnt -rbno TARGET,AVAIL,FSTYPE | awk '$3!~/tmpfs|overlay|squashfs|nfs|fuse|autofs/ && $2+0>20e9{print $2,$1}' | sort -rn | head -1 | awk '{print $2}') docker run -d --rm --name <job> --gpus all \ --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --shm-size=128g \ --cap-add=SYS_PTRACE --env-file .env \ -e HF_HOME="$D/hf" -e HF_DATASETS_CACHE="$D/hf/datasets" \ -e TMPDIR="$D/tmp" -e HALO_DATA_ROOT="$D" \ -v $(pwd):/workspace -v "$D:$D" -v ~/.aws:/root/.aws -w /workspace \ halo:blackwell \ bash -lc "torchrun --nproc_per_node=<nproc> <script> <config> > \"$D/<job>.log\" 2>&1" ``` Mandatory pieces (do NOT drop them): - **`--env-file .env`** — the repo-root `.env` supplies `WANDB_API_KEY`, `HF_TOKEN`, `AWS_DEFAULT_REGION`. The code does **not** auto-load it. - **No prefix** inside the container — call `python`/`torchrun`/`accelerate`/ `pytest` directly (they are on `PATH`; the image is uv-built, no Poetry, no venv). - **`-v $(pwd):/workspace -w /workspace`** — the code mounts live; edits on the host take effect immediately. - **`-v "$D:$D"`** — outputs, logs, HF cache all go here. Verify `$D` is a large volume before any multi-GB write (`findmnt` / `df -h` / `readlink -f`); the root FS is small and on some hosts shares a device with `/mnt`. - **`--cap-add=SYS_PTRACE`** — lets py-spy attach for hang triage (`agent-docs/reference/debugging.md`). - **`-e HALO_DATA_ROOT="$D"`** — toolkit scratch root: S3 dataset cache → `$D/s3_datasets`, profiler artifacts → `$D/profiling` (one knob; defaults to `~/.cache/halo` when unset). - **`-v ~/.aws:/root/.aws`** — S3 dataset/checkpoint access. Pre-cached S3 datasets live under `$HALO_DATA_ROOT/s3_datasets` (keyed by `md5("<bucket>/<key>")`) and load without live AWS. Single-GPU / accelerate / pytest variants just swap the `bash -lc "..."` body, e.g.: ```bash # Single GPU, plain python bash -lc "python <script> <config> > \"$D/<job>.log\" 2>&1" # Accelerate DDP / FSDP (prefer the FSDP2 config; CLAUDE.md flags an Accelerate FSDP-v1 corruption bug) bash -lc "accelerate launch --config_file launcher-configs/accelerate/fsdp2_gradop_config.yaml \ <script> <config> > \"$D/<job>.log\" 2>&1" # A specific pytest selection not covered by the make targets bash -lc "pytest -m gpu tests/gpu/test_suite.py -k <pattern> > \"$D/<job>.log\" 2>&1" ``` A **CPU-only** job (CPU tests, lint-in-image, one-off Python with no GPU) drops `--gpus all` and the GPU ulimits/shm, matching the Makefile's `DOCKER_RUN_CPU`: ```bash docker run -d --rm --name <job> \ -e PYTHONPATH=/workspace -e HF_HOME="$D/hf" \ -v $(pwd):/workspace -v "$D:$D" -w /workspace \ halo:blackwell \ bash -lc "pytest -m cpu tests/cpu > \"$D/<job>.log\" 2>&1" ``` Both mounts are load-bearing: without `-v "$D:$D"` the log is written into the container's ephemeral filesystem and destroyed by `--rm`, and without the HF cache the CPU tests that call `from_pretrained` directly **hard-fail** (not skip) the moment the Hub is unreachable. ## 3. Image selection — by GPU architecture | GPU | Arch | Image | |-----|------|-------| | B200 / B300 (x86_64 hosts; no arm64 image for GB200/GB300) | SM100 / SM103 (Blackwell) | `halo:blackwell` — FA4 (default) + FA2 + DeepEP, no FA3 | | H100 / H200 | SM90 (Hopper) | `halo:hopper` — FA2 + FA3 + DeepEP, no FA4 | Default to **blackwell** (the Makefile `IMAGE` default); confirm the host arch with `nvidia-smi`. Confirm with `docker run --gpus all --rm halo:blackwell nvidia-smi` (or `nvidia-smi` if available) when unsure of the host arch. If the local tag is missing, pull it from ECR — see `agent-docs/infrastructure/docker.md`. For `make`, override with `IMAGE=halo:hopper`. For the raw template, swap the image line. ## 4. GPU count and disjoint / concurrent runs - **GPU count** is the `torchrun --nproc_per_node=<n>` value (and `NPROC=<n>` for `make`). EP/CP/TP parallelism flags (e.g. `--expert_parallel_size`, `--context_parallel_size`, `--tensor_parallel_size`) pass through `EXTRA=` or directly after the config in the raw template. - **Subset of GPUs** — restrict the devices the container sees with `--gpus '"device=0,1,2,3"'` (note the nested quoting). Inside, torchrun uses `--nproc_per_node` equal to that count. - **Two jobs at once on disjoint GPUs** — give each its own device set, a **unique `--master_port`**, and a distinct `--name`/log: ```bash # Job A on GPUs 0-3 docker run -d --rm --name jobA --gpus '"device=0,1,2,3"' \ --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --shm-size=128g \ --env-file .env -e HF_HOME="$D/hf" -e HF_DATASETS_CACHE="$D/hf/datasets" \ -e TMPDIR="$D/tmp" -e HALO_DATA_ROOT="$D" -v $(pwd):/workspace -v "$D:$D" -v ~/.aws:/root/.aws \ -w /workspace halo:blackwell \ bash -lc "torchrun --nproc_per_node=4 --master_port=29801 <script> <config> > \"$D/jobA.log\" 2>&1" # Job B on GPUs 4-7 — different port + name + log docker run -d --rm --name jobB --gpus '"device=4,5,6,7"' \ --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --shm-size=128g \ --env-file .env -e HF_HOME="$D/hf" -e HF_DATASETS_CACHE="$D/hf/datasets" \ -e TMPDIR="$D/tmp" -e HALO_DATA_ROOT="$D" -v $(pwd):/workspace -v "$D:$D" -v ~/.aws:/root/.aws \ -w /workspace halo:blackwell \ bash -lc "torchrun --nproc_per_node=4 --master_port=29802 <script> <config> > \"$D/jobB.log\" 2>&1" ``` `CUDA_VISIBLE_DEVICES` (`-e CUDA_VISIBLE_DEVICES=0,1`) also works to scope