Nottlespike

DeepSeek 3FS on a single Ubuntu 26.04 LTS machine (soft-RoCE, no RDMA NIC)

How to run DeepSeek 3FS all-on-one-node on Ubuntu 26.04 LTS with an ordinary Ethernet NIC, using soft-RoCE (rdma_rxe). 3FS only documents Ubuntu 20.04/22.04; this is the full working path on a current LTS, including the two build-container gotchas that are not written down anywhere.

Tested on: AMD B550 board, Ryzen (32 threads), 2× NVMe + SATA SSD, 91 GB RAM, single Realtek RTL8125 2.5GbE (no RDMA hardware), Ubuntu 26.04.1 LTS, kernel 7.0.0-34-generic.

Why this shape

• 3FS's single-machine entry point is tests/fuse/run.sh (see upstream issue #97): it starts FoundationDB (memory mode) + mgmtd + meta + storage + FUSE client on one host with four file-backed targets and a replica-1 chain table.
• The official build toolchain is clang-14, which no longer exists in Ubuntu 24.04+ repos — so build inside the repo's ubuntu:22.04 dev container and run the binaries natively. Binaries built a

Nottlespike

#!/usr/bin/env bash
# setup-agent-guides.sh
# ---------------------
# Bootstraps agent guide files for a repository.
#
# This script is intended for "new repo setup" and repeat-safe updates: 
# 1) Reads a source guide file (`.md` by default, or a provided path)
# 2) Copies it to `AGENTS.md` (canonical guide file)
# 3) Creates `CLAUDE.md` and `GEMINI.md` aliases to `AGENTS.md`
# 4) Falls back to plain file copies if symlinks are unavailable

Nottlespike

#!/usr/bin/env bash
set -euo pipefail

# update_agents.sh
# Regenerates agents.md by sending (existing agents.md + new info) to Claude using a fixed system prompt.
#
# Requirements:
#   - bash, curl, jq
#   - ANTHROPIC_API_KEY environment variable set
#

Nottlespike

// paged_attention.metal - Paged Attention for Apple Silicon
//
// Implements vLLM-style paged attention adapted for Metal simdgroup architecture.
// Paged attention decouples logical token positions from physical memory layout,
// enabling efficient batch serving with variable-length sequences.
//
// Key differences from flash_attention.metal:
//   - KV cache is organized in fixed-size blocks (pages)
//   - Block tables map logical block indices to physical block addresses
//   - Each sequence can have a different context length

Nottlespike

"""
Accurate comparison of MLX nvfp4 vs NVIDIA NVFP4 implementation.

Key architectural difference:
NVIDIA NVFP4 uses a TWO-LEVEL SCALING strategy:
  1. Global FP32 per-tensor scale: s_enc = (6 * 448) / tensor_amax
  2. Local E4M3 per-block scale: one scale per 16 elements

MLX appears to use only single-level E4M3 block scaling without the FP32 tensor scale.

Nottlespike

# P100 Extended Implementation Tasks - 100+ tasks for full agent utilization
# Tesla P100 (GP100) - 56 SMs, 3584 CUDA cores, 16GB HBM2 @ 732 GB/s

tasks:
  # ============================================
  # CUDA Kernels (P0 - Critical) - 20 tasks
  # ============================================
  
  - name: kernel-vecadd-sm60
    prompt: |

Nottlespike

                     Patchy 2.0 License with OpenAI and California State Exclusion
                           Version 2.0, May 2024

TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION

1. Definitions.

   "License" shall mean the terms and conditions for use, reproduction,
   and distribution as defined by Sections 1 through 9 of this document.