[sglang] Port Foundry to main and validate TP2/TP4

[rchalamala]

Summary

• port Foundry to current SGLang main through its native plugin framework and pluggable full-decode graph backend
• retain the legacy SGLang fork path and fail closed for unsupported main configurations
• add one-shape torch symmetric-memory TP with a stable all-reduce/all-gather buffer
• persist hybrid Qwen3.5 Mamba pool sizing and harden graph-load/state-write failure handling
• consolidate prior TP, IPC, DeepEP, and archive-validation work relative to Foundry upstream main

TP contract

Foundry SGLang-main TP requires:

• exactly one explicit padded decode graph
• --enable-torch-symm-mem
• custom/fused alternate all-reduce paths disabled

The graph uses torch symmetric-memory all-reduce and PyTorch multimem_all_gather_out over the same stable symmetric buffer. Multi-shape TP remains fail-closed because later captures retain collective state that is not portable to a fresh process.

Pinned validation

SGLang: 9b853e6832e71a3058212df02a025232a453e146

• Single GPU Qwen3-1.7B: 52 graphs restored without recapture; outputs match baseline.
• Qwen3-8B TP2 on 2×H100: sequential outputs and prompt-discriminating eight-token fixed-batch IDs match baseline/LOAD; archive fingerprints and offsets reproduce; each rank graph contains 73 symmetric all-reduces, one symmetric all-gather, and zero NCCL kernels.
• Qwen3.5-122B-A10B-GPTQ-Int4 TP4 on 4×H100, run d124c611437f2b1dd2286ca6-1784833491971742743: every check passed, including discriminating batched token IDs, all-rank symmetric collective inspection, reproducible archives/offsets, restore without recapture, and no runtime errors.
• DeepEP semantic SAVE/LOAD: passed on 2×H100 with injective BF16 payload validation for every local expert.
• RNG replay: exact golden sequences passed through both single and parallel load APIs.
• CUDA suite: 24 passed (vLLM import test excluded because the Foundry vLLM fork is not installed; DeepEP run separately).
• CPU contracts: 10 passed.
• Full branch pre-commit: all hooks passed.
• Final code review: no Critical or Important findings; ready to merge.

Unsupported SGLang-main modes

Prefill graphs, speculative decoding, LoRA variants, PD multiplexing, pipeline parallelism, hidden-state output, diffusion models, DeepEP through the new main backend, TP without symmetric memory, and multi-shape TP.