fix(sweep): filter invalid vLLM GPU topologies - #375

Summary

• filter vLLM sweep variants whose TP × PP × PCP × DP world cannot run within the GPUs allocated to each worker
• preserve valid mixed topologies such as TP4/DP2 on an eight-GPU worker
• keep intentional TP-only configurations that use fewer GPUs than the reserved node
• apply the guard before full schema validation so one oversized vLLM Router variant does not abort the entire sweep
• leave SGLang and TensorRT-LLM expansion unchanged because their parallel dimensions have different semantics

Motivation

Cartesian topology sweeps in srt-slurm-recipes currently expand combinations such as TP8/PP8/DP8 on an eight-GPU node. Those jobs cannot run, but constraining the recipes to TP8/PP1 also removes useful coverage such as TP4/DP2.

This addresses the root cause requested in NVIDIA/srt-slurm-recipes#9 by filtering incompatible vLLM variants during sweep generation.

Validation

• pytest tests/ -q on Linux: 1551 passed, 2 skipped, 6 deselected
• pytest tests/test_sweep.py -q on Linux and Windows: 25 passed
• ruff check src/srtctl/
• ruff format --check src/srtctl/
• ty check src/srtctl/core/sweep.py
• expanded all six recipes from srt-slurm-recipes PR Add Kimi K2.5 disagg STP and MTP recipes for GB200 NVfp4 (ISL8K_OSL1K and ISL1K_OSL1K) #9:
  • DeepSeek-R1 and Nemotron retain TP4/DP2 and cap retained worlds at eight GPUs
  • MiniMax retains its allocation-compatible DP8 variant

Checklist

• Added regression coverage for oversized TP/PP, mixed TP/DP, vLLM Router, non-vLLM isolation, and fully invalid sweeps
• Commit includes a DCO sign-off