Consolidate TP support and validate vLLM and SGLang

[rchalamala]

Summary

• Consolidate the TP, VMM-IPC, graph-loader, SGLang-main, DeepEP, and validation work into one branch based directly on main.
• Add an experimental vLLM Qwen3-8B TP=2 recipe and 2×H100 baseline/SAVE/SAVE2/LOAD proof for the pinned NCCL P2P/IPC path.
• Keep the existing experimental SGLang NCCL TP=2 full-shape path and add a replay-safe SGLang-main torch symmetric-memory option.
• Harden VMM reservations and IPC ownership, pitched-allocation export, no-fabric DeepEP coverage, and graph-capture contracts.
• Align status and scope with upstream discussion Consolidate TP support and validate vLLM and SGLang #5 and issues [sglang] No-fabric DeepEP EP: port the NVL CUDA-IPC path to the SGLang integration #1, [sglang] Experimental dense tensor-parallel recipe (NCCL all-reduce over VMM-IPC) #2, and [tp] Consolidate branches and validate vLLM/SGLang #6.

SGLang symmetric-memory fix

PyTorch symmetric memory owns process-local virtual mappings and device pointer tables. SAVE preserves full graph JSON and a versioned backend/state record; LOAD rebuilds required SGLang warmup state, validates the exact TP=2 BF16 two-shot ABI and copy/dependency topology, and relocates every supported external operand to the live communicator.

The supported scope is intentionally fail-closed:

• TP must be 2.
• Exactly one batch-1 CUDA graph is persisted (SGLANG_CUDA_GRAPH_MAX_BS=1).
• Larger runtime batches fall back to eager execution.
• Multi-shape symmetric capture is rejected because later shapes retain mutable FlashInfer state established by earlier captures.
• Backend, dtype, capacity, rank/world size, specialization, pointer inventory, and graph layout drift all abort before replay.

Verification

• 42 passed across the final CPU TP contract suite.
• pre-commit run --all-files: all hooks passed.
• vLLM TP=2 on 2×H100: 64 graphs restored per rank, reproducible offsets/inventories, no recapture, deterministic sequential and concurrent outputs (vllm-tp-8afbfda61de-1784771242076820717).
• SGLang symmetric TP=2 on 2×H100: all checks passed in baseline/SAVE/SAVE2/LOAD run 8383c4db4d9e6459a6238960-1784839292861701133, including reproducible semantic fingerprints and offsets, one graph restored per rank at 68555898880, byte-identical deterministic outputs, successful eager batch fallback, successful graph replay afterward, and no runtime errors.
• Final validated code reloaded that retained H100 archive successfully with the hardened validator and reproduced deterministic and post-fallback outputs.

Multi-shape follow-up

• Approved design: docs/superpowers/specs/2026-07-23-sglang-multishape-symmetric-state-design.md
• Task-by-task TDD plan: docs/superpowers/plans/2026-07-23-sglang-multishape-symmetric-state.md

The plan retains SGLang's shared graph pool while introducing explicit per-shape FlashInfer state capsules, typed graph-operand manifests, exact pinned ABI validation, and staged 1+8 → 1+8+32 → full-inventory H100 gates.

Scope

This validates pinned experimental TP recipes, not general TP across arbitrary models, GPU counts, graph-shape inventories, or collective versions. General/official TP remains tracked upstream.