[wai-bao]

Summary

Unifies the four per-site decode row-construction copies (MTP validate, MTP draft-extend, DFlash query, fixed/dflash2 verify) into a single decode_input_builder with a declarative plan, and consolidates the shared seq-lens / paged-KV layout knowledge that those copies each reimplemented.

• Deletes spec_input_builder (852 lines) in favor of framework/batch/decode_input_builder.{h,cpp}: RowSpec rows, DecodeLayout (flat / chunked / block-parallel / chunked-decode), plan-driven commit (update_block_table / commit_decode_input / commit_decode_geometry), and a named expand_idle_rank_block entry for idle DP-rank synthetic blocks.
• Adds kv_cache/layout/seq_lens_layout.h: the backend seq-lens layout policy (NPU per-seq vs cu_seqlens), single source for seq_len_at, append_seq_len(s), append_qo_indptr, seq_lens_capacity — replaces three get_q_seq_len copies and per-builder build_q_cu_seq_lens_vec copies.
• Adds kv_cache/layout/paged_kv_layout.h: row-level paged-KV geometry (kv_page_count, kv_last_page_len, append_paged_kv_entries, slot_for_position, compute_slot_mapping) shared by the batch builders, the decode-row builder, the expanded-decode metadata builder, and the DSA metadata builders.
• Merges the two ~90-line update_decode_step_input copies (MTP / DFlash) behind ExpectedKvPolicy in mtp_runtime_helpers, with the chunked-verify clamp parameters carried inside the policy.
• Hot-path wins along the way: cached device index tensors (CachedDeviceIndices) for draft-query / draft-extend / validate indices, model_type resolved once per worker instead of re-parsing config.json per predicate call, per-sequence hoists in draft-prefix bounds, and the per-step result.lengths.to(int32) cast removed from the verify commit.

Net −276 lines across 53 files.

Testing

• decode_input_builder_test (21), task_execution_pipeline_speculative_test (152), mtp_decode_input_test (25), mtp_decode_metadata_test (18), schedule_overlap_test (42), batch_test (134) all pass on NPU.

[netlify]

✅ Netlify Deploy Preview ready

Built without sensitive environment variables

Name	Link
🔨 Latest commit	f3e7ffb
🔍 Latest deploy log	https://app.netlify.com/projects/xllm-docs-preview/deploys/6aec9fb656a2976df8a60454
😎 Deploy Preview	https://deploy-preview-2640--xllm-docs-preview.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.
🤖 Make changes	Run an agent on this branch

---

To edit notification comments on pull requests, go to your Netlify project configuration.