fix(slurm): support co-located proxy in sglang-disagg (min 2 nodes)

[mkuznet1]

Motivation

The SGLang disaggregated launcher hard-required a dedicated proxy node and a
minimum of 3 nodes (1 proxy + 1 prefill + 1 decode). This blocked the minimal
and more flexible co-located layout, where the proxy/router runs on the first
prefill node (started by the model run.sh), as already done by the vllm-disagg
launcher. The goal is to support both topologies and lower the minimum cluster
size to 2 nodes.

Technical Details

All changes are in src/madengine/deployment/slurm.py:

• Relax custom-split validation to accept two topologies instead of one:
  • dedicated proxy node: 1 + prefill + decode == nnodes
  • co-located proxy on the first prefill node: prefill + decode == nnodes
xP/yD and the exported topology env are unchanged; the proxy is launched by
    the model script, so both layouts are valid. Error message updated to state the
    expected node count for each topology.
• Lower the minimum node-count guard from < 3 to < 2, since the validation
  above already permits prefill + decode == nnodes. Updated the docstring
  (Minimum cluster: 2 nodes, nnodes must be >= 2) and the corresponding
  ValueError message accordingly.

[Copilot]

Pull request overview

This PR updates the SLURM SGLang-disaggregated launcher setup to support a co-located proxy/router topology (proxy runs on the first prefill node) in addition to the existing dedicated proxy-node topology, reducing the minimum cluster size to 2 nodes.

Changes:

• Relaxes custom-split validation to accept either 1 + xP + yD == nnodes (dedicated proxy) or xP + yD == nnodes (co-located proxy).
• Lowers the minimum-node guard and updates related docstrings/error messages accordingly.
• Updates the generated “Cluster Configuration” comment block to describe the two supported topologies.

---

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

[mkuznet1]

Test run (8a166ba)

707 passed / 17 failed / 2 skipped (unit 522/2, integration 139/1, e2e 46/14).

Cat	#	Root cause
A	13	user not in render/video → amd-smi warnings break GPU count (17 vs 8) → empty product name
B	2	stale keepAlive tests expect container name without _node0 suffix
C	2	node has /nonexistent dir (0750) → PermissionError not FileNotFoundError

A (13): integration test_renderD_count_matches_gpu_count; e2e Discover test_dynamic/test_multiple; 9 profiling tests; test_gpu_product_name_matches_arch.
B (2): test_keepAlive_keeps_docker_alive, test_keepAlive_preserves_model_dir.
C (2): unit test_upload_file_to_mongodb_file_not_found, test_validate_additional_context_file_not_found.

Bottom line: 15/17 fixable without code (groups or root-in-container); 2 keepAlive need a test fix.

[mkuznet1]

The default-split path now special-cases nnodes == 2 to a co-located 1 prefill + 1 decode split (xP=1, yD=1) instead of yD=0. For nnodes >= 3 the dedicated-proxy default already yields yD >= 1.

Added test coverage: new TestSglangDisaggDefaultSplit exercises the default-split path — nnodes=2 is co-located (xP=1, yD=1), parametrized nnodes in [2,3,4,5,6,8] all yield >=1 prefill and >=1 decode, and nnodes<2 raises ValueError. Local unit run: 28 passed.