llama.cpp VS mlc-llm

> 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

It depends on what you mean by "this." MLC's catch is that you need to define/compile models for it with TVM. Here is the list of supported model architectures: https://github.com/mlc-ai/mlc-llm/blob/main/python/mlc_llm/m...

llama.cpp has a much bigger supported model list, as does vLLM and of course PyTorch/HF transformers covers everything else, all of which work w/ ROCm on RDNA3 w/o too much fuss these days.

For inference, the biggest caveat is that Flash Attention is only an aotriton implementation, which besides being less performant sometimes, also doesn't support SWA. For CDNA there is a better CK-based version of FA, but CK doesn't not have RDNA support. There are a couple people at AMD apparently working on native FlexAttention, os I guess we'll how that turns out.

(Note the recent SemiAccurate piece was on training, which I'd agree is in a much worse state (I have personal experience with it being often broken for even the simplest distributed training runs). Funnily enough, if you're running simple fine tunes on a single RDNA3 card, you'll probably have a better time. OOTB, a 7900 XTX will train at about the same speed as an RTX 3090 (4090s blow both of those away, but you'll probably want more cards and VRAM of just move to H100s).

This one uses gpu, it doesn't support Mistral yet: https://github.com/mlc-ai/mlc-llm

I have tried running mistral 7B with MLC on my m1 metal. And it kept crushing (git issue with description). Memory inefficiency problems.

Project: https://github.com/mlc-ai/mlc-llm

For LLM inference, a shoutout to MLC LLM, which runs LLM models on basically any API that's widely available: https://github.com/mlc-ai/mlc-llm

One of your problems might be that gfx1032 is not supported by AMD's ROCm packages, which has a laughably short list of supported hardware: https://rocm.docs.amd.com/en/latest/release/gpu_os_support.h...

The normal workaround is to assign the closest architecture, eg gfx1030, so `HSA_OVERRIDE_GFX_VERSION=10.3.0` might help

Also, it looks like some of your tested projects are OpenCL? For me, I do something like: `yay -S rocm-hip-sdk rocm-ml-sdk rocm-opencl-sdk` to cover all the bases.

My recent interest has been LLMs and this is my general step by step for those (llama.cpp, exllama) for those interested: https://llm-tracker.info/books/howto-guides/page/amd-gpus

I didn't port the docs back in, but also here's a step-by-step w/ my adventures getting TVM/MLC working w/ an APU: https://github.com/mlc-ai/mlc-llm/issues/787

From my experience, ROCm is improving, but there's a good reason that Nvidia has 90% market share even at big price premiums.

Maybe they're talking about https://github.com/mlc-ai/mlc-llm which is used for web-llm (https://github.com/mlc-ai/web-llm)? Seems to be using TVM.

you already have TVM for the cross platform stuff

see https://tvm.apache.org/docs/how_to/deploy/android.html

or https://octoml.ai/blog/using-swift-and-apache-tvm-to-develop...

or https://github.com/mlc-ai/mlc-llm