Aleksei Romanov (@gfactor)

Quantization, Explained: From FP32 to INT4, and What Each Precision Costs You

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 readgfactor.hashnode.dev

6d ago

Teaching a 27B Model to Write Triton Kernels: SFT + RL Gym

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 readgfactor.hashnode.dev

Nov 3

GPU Kernels, Explained: CUDA, Triton, cuBLAS, torch.compile and KernelBench

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 readgfactor.hashnode.dev

Nov 3

Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning

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 readgfactor.hashnode.dev

Oct 9

Pushing the Limits: Extreme Inference Speedup of Qwen 3.8 27B on NVIDIA B300 (100 to 10k+ tok/s)

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 readgfactor.hashnode.dev

Oct 7

Aleksei Romanov

About

Available for

Aleksei Romanov's blogs

Quantization, Explained: From FP32 to INT4, and What Each Precision Costs You

Teaching a 27B Model to Write Triton Kernels: SFT + RL Gym

GPU Kernels, Explained: CUDA, Triton, cuBLAS, torch.compile and KernelBench

Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning

Pushing the Limits: Extreme Inference Speedup of Qwen 3.8 27B on NVIDIA B300 (100 to 10k+ tok/s)

Aleksei Romanov

About

Available for

Aleksei Romanov's blogs

Quantization, Explained: From FP32 to INT4, and What Each Precision Costs You

Teaching a 27B Model to Write Triton Kernels: SFT + RL Gym

GPU Kernels, Explained: CUDA, Triton, cuBLAS, torch.compile and KernelBench

Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning

Pushing the Limits: Extreme Inference Speedup of Qwen 3.8 27B on NVIDIA B300 (100 to 10k+ tok/s)

Aleksei Romanov

About

Available for

Aleksei Romanov's blogs

Quantization, Explained: From FP32 to INT4, and What Each Precision Costs You

Teaching a 27B Model to Write Triton Kernels: SFT + RL Gym

GPU Kernels, Explained: CUDA, Triton, cuBLAS, torch.compile and KernelBench

Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning

Pushing the Limits: Extreme Inference Speedup of Qwen 3.8 27B on NVIDIA B300 (100 to 10k+ tok/s)

Search Hashnode

Aleksei Romanov

About

Available for

Aleksei Romanov's blogs

Quantization, Explained: From FP32 to INT4, and What Each Precision Costs You

Teaching a 27B Model to Write Triton Kernels: SFT + RL Gym

GPU Kernels, Explained: CUDA, Triton, cuBLAS, torch.compile and KernelBench

Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning

Pushing the Limits: Extreme Inference Speedup of Qwen 3.8 27B on NVIDIA B300 (100 to 10k+ tok/s)