Support public K2-Horizon base models in SGLang #49

Public source requirements

Use the public Hugging Face artifacts below for all K2-Horizon work in this issue.
These identities define the model consumer. They do not add model bytes, runtime execution, or another owner to a source-only issue.
These requirements supersede earlier internal checkpoints, source revisions, private locators, model aliases, and conversion requirements for this campaign.

The source catalog and artifact cookbooks bind the accepted 2026-09-13 source epoch.
Use their exact content identities, manifests, input roles, and loader requirements.
A changed upstream head requires a separately reviewed source epoch.
A local cache or storage path is a physical input. It does not identify the model.

Earlier internal checkpoint tests, image runs, and qualification results do not qualify these public bytes.
Repeat the applicable source, engine, topology, request, and cleanup checks for each selected public identity.
Keep model bytes unchanged and outside Git. Keep admitted model inputs read-only.
Keep remote checkpoint code disabled. Use reviewed installed runtime support.

Keep native-8192 and yarn-131072 as separate checked runtime identities.
They can use the same pinned source configuration bytes. #860 owns native qualification, and #871 owns YaRN qualification.

Problem and scope

SGLang needs a complete native loading and serving contract for the six public K2-Horizon base artifacts.
Use the pinned public inputs above and synthetic fixtures derived from their documented configuration and tensor shapes.
No private validation package is required for the public source contract.

Required behavior and acceptance

• Provide one native K2HorizonForCausalLM model entry point and installed K2 configuration support.
• Derive dense, routed-expert, shared-expert, and MoVA behavior from typed source fields.
• Support grouped normalization and partial rotary embeddings where the accepted public configuration requires them.
• Keep MoVA value routing distinct from MLP expert routing.
• Test dense, routed-expert, shared-expert, and pipeline-partition weight loading.
• Test the applicable normalization and rotary forward behavior.
• Support and test tensor parallelism, expert parallelism, and DeepEP selection for the selected topology.
• Preserve ordinary text, reasoning content, and structured tool calls in streaming and non-streaming responses.
• Test malformed calls and streaming boundaries without inventing calls or losing visible content.
• Reject contradictory or unsupported configurations before weight loading and GPU allocation.
• Load configuration with remote checkpoint code disabled and network access disabled after staging.
• Run fresh H200 qualification for every required public architecture variant.
• Bind results to the public source revision, runtime revision, immutable image, topology, probe set, and cleanup evidence.
• Publish sanitized receipt identities and outcomes. Public model names, revisions, and documented architecture facts can appear in tests and reports.
• Keep credentials, private storage paths, hosts, endpoints, and private runtime data out of public records.

Dependencies and boundaries

LLM360/comet#812 supplies Comet's immutable compatibility inputs.
LLM360/comet#811 owns Comet image admission and full service qualification.
LLM360/sglang#50 owns published FP8 formats.
Uno remains under pairing and diffusion decoding.
Do not change model bytes, publish a Comet route, or claim Comet qualification from a standalone engine test.