I have implemented speculative drafting via TCP/IP. 50%+ speedup of main model. Any interest in this for llama.cpp?

[xkmire]

I have implemented speculative drafting via model over TCP/IP using existing llamaserver API /completion that allows to send context as tokenids between models with the same token library. Speedup can be more than 50% in good situations (see test results below).

Good thing was that it was easier than I tought, and patch is fairly small. I am not especially familiar with github and doing PR and things. Can this be of interest?
How do I proceed?

BR

Description
CLI options for remote draft:
• --spec-type draft-remote (added draft-remote to existing --spec-type)
• --spec-draft-model-url (new)
• --spec-draft-model-name (new, optional, default empty)
• --spec-draft-api-key (new, optional, default empty)
• --spec-draft-context-limit (new, optional, default: 8192)
• --spec-draft-n-max (reused, optional - now works for both local and remote draft, default 3)
Test setup
Two computers with direct connection using a 10Gbit/s ethernet point to point cable. Main model is running on an RTX 6000 Pro Blackwell and draft model is running on a RTX 5090.

Main model: Devstral-2-123B-Instruct-2512-Q4_K_M (KV: q8_0, ctx_length: 128k), standard 15-20 tokens/s slowing to below 10 on large contexts
Draft model: Ministral-3-3B-Instruct-2512-Q4_K_M (KV: q8_0, ctx_length: 8k), 300-400 tokens/second
Result
Programming task writing a python program:
draft acceptance = 0.81300 ( 2013 accepted / 2476 generated), mean len = 4.5
Speedup: 25-30 tokens per second, increasing speed 50%+. When context grows, speedup increases since main model slows a lot.
For natural language tasks, acceptance drops to 0.3-0.5 depending on question and speed up is between 0-20% only.

My test commands
draft model start:
/home/1/llama.cpp/build/bin/llama-server
--model /home/mir/services/models/Ministral-3-3B-Instruct-2512-Q4_K_M.gguf
-fa on
-ctk q8_0
-ctv q8_0
-c 8192
-ngl 99
-b 1024
-ub 1024
--jinja
--kv-unified
--no-mmap
--parallel 4
--threads 8
--temp 0.0
--top-k 1
--host 0.0.0.0
--port 3535

main model start:
/home/2/llama.cpp/build/bin/llama-server
--model /home/mir/services/models/Devstral-2-123B-Instruct-2512-Q4_K_M-00001-of-00002.gguf
-fa on
-ctk q8_0
-ctv q8_0
-c 131072
--no-mmap
--parallel 4
--jinja
--threads 16
--spec-type draft-remote
--spec-draft-model-url http://x.x.x.x:3535/v1
--spec-draft-n-max 5
--port 3536

diff.patch