[Docs] Add 5-node MiniMax example files

[yubofredwang]

Add 5-node MiniMax example files

[Copilot]

Pull request overview

Adds a new 5-node MiniMax-M2.5 (Eagle3) example configuration and launcher script to the TorchSpec examples/configs set, intended to document a 40-GPU (5-node) training/inference split workflow.

Changes:

• Added a new example run.sh for launching MiniMax-M2.5 5-node training with configurable training/inference GPU allocation.
• Added a new sglang_minimax_m25_5node.yaml config for the 5-node setup (model/dataset/training/inference/mooncake settings).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
examples/minimax-m25-5node-h200/run.sh	New launcher script for the 5-node MiniMax-M2.5 setup, including logging and (attempted) resume behavior.
configs/sglang_minimax_m25_5node.yaml	New example config defining the MiniMax-M2.5 5-node training + SGLang inference + mooncake parameters.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The resume logic is effectively always enabled because CHECKPOINT_DIR is derived from output_dir (so it will be non-empty whenever output_dir is set), which makes the script print "No checkpoint tracker found" on fresh runs and always passes training.load_path=... even when the directory doesn’t exist. Gate resume behavior on the checkpoint directory existing and containing latest_checkpointed_iteration.txt (or another explicit resume flag) before setting RESUME_ARGS / printing resume messages.

Suggested change

	if [ -n "$CHECKPOINT_DIR" ]; then
	RESUME_ARGS=(training.load_path="$CHECKPOINT_DIR")
	if [ -f "$CHECKPOINT_DIR/latest_checkpointed_iteration.txt" ]; then
	LAST_STEP=$(<"$CHECKPOINT_DIR/latest_checkpointed_iteration.txt")
	echo "Resuming from checkpoint: $CHECKPOINT_DIR/iter_$(printf '%07d' "$LAST_STEP")"
	if [ -n "$CHECKPOINT_DIR" ] && [ -d "$CHECKPOINT_DIR" ]; then
	if [ -f "$CHECKPOINT_DIR/latest_checkpointed_iteration.txt" ]; then
	LAST_STEP=$(<"$CHECKPOINT_DIR/latest_checkpointed_iteration.txt")
	echo "Resuming from checkpoint: $CHECKPOINT_DIR/iter_$(printf '%07d' "$LAST_STEP")"
	RESUME_ARGS=(training.load_path="$CHECKPOINT_DIR")

[Copilot]

PYTORCH_CUDA_ALLOC_CONF is exported here, but TorchSpec Ray actors only receive a small allowlist of environment variables (via get_torchspec_env_vars()), which currently does not include PYTORCH_CUDA_ALLOC_CONF. As a result, the allocator setting likely won’t apply to the training/inference worker processes where it matters. Either remove this export to avoid a false sense of effect, or plumb it through (e.g., add it to the env allowlist / train_env_vars) so it reaches all Ray actors.

Suggested change

export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

[Copilot]

This example targets a multi-node (5-node) run, but the script doesn’t preflight that it’s connected to a Ray cluster. If Ray isn’t running / RAY_ADDRESS isn’t set, TorchSpec may start a local Ray instance and then fail after waiting for 40 GPUs, which is slow and not very actionable. Consider adding a ray status (as in the other multi-node examples) with clear instructions to start/join the cluster or set RAY_ADDRESS before invoking torchspec.train_entry.