[sglang] Persist TP, prefill, NEXTN, and DeepEP graphs on main

[rchalamala]

Summary

• port Foundry to pinned SGLang main through its plugin and pluggable full-graph backend APIs
• add versioned per-rank graph-session catalogs and allowlisted nested output serialization
• persist full decode and full-prefill body graphs while preserving SGLang's eager prefill tail
• restore Qwen3.5 NEXTN target/draft graphs, hybrid GDN/Mamba metadata, symmetric-memory TP, and DeepEP low-latency NVL/IPC graphs
• fail safely on archive/config drift; non-full prefill backends pass through upstream, while unsafe non-GDN hybrid/forward-hook full-prefill configurations use eager prefill

Validated configurations

Pinned SGLang: 9b853e6832e71a3058212df02a025232a453e146

• Full prefill, Qwen3-1.7B on H100: prefill 16/64 plus decode restored with exact token IDs (ap-21ZFLYka0y7xiAJhAeRBOv).
• NEXTN, Qwen3.5-9B TP2 on 2×H100: target verify, draft, hidden-state schemas, and prefill restored; exact seeded IDs and nontrivial acceptance (ap-sOMcOXV1i1rzFvBpaE5QSf).
• DeepEP, Qwen3-30B-A3B-FP8 TP2/EP2 on 2×H100: semantic per-expert payload oracle and SGLang prefill/decode dispatch+combine replay passed (ap-cHBR8KjDjQVceSSbl5beA, ap-eNHOvES59jX3TVV4MuAylQ).
• Combined, Qwen3.5-35B-A3B online FP8 TP2/EP2 + NEXTN + DeepEP + full prefill: all strict checks and empty semantic archive differences (ap-LrcdL80HqontHF8EDxqC5s).
• Target deployment, Qwen3.5-122B-A10B-GPTQ-Int4 TP4 + full prefill on 4×H100: three graphs/rank, exact outputs, symmetric collectives, reproducible archives and offset 59301167104 (ap-BirUveQeDvzsIFRkwyy8yW).

Constraints

• TP persistence uses one explicit decode shape and --enable-torch-symm-mem; full prefill may use multiple token buckets.
• Only the full prefill backend is persisted. Breakable/tc-piecewise remain upstream pass-through modes.
• NEXTN is the GPU-validated speculative algorithm.
• Persisted hybrid prefill is validated for GDN. Other hybrid-linear backends and configurations with forward hooks fall back to eager prefill.
• Hopper DeepEP validation uses low-latency legacy NVL/IPC with FA3 and the documented NVSHMEM settings; modern use_fabric signatures remain supported but are not claimed as validated here.
• PP, LoRA graph variants, PDMux, diffusion models, runtime recapture, and arbitrary multi-shape TP remain out of scope.

Verification

• 111 focused CPU contracts passed.
• Ruff check/format, py_compile, git diff --check, and every pre-commit run --all-files hook passed.
• Native validation compares baseline, SAVE, a second cold SAVE, and fresh-process LOAD with exact tokens, catalogs, semantic fingerprints, offsets, graph-node collectives/DeepEP kernels, and runtime-error scanning.

Review notes

The archive keeps complete graph dependencies (strip_dependencies=False) because shared-executor optimization is not yet safe for distributed collective replay. The branch has passed task-scoped reviews and a final whole-branch review with no open Critical or Important findings.