llama_decode_eagle latency issue solved

[mkjsym]

Make sure to read the contributing guidelines before submitting a PR

llama_decode_eagle 함수의 latency issue를 해결하였습니다.

[Copilot]

Pull Request Overview

This PR adds support for an EAGLE input layer to address latency issues in the llama_decode_eagle function and updates related tensor loading, normalization, and RoPE settings. Key changes include:

• Introduce LLM_TENSOR_LAYER_INPUT_EAGLE and map it in tensor loading and architecture definitions.
• Comment out the redundant RMS normalization step in llm_build_eagle to reduce overhead.
• Adjust RoPE type mapping for the EAGLE architecture and update example integration under examples/speculative-eagle.

Reviewed Changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
src/llama-model.cpp	Added handling for LLM_TENSOR_LAYER_INPUT_EAGLE and skipped norm.
src/llama-arch.h	Defined new enum value LLM_TENSOR_LAYER_INPUT_EAGLE.
src/llama-arch.cpp	Updated tensor info for LLM_TENSOR_EMBD_FC to use the new layer.
examples/speculative-eagle/*	Added complete example and build integration for EAGLE decoding.
examples/CMakeLists.txt	Registered the new speculative-eagle example in the build.

Comments suppressed due to low confidence (4)

src/llama-arch.h:378

• The new enum value lacks a descriptive comment; add a brief explanation for LLM_TENSOR_LAYER_INPUT_EAGLE so its purpose is clear to future maintainers.

    LLM_TENSOR_LAYER_INPUT_EAGLE,

src/llama-arch.h:374

• New tensor layer functionality for EAGLE is not covered by existing tests; add unit tests for load_tensors and the EAGLE decode path to ensure correctness.

enum llm_tensor_layer {

examples/speculative-eagle/speculative-eagle.cpp:1

• [nitpick] Comments are written in Korean and English; for consistency and to accommodate a global contributor base, translate or unify comments in English.

//Tree-based EAGLE 구현 코드