alexellis

RiNGSiDE TP4: cold 35K exact-key copy error

Observed on 27 September 2026. This concerns othexmr's RiNGSiDE NVFP4 TP4 recipe. This is a single observed functional failure, not a RigMark failure, crash, proven kernel bug, or established failure rate. The cause is unresolved; temperature zero does not establish bitwise determinism for this runtime.

Observed

alexellis

Practical fixes that actually work, drawn from the same NVIDIA forum threads, Reddit reports, GitHub issues, and dual-Spark recipes people are using right now.

1. Kill the most common progressive leak (multi-node / TP=2)

Add these two environment variables to every vLLM launch (or docker-compose / start script):

UCX_MEM_MMAP_HOOK_MODE=none
UCX_RCACHE_MAX_UNRELEASED=1024

alexellis

Rebuilt llama.cpp:

llama.cpp/build/bin/llama-server \
  -hf unsloth/GLM-4.7-Flash-GGUF:UD-Q4_K_XL \
  --jinja --threads -1 --ctx-size 96768 \
  --temp 1.0 --top-p 0.95 --min-p 0.01 --fit on \
  --flash-attn off

alexellis

This final prompt ended up having to be so specific, even giving the implementation detail as the model was so dumb.

Prompt:

We're buying licenses from slicervm.com for their microVM product.

Find out the pricing for 2x home edition licenses and 5x commercial ones the page
is https://slicervm.com/pricing

alexellis

config:
  host_groups:
    - name: coral
      userdata_file: ./userdata.sh
      storage: image
      storage_size: 25G
      count: 1
      vcpu: 2
      ram_gb: 4
      network:

alexellis

# Example from https://docs.slicervm.com/getting-started/walkthrough/

# Then under the hostgroup instead of userdata:
# userdata_file: ./userdata.sh


# Then add a DNAT rule from the host to the microVM via iptables (ask ChatGPT for this!)

# Or simply add a static route on your router via your Slicer host to the microVM - easy to do 
# in most router admin interfaces, then also set the default DNS to the VM IP there too.

alexellis

package main

import (
	"context"
	"flag"
	"fmt"
	"os"
	"reflect"
	"time"

alexellis

From my blog post: The Internet is my computer

Equinix Metal is my computer

Run hosted VSCode on Equinix Metal's huge: AMD Epyc instances with 64GB RAM and 24 Cores, coupled with a bonded 2 x 10 Gbps uplink to the Internet.

• Provision your Ubuntu 20.04 LTS server using the dashboard and add your SSH key
• Once you have the public IP, log in over SSH

alexellis

Installing OpenFaaS with libkrun-based isolation on Fedora 33

Setting up the system

Revert to cgroup v1

The version of containerd that OpenFaaS will download during the installation doesn't support cgroup v2, so we need to configure systemd to avoid using the Unified cgroup Hierarchy, and reboot the system:

alexellis

Resources

Side-project thread

If you were starting a side-project or two today...

GitLab 5-minute rails

5 Mins to free-tier deployments