sglang profiling

[AakarshAMD]

Motivation

Add an opt-in profiling workflow for SGLang disaggregated serving using rocprofv3, ROCtx, and MoRI instrumentation, while preserving the normal non-profiling workflow.

Technical Details

• Added a dedicated ROCm 7.2.3 profiling Dockerfile with SGLang and MoRI instrumentation patches.
• Added benchmark_xPyD_profile.sh for profiling-specific request IDs, timing manifests, warmup control, and multi-sweep isolation.
• Added conditional rocprofv3 kernel/marker capture and graceful worker finalization.
• Added post-processing for clean traces, request statistics, MoRI correlation, kernel categorization, and local-rank-0 analysis.
• Added validation for capture completeness, request identity, malformed durations, and MoRI transfer attribution.
• Profiling is enabled with RUN_PROFILE=1; the standard benchmark and Dockerfile remain unchanged for normal runs.

Test Plan

• Build the profiling image with ENABLE_ROCTX=1.
• Run profiled 1P1D and 2P2D configurations across TP and EP models.
• Verify benchmark completion, worker capture finalization, request-ID consistency, MoRI attribution, clean trace generation, and kernel-analysis outputs.

Test Result

All tested configurations passed:

• DeepSeek-V3 — 1P1D, EP8
• DeepSeek-V3 — 2P2D, EP16
• DeepSeek-R1 — 1P1D, EP8
• DeepSeek-R1 — 2P2D, EP16
• Llama-3.1-8B-Instruct — 1P1D, TP8
• Llama-3.3-70B-Instruct — 1P1D, TP8
• Mixtral-8x7B-Instruct-v0.1 — 1P1D, TP8

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.