---
title: "Aleksei Romanov (@gfactor) | Hashnode"
description: "Ex-Apple GenAI engineer, building an LLM post-training platform at g factor technologies. Read the latest articles by Aleksei Romanov on Hashnode."
url: "https://hashnode.com/@gfactor"
---

@gfactor

# Aleksei Romanov

@gfactor · United States · Joined September 2026

Ex-Apple GenAI engineer, building an LLM post-training platform at g factor technologies

[Website](https://www.g-ftech.com/)

### [Quantization, Explained: From FP32 to INT4, and What Each Precision Costs You](https://gfactor.hashnode.dev/quantization-explained-from-fp32-to-int4-and-what-each-precision-costs-you?utm_source=hashnode&utm_medium=feed)

In LoRA & DoRA and our other posts on large models, we keep running into the same wall: memory. This post is about the number formats behind that wall. When you run a modern language model, memory is

14 min read · [gfactor.hashnode.dev](https://gfactor.hashnode.dev/quantization-explained-from-fp32-to-int4-and-what-each-precision-costs-you?utm_source=hashnode&utm_medium=feed)

[6d ago](https://hashnode.com/posts/quantization-explained-from-fp32-to-int4-and-what-each-precision-costs-you/6aed0294c8a8a0842336cfe1)

### [Teaching a 27B Model to Write Triton Kernels: SFT + RL Gym](https://gfactor.hashnode.dev/teaching-a-27b-model-to-write-triton-kernels-sft-rl-gym?utm_source=hashnode&utm_medium=feed)

In Part 1, we explored how GPU kernels work under the hood — why operator fusion and memory tiling matter, and how benchmarks like KernelBench evaluate whether an LLM can write kernels that actually b

27 min read · [gfactor.hashnode.dev](https://gfactor.hashnode.dev/teaching-a-27b-model-to-write-triton-kernels-sft-rl-gym?utm_source=hashnode&utm_medium=feed)

[Nov 3](https://hashnode.com/posts/teaching-a-27b-model-to-write-triton-kernels-sft-rl-gym/6aea252b0aa43068615c154e)

### [GPU Kernels, Explained: CUDA, Triton, cuBLAS, torch.compile and KernelBench](https://gfactor.hashnode.dev/gpu-kernels-explained-cuda-triton-cublas-torch-compile-and-kernelbench?utm_source=hashnode&utm_medium=feed)

When you run a neural network in PyTorch, Python is only the coordinator. Behind every tensor operation — from a simple activation to a multi-head attention block — sits one or more GPU kernels: speci

28 min read · [gfactor.hashnode.dev](https://gfactor.hashnode.dev/gpu-kernels-explained-cuda-triton-cublas-torch-compile-and-kernelbench?utm_source=hashnode&utm_medium=feed)

[Nov 3](https://hashnode.com/posts/gpu-kernels-explained-cuda-triton-cublas-torch-compile-and-kernelbench/6ae986128db764f898d76a12)

### [Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning](https://gfactor.hashnode.dev/multi-reward-rl-part-3-gdpo-cispo-repo-r-at-27b-and-the-advantage-floor-that-stopped-learning?utm_source=hashnode&utm_medium=feed)

Qwen3.8-27B, 600 steps, 12 reward channels: CISPO, REPO-R and prompt filtering lifted holdout from −0.47 to +1.80, and an advantage floor silently zeroed 96% of the signal.

15 min read · [gfactor.hashnode.dev](https://gfactor.hashnode.dev/multi-reward-rl-part-3-gdpo-cispo-repo-r-at-27b-and-the-advantage-floor-that-stopped-learning?utm_source=hashnode&utm_medium=feed)

[Oct 9](https://hashnode.com/posts/multi-reward-rl-part-3-gdpo-cispo-repo-r-at-27b-and-the-advantage-floor-that-stopped-learning/6ac929e55952e02b451b58c9)

### [Pushing the Limits: Extreme Inference Speedup of Qwen 3.8 27B on NVIDIA B300 (100 to 10k+ tok/s)](https://gfactor.hashnode.dev/pushing-the-limits-extreme-inference-speedup-of-qwen-3-8-27b-on-nvidia-b300-100-to-10k-tok-s?utm_source=hashnode&utm_medium=feed)

How tensor parallelism, MTP speculative decoding, prefix caching and suffix decoding move Qwen 3.8 27B from 104 tok/s to nearly 10,000 tok/s on our stack, and where Fireworks AI went further.

22 min read · [gfactor.hashnode.dev](https://gfactor.hashnode.dev/pushing-the-limits-extreme-inference-speedup-of-qwen-3-8-27b-on-nvidia-b300-100-to-10k-tok-s?utm_source=hashnode&utm_medium=feed)

[Oct 7](https://hashnode.com/posts/pushing-the-limits-extreme-inference-speedup-of-qwen-3-8-27b-on-nvidia-b300-100-to-10k-tok-s/6ac67afe28ee0bceeb9e0f74)