docs(skills): add GPU performance analysis and inference optimization skills

[shanmugamr1992]

Summary

Adds two skills/ guides covering the profile-to-optimization workflow for Megatron-Core inference performance work. Documentation only — no changes to any library or training code.

• nsight-system-analysis — analyze nsys traces to find the per-iteration gap between two implementations, or break down a single profile. Covers iteration anchoring, GPU busy/idle split (union-of-intervals, not per-stream sums), per-category kernel breakdown, module slicing between GEMM anchors, exposed communication, and single-forward-pass kernel composition. Ships scripts for each step plus a report template and SQL recipes.

• optimize-inference-siddharth — what to change, in what order, and what never to break when an architecture is slower than vLLM. Encodes the patterns behind Siddharth Singh's (@sidsingh-nvidia) 2026 inference work: CUDA-graph scope and bucket coverage, the inference_optimized MoE stack (NVLS AllGather-V, fused grouped GEMM, shared-expert overlap, pad-row masking), Mamba/SSM scratch sizing, Triton production hygiene, and per-step host overhead — as eleven hard rules with the failure each one prevents.

  It also adds methodology validated on a Qwen3-30B-A3B EP4 campaign that took mcore from 66% to 81% of the vLLM DP4+EP baseline on 4xGB200:

  • Decision gates — quantify a lever's ceiling before writing code. Three worked gates each avoided a multi-week effort, including a hand-written grouped GEMM whose entire ceiling was 1.45x, most of it reachable by tile tuning alone.
  • Competitor-trace differential — is an individual kernel slower, or are there just more of them? Opposite fixes.
  • A/B protocol — session drift reached 1.6%, larger than most remaining wins, so acceptance requires same-session back-to-back arms with non-overlapping distributions rather than a mean delta.
  • nsys failure modes — the trace flag combinations that deadlock finalization on MoE decode under CUDA graphs, and the working substitutes.

The second skill is self-maintaining: references/updating-this-skill.md documents the triggers, routing, quality bar, deletion policy, and revision log, so campaigns extend and prune it rather than letting it go stale.

Test plan

Documentation only; no code paths are touched and nothing is imported by the library or tests.

• Both SKILL.md frontmatter blocks parse as YAML, with name matching the directory
• All intra-skill relative links resolve
• All scripts/*.py byte-compile; taxonomy_template.yml parses
• license: Apache-2.0 and SPDX headers consistent with the repo, matching the convention of the existing skills/

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.