---
title: "Yuri Pocepaev (@badbat4560) | Hashnode"
description: "Lead developer at Neuroprem | LLM systems & inference. Read the latest articles by Yuri Pocepaev on Hashnode."
url: "https://hashnode.com/@badbat4560"
---

@badbat4560

# Yuri Pocepaev

@badbat4560 · Joined September 2026

Lead developer at Neuroprem | LLM systems & inference

[Website](https://neuroprem.com/)

### [When Offloading Stalls Inference: Queue Capacity, CUDA Event Ownership, and Misleading Health Checks](https://badbat4560.hashnode.dev/when-offloading-stalls-inference-queue-capacity-cuda-event-ownership-and-misleading-health-checks?utm_source=hashnode&utm_medium=feed)

An inference server stopped generating tokens while more than 200 GiB of host memory remained available. Its health endpoint kept returning HTTP 200. The first visible exception was queue.Full inside

11 min read · [badbat4560.hashnode.dev](https://badbat4560.hashnode.dev/when-offloading-stalls-inference-queue-capacity-cuda-event-ownership-and-misleading-health-checks?utm_source=hashnode&utm_medium=feed)

[Oct 5](https://hashnode.com/posts/when-offloading-stalls-inference-queue-capacity-cuda-event-ownership-and-misleading-health-checks/6ac35f8a880537919b5bb9f2)

### [One Model, Many Roles: Specializing LLM Inference Without Training More Models](https://badbat4560.hashnode.dev/one-model-many-roles-specializing-llm-inference-without-training-more-models?utm_source=hashnode&utm_medium=feed)

A 40-request pilot tests focused tool sets, plus an architecture for chat, reasoning, code, and evidence roles sharing one model.

16 min read · [badbat4560.hashnode.dev](https://badbat4560.hashnode.dev/one-model-many-roles-specializing-llm-inference-without-training-more-models?utm_source=hashnode&utm_medium=feed)

[Sep 29](https://hashnode.com/posts/one-model-many-roles-specializing-llm-inference-without-training-more-models/6abc0307b28a18cdeba77e06)

### [Prevent Context Leakage in Multi-Tenant LLM Systems](https://badbat4560.hashnode.dev/prevent-context-leakage-in-multi-tenant-llm-systems?utm_source=hashnode&utm_medium=feed)

Practical tenant isolation for LLM apps: trusted identity, scoped memory and RAG, safe semantic caching, and cross-user regression tests.

8 min read · [badbat4560.hashnode.dev](https://badbat4560.hashnode.dev/prevent-context-leakage-in-multi-tenant-llm-systems?utm_source=hashnode&utm_medium=feed)

[Sep 29](https://hashnode.com/posts/prevent-context-leakage-in-multi-tenant-llm-systems/6abb12996ef26cad7bbdab5c)

### [Smaller Context, Recoverable History: Inside an Agent Memory Handoff](https://badbat4560.hashnode.dev/smaller-context-recoverable-history-inside-an-agent-memory-handoff?utm_source=hashnode&utm_medium=feed)

A memory handoff replaces a large conversation window while keeping retained source segments available for recall. The goal is to reduce how much history the LLM has to carry in its active context, wh

11 min read · [badbat4560.hashnode.dev](https://badbat4560.hashnode.dev/smaller-context-recoverable-history-inside-an-agent-memory-handoff?utm_source=hashnode&utm_medium=feed)

[2](https://hashnode.com/posts/smaller-context-recoverable-history-inside-an-agent-memory-handoff/6ab9cf57f3436455b835285b) [Sep 28](https://hashnode.com/posts/smaller-context-recoverable-history-inside-an-agent-memory-handoff/6ab9cf57f3436455b835285b)

### [Running Qwen Flash-Next NVFP4 in vLLM: PLE Loading, B12x Fixes, and Stable Inference](https://badbat4560.hashnode.dev/running-qwen-flash-next-nvfp4-in-vllm-ple-loading-b12x-fixes-and-stable-inference?utm_source=hashnode&utm_medium=feed)

A deployment walkthrough covering PLE loading, FlashInfer B12x route fixes, shared workspaces, and concurrent inference memory tuning.

17 min read · [badbat4560.hashnode.dev](https://badbat4560.hashnode.dev/running-qwen-flash-next-nvfp4-in-vllm-ple-loading-b12x-fixes-and-stable-inference?utm_source=hashnode&utm_medium=feed)

[Sep 26](https://hashnode.com/posts/running-qwen-flash-next-nvfp4-in-vllm-ple-loading-b12x-fixes-and-stable-inference/6ab7b1b697317aa06660975f)
