How to reproduce the Needle benchmark numbers on Cactus?

Hi Cactus team,

I am trying to reproduce the Needle performance numbers mentioned on the Hugging Face page:

https://huggingface.co/Cactus-Compute/needle

It says:

> In production, Needle runs on Cactus at 6000 toks/sec prefill and 1200 decode speed.

I tested Needle on the `align_needle_inference_logic` branch with the Cactus old-style INT4 weights format.

My setup:

- Device: NVIDIA Jetson Thor / ARM Linux
- Branch: `align_needle_inference_logic`
- Commit: `71fc11bc`
- Model format: `config.txt + *.weights`
- Quantization: INT4 weights, FP16 compute

My measured speed is around:

```text
prefill: ~1200 tok/s
decode: ~100-140 tok/s
```

Could you share the exact setup used to get the reported numbers?
Specifically:
Which device/hardware was used?
Which Cactus branch/commit was used?
Which model weights were used?
What benchmark command or script was used?
What input length and output length were measured?
Was this CPU-only, GPU, Apple NPU/ANE, Android NPU, or another backend?
Thanks，I look forward to hearing from you!

Start with the Hugging Face Needle page and the align_needle_inference_logic branch at commit 71fc11bc, then compare the reported figures with the old-style INT4 weights setup described here. Document the hardware, weights, benchmark command or script, input and output lengths, and backend used so the 6000 tok/s prefill and 1200 tok/s decode numbers can be reproduced or explained.

Written by the indexing model from the issue text.