[Septend-fun]

Overview

Add conversion and inference support for Spark-X2.6 DSpark drafts.

Spark-X2.6 shares the Spark2.5 architecture and uses the existing Spark2_5ForCausalLM conversion and spark2_5 runtime.

• Export the embedded draft_model.* weights with --dspark, and exclude them from normal target conversion.
• Support feature normalization before fusion, headwise sigmoid attention gates, GELU FFN, and per-layer full/sliding-window RoPE.
• Reuse the existing DSpark Markov head, confidence head, and target verification flow.

Additional information

Validation with the supplied checkpoint:

• Built llama-server, llama-completion, and test-llama-archs.
• Verified all 59 exported draft tensors against the original checkpoint.
• Compared draft outputs with an independent PyTorch FP32 reference on CPU and CUDA: 24/24 top-1 matches.
• With greedy sampling and f16 KV cache, all 12 measured responses matched target-only token sequences.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. Codex assisted in verifying the correctness of the DSpark draft integration.