Add experimental SGLang 0.5.19 integration

[zhaoyboo]

Summary

• Add third_party/sglang-integration, pointing to the newly published ProjectDMX/DMI-SGLang-Integration at 97c8fb4fea1184fb06e1f9df4e73e4cc93899fa9.
• Add docs/sglang.md and update README/install guidance for a dedicated SGLang environment, the offline DMIEngine API, and explicit flush before shutdown.
• Keep backend-specific code in the integration repository: this core PR changes only the submodule and documentation. It does not fork SGLang or change DMI native/API implementation.

Scope and compatibility identity

Experimental offline, non-streaming Qwen3 dense, TP1/PP1 integration for official SGLang v0.5.19, source commit 0bcd822377da7b5718e674eaf9c870d349424dd1.

Runtime used for qualification: DMI 1.2.0 / API v1, Python 3.12, torch 2.13.0+cu130, sgl-kernel 0.4.6.post1, FlashInfer 0.6.18, RTX 4090, BF16, ClickHouse 25.12.2.54. A matching native build and a version-isolated environment are required.

The original implementation report includes Qwen3-0.6B eager/graph/padded-graph/overlap and Qwen3-8B graph cells. The follow-up review independently reran only the bounded Qwen3-0.6B padded-graph cell; it did not independently reproduce all five cells.

The tested core baseline was c222f18c4db55f6f2d7136b1d6608c0b540b273a (the SGLang branch adds documentation/submodule changes on top). Current main has since advanced; this PR does not claim a full runtime requalification of those later core changes.

Review fixes included in the pinned integration

• Restore parent DMX/LD_PRELOAD environment after worker startup; prevent sequential engines from inheriting auto-generated capture IDs or per-engine overrides.
• Preserve failed flush status on both engine and worker; a repeated stop cannot turn a failed capture into an apparent success.
• Reject positional as well as keyword stream=True before generation.
• Check every storage row for the unique capture, rather than filtering away unexpected request IDs before validation.
• Correct initialization-order documentation and explicitly exclude concurrent engine construction in one process.

Validation

• Final CPU gate: 101 passed (16 new regressions).
• Runtime audit: 23 resolved targets, 0 errors / 0 audit warnings. Known optional bailing MoE imports lack vLLM; those architectures are out of scope.
• Independent stock/monitored GPU processes: Qwen3-0.6B, captured decode buckets [1, 2, 4], 3 requests, 8 greedy tokens each, overlap off, prefill graph off in both modes.
• Public output equality; 234 logprob values equal at zero tolerance, max difference 0.0.
• 744 exact-once stored rows, with token/range/type/argmax checks and a capture-wide storage query. Selected hooks: resid_pre,final_ln,token_ids,final_logits.
• git diff --check; integration commit fetched successfully from the public remote before pinning.

The GPU rerun covered environment/flush/storage fixes. The later positional-stream guard was covered by CPU rejection and nonstreaming-forwarding tests, not another GPU run. Full stock logits bitwise equality and all-hook capture are not claimed.

Evidence at the pinned integration commit:

• Follow-up review and reproduction commands
• Independent GPU comparator receipt
• Original port audit and limits

Limits

TP/PP/DP beyond the supported single-device topology, MoE/other architectures, speculative decoding, LoRA, PD disaggregation, TBO, torch.compile/tc_piecewise decode, and non-generation models are rejected. Prefill graph is forced eager. Serving, streaming, cache-hit/chunked-prefill interaction matrices, quantization and full-hook stress are not qualified. Overlap can capture one executed lookahead forward beyond public completion; disable overlap for exact public-step alignment.

No release tag is created and no automatic merge is requested.