feat(pytorch): support Kimi K2.6

[qescccczmr]

Motivation

Add Kimi-K2.6 support to the PyTorch backend, including text and multimodal inference, compressed-tensors W4A16 MoE, EAGLE3 speculative decoding, and serving integration.

The implementation supports Kimi-K2.6's MLA attention, MoE routing, vision encoder, multimodal projector, and external EAGLE3 draft model.

Changes

• Add Kimi-K2.6 configuration, model loading, native language model, vision tower, and multimodal projector.
• Add compressed-tensors W4A16 MoE model construction, weight loading, and Triton kernels.
• Add Kimi-K2.6 EAGLE3 speculative decoding with shared target embeddings and auxiliary hidden states.
• Add coverage for compressed-tensors W4A16 kernels, MoE routing, FA3 attention, model loading, and serving.

Acc Benchmarks

Academic

Dataset	Metric	LMDeploy	HF reference
GPQA Diamond	Accuracy, TP8, 4-run average, CompassAcademic	91.67%	90.50%
AIME 2026	Accuracy, TP8	97.60%	96.40%

LMDeploy and Hugging Face were evaluated with TP8 under aligned evaluation settings.

LMDeploy reaches or slightly exceeds the Hugging Face reference on both academic benchmarks, with no observed text accuracy regression.

Multimodal

Dataset	Metric	LMDeploy	HF reference
MathVision	Accuracy	84.54%	87.40%

Kimi-K2.6 successfully completes the MathVision multimodal evaluation.

Acceptance of EAGLE3

Draft model:

lightseekorg/kimi-k2.6-eagle3.1-mla

Aggregate results

Engine	Mean accept length	Strict draft acceptance
LMDeploy	2.93204	64.4015%
SGLang	2.93131	64.3769%
LMDeploy − SGLang	+0.00074	+0.0246 pp

Per-position acceptance rates:

Engine	Position 0	Position 1	Position 2
LMDeploy	81.43%	63.41%	48.36%
SGLang	81.52%	63.21%	48.40%

Runtime Validation

• Kimi-K2.6 TP8 text inference passed on 8 NVIDIA H200 GPUs.
• Kimi-K2.6 TP8+EP8 text inference completed model loading, warmup, and generation successfully.
• EAGLE3 completed all 30 AIME24 requests without engine errors.
• MathVision multimodal evaluation completed successfully.

Assistance

Assisted with Codex, reviewed manually.

[grimoire]

You don't have to patch the builder of dsv2. You can just remove 'kimi_k2' in it's condition.

[qescccczmr]

I've removed kimi_k2 from DeepseekV2ModelConfigBuilder.

[grimoire]

LGTM

[grimoire]

Please fix lint