feat: add Kimi K3 DSpark disaggregated reproduction

[maocheng23]

Motivation

Migrate the prior four-node colocated Kimi K3 V1C continual DSpark run to current SpecForge as a two-node disaggregated job: one TP8 capture server plus one four-rank FSDP trainer.

The recipe preserves the observable training contract: five auxiliary layers, weights-only warm start, regenerated prompt order, effective global batch 128, constant LR, DSpark loss weights, 64K context, and checkpoint semantics.

Modifications

• Add configurable LR schedules, including the constant schedule required by the reproduced run, with checkpoint compatibility tests.
• Add an independent prompt_seed so online prompt order is reproducible without coupling it to model/trainer RNG.
• Add dedicated DSpark server capture wiring and a versioned SGLang patch with Mooncake metadata transport.
• Validate the full K3 patch against the current public K3 SGLang branch tip 9acd9cba39f522da71c6d9b9695f2ceb41d36b18; retain the older target name as a compatibility alias.
• Carry the K3-specific 64K correctness guards: int64 token offsets, scale-stable residual scoring, and the old-fork Marlin grid.y fallback.
• Add the typed Kimi K3 V1C config and an executable two-node disaggregated recipe, including routable Mooncake bind settings, five-minute leases, watermarks, and a one-update smoke procedure.

Upstream SGLang main currently has no kimi_k3.py; this recipe therefore targets the latest public K3 serving branch rather than claiming upstream-main K3 support.

Related Issues

N/A

Accuracy Test

Completed:

• 94 targeted unit/integration tests: PASS.
• Repository pre-commit suite: PASS.
• K3 SGLang patch git apply --check against 9acd9cb: PASS.
• RunPod trainer-side config plan: PASS (one producer + four FSDP consumers).
• Prepared data verification: 4 smoke rows and 462 full rows; full-data SHA-256 6d50e6bb9ee59095eed91bfba035081efef9fea43bece9ad5dd01c6648a8ef24.

Pending before marking ready:

• TP8 64K capture -> Mooncake -> four-rank finite optimizer update -> checkpoint E2E smoke.
• Full-run loss/accuracy/tau trajectory comparison with the prior V1C run.

Benchmark & Profiling

No performance claim yet. Throughput and numerical-parity evidence will be added after the RunPod E2E gate.

Checklist

• Format code with the repository pre-commit configuration.
• Add unit tests for scheduler, prompt ordering, capture contracts, topology, and model loading.
• Update documentation and add a complete example recipe.
• Add RunPod E2E accuracy and throughput evidence before requesting final review.
• No credentials are stored in config or documentation.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.