eagle3: add qwen3.5 4B 9B 35B-A3B support

[36330]

Summary

This PR extends the EAGLE3 implementation with recurrent verification-state support.

Compared with the earlier EAGLE3 work, this version keeps target-side verification in a single batched decode path and adds the recurrent state handling needed to make that flow work correctly for hybrid / recurrent models.

Main changes:

• add EAGLE3 recurrent round-state APIs
• preserve single-batch target verification in speculative-simple
• promote the accepted recurrent depth state directly after verification
• keep KV cleanup on the existing seq_rm() path
• add related Qwen3.5 / Qwen3.5-MoE integration
• include small converter / model wiring updates needed by this flow

Details

Core files changed:

• examples/speculative-simple/speculative-simple.cpp
• include/llama.h
• src/llama-context.cpp
• src/llama-memory.h
• src/llama-memory-hybrid.h
• src/llama-memory-hybrid.cpp
• src/llama-memory-recurrent.h
• src/llama-memory-recurrent.cpp
• src/models/qwen35.cpp
• src/models/qwen35moe.cpp

Additional related updates:

• convert_hf_to_gguf.py
• examples/speculative/speculative.cpp
• src/models/qwen3vl-moe.cpp

test examples:
Qwen3.5-9B-BF16.gguf
eagle3-qwen3.5-9b-eagle.gguf

draft1 accept = 61.463%
encoded 26 tokens in 0.390 seconds, speed: 66.656 t/s
decoded 259 tokens in 15.686 seconds, speed: 16.512 t/s
无eagle3
[ Prompt: 66.2 t/s | Generation: 9.8 t/s ]
speedup:1.68x

Qwen3.5-4B-Q4_K_M.gguf
eagle3-qwen35-4b-draft-Q4_K_M.gguf

draft2 accept = 53.140%
encoded 26 tokens in 0.080 seconds, speed: 326.052 t/s
decoded 257 tokens in 3.047 seconds, speed: 84.339 t/s
无eagle3
[ Prompt: 437.6 t/s | Generation: 64.2 t/s ]
speedup:1.31x

[ggml-gh-bot]

Hi @36330, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• Large PR: Large changes require prior discussion (e.g. an issue or RFC) and maintainers may not be able to review this PR as-is. Consider splitting it into smaller, focused PRs.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[sorasoras]

how does this compare to MTP?

[jacekpoplawski]

"extends the EAGLE3 implementation" - what does it mean? Has EAGLE3 ever been implemented in llama.cpp?

[36330]

EAGLE3 has been implemented in llama.cpp. However, due to the particularity of the qwen3.5 linear attention architecture, some adaptations have been made.

[ngxson]

You're pushing to wrong branch, I suppose this is not yet ready

[36330]

I successfully merged the latest version.

[jacekpoplawski]

EAGLE3 has been implemented in llama.cpp

this looks like a draft
#18039
(please correct me if I am wrong)

[36330]

EAGLE3 has been implemented in llama.cpp

this looks like a draft #18039 (please correct me if I am wrong)

qwen3.5 not support

[jacekpoplawski]

EAGLE3 has been implemented in llama.cpp

this looks like a draft #18039 (please correct me if I am wrong)

qwen3.5 not support

maybe you wanted to push into ichbinhandsome:eagle3-adapt-new-arch instead into master?