llama.cpp VS exllamav2

> The benchmark prompt was:

> Write a compact Python function that parses a unified diff and returns the changed file paths. Then explain two edge cases.

> Each benchmark generated about 128 tokens.

Generating 128 tokens is probably not enough for good benchmark results. MTP speedup depends on how often the predicted tokens are accepted. In my experience, the very early output has a higher acceptance rate, so short testing can give false positive speedups.

Also llama.cpp includes a tool specifically for benchmarking:

https://github.com/ggml-org/llama.cpp/blob/master/tools/llam...

In my build, MTP came from mainline llama.cpp, not ik_llama. ik_llama got me to ~47 (engine + quant), but I couldn't get MTP running there — my build rejected the -mtp flags and ignored the model's nextn tensors. Mainline llama.cpp added MTP fairly recently (PR #22673, merged 2026-05-16), and that's where it worked for me. (There may well be an ik_llama path I missed — this is just what got it going on my box.)

That script grew up. Today I'm releasing LlamaStash, the first public release of a fast, cross-platform, terminal-native launcher for llama.cpp with zero overhead.

LlamaStash spawns the unmodified upstream llama-server. So three different questions follow from that, and there is a benchmark suite for each.

llama.cpp includes a benchmarking tool called llama-bench https://github.com/ggml-org/llama.cpp/blob/master/tools/llam...

ik_llama includes llama-sweep-bench https://github.com/ikawrakow/ik_llama.cpp/blob/main/examples...

When comparing hardware, the output of these tools is very helpful to let others put it into context. The post says the output is "reading speed" but knowing the prefill and token generation speeds would be a lot more helpful.

lol the same way we implement all of the reduced precision fp8, fp4 types today: by storing them in the corresponding uint:

https://github.com/ggml-org/llama.cpp/discussions/15095

# Build llama.cpp with Metal backend git clone https://github.com/ggml-org/llama.cpp cd llama.cpp && cmake -B build -DGGML_METAL=ON && cmake --build build -j # Community-quantized GGUFs (Google ships safetensors; unsloth ships GGUF) huggingface-cli download unsloth/gemma-4-31B-it-GGUF \ gemma-4-31B-it-Q4_K_M.gguf --local-dir . huggingface-cli download unsloth/gemma-4-26B-A4B-it-GGUF \ gemma-4-26B-A4B-it-Q4_K_M.gguf --local-dir . # Benchmark: 200 generations of 512 tokens, log per-call timing ./build/bin/llama-bench -m gemma-4-31B-it-Q4_K_M.gguf -n 512 -r 200 -o json > dense.json ./build/bin/llama-bench -m gemma-4-26B-A4B-it-Q4_K_M.gguf -n 512 -r 200 -o json > moe.json

Today, I’m going to focus on my most popular project, TabbyAPI. TabbyAPI is a python based FastAPI server that allows users to interact with Large Language Models (or LLMs) using the ExllamaV2 library and adheres to the OpenAI API specification.

There are already some implementations out there which attempt to accomplish this!

Here's an example: https://github.com/silphendio/sliced_llama

A gist pertaining to said example: https://gist.github.com/silphendio/535cd9c1821aa1290aa10d587...

Here's a discussion about integrating this capability with ExLlama: https://github.com/turboderp/exllamav2/pull/275

And same as above but for llama.cpp: https://github.com/ggerganov/llama.cpp/issues/4718#issuecomm...

Sorry, I'm somewhat familiar with this term (I've seen it as a model loader in Oobabooga), but still not following the correlation here. Are you saying I should instead be using this project in lieu of llama.cpp? Or are you saying that there is, perhaps, an exllamav2 "extension" or similar within llama.cpp that I can use?

EDIT: I reported the bug to the exllamav2 Github. It's actually already fixed, just not on any current built release.

install exllamav2 from git with pip install git+https://github.com/turboderp/exllamav2.git. Make sure you have flash attention 2 as well.

Recent releases for exllamav2 brings working fp8 cache support, which I've been very excited to test. This feature doubles the maximum context length you can run with your model, without any visible downsides.

Without batching, I was actually thinking that's kind of modest.

ExllamaV2 will get 48 tokens/s on a 4090, which is much slower/cheaper than an H100:

https://github.com/turboderp/exllamav2#performance

I didn't test codellama, but the 3090 TI figures are in the ballpark of my generation speed on a 3090.

First, you need the convert.py script from turboderp's Exllama2 repo. You can read all about the convert.py arguments here.