llama.cpp VS unsloth

> 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

Training: unsloth + trl (SFTTrainer). Unsloth handles the 4-bit quantization and gradient checkpointing; trl handles the training loop.

pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git" pip install --no-deps xformers trl peft accelerate bitsandbytes

Here's the thing that made this experiment actually feasible: Unsloth.

For the actual training, I rely on Unsloth. It makes fine-tuning 2x faster and uses 70% less memory.

Unsloth - Efficient fine-tuning

Oh for finetuning - we do have some code for MoE finetuning for Qwen at https://github.com/unslothai/unsloth, but we haven't yet announced it yet!

Lot's of tools for each of those separately (RAG and fine-tuning). We're working on combining them but it's not ready yet.

You don't need a big GPU cluster. Fine-tuning is quite accessible via both APIs and local tools. Some suggestions:

- getkiln.ai (biased, my tool): let's you try all of the below, and compare/eval the resulting models

- API based tuning for closed models: OpenAI, Google Gemini

- API based tuning for open models: Together.ai, Fireworks.ai

- Local tuning for open models: https://unsloth.ai (can be run on Google Collab instances if you don't have local Nvidia GPUs).