Skip to content

Instantly share code, notes, and snippets.

@solatticus
solatticus / vulkanfix.md
Created January 3, 2026 20:15
FIXED: Vulkan 'vkCreateDevice' failures with the GB10 (Blackwell) GPU

Fixing Vulkan on NVIDIA DGX Spark (GB10/Blackwell)

A guide for DGX Spark developers experiencing Vulkan vkCreateDevice failures with the GB10 (Blackwell) GPU.

The Problem

You have a new spark, but Vulkan applications fail with errors like:

vkCreateDevice failed with VK_ERROR_INITIALIZATION_FAILED
@solatticus
solatticus / vulkanfix_02.md
Created January 4, 2026 13:50
Update: Fixing Vulkan on NVIDIA DGX Spark

Fixing Vulkan on NVIDIA DGX Spark (GB10/Blackwell)

A guide for DGX Spark developers experiencing Vulkan vkCreateDevice failures with the GB10 (Blackwell) GPU.

The Problem

You have a new spark, but Vulkan applications fail with errors like:

vkCreateDevice failed with VK_ERROR_INITIALIZATION_FAILED
@solatticus
solatticus / nvidia-false-advertising.md
Last active August 13, 2026 17:33
FlashAttention-4 Cannot Run on RTX 5090 (SM120) — A Deep Investigation

FlashAttention-4 Cannot Run on RTX 5090 (SM120) — A Deep Investigation

TL;DR

FlashAttention-4 will not run on the NVIDIA RTX 5090 (SM120, "desktop Blackwell") and no amount of software patching can fix it. Despite sharing the "Blackwell" brand with data center GPUs like the B200 (SM100), the RTX 5090 uses a fundamentally different tensor core architecture. SM100 has a dedicated tensor memory (TMEM) subsystem with its own instruction family (UTCHMMA, UTMALDG, etc.) that FA4's warp-specialized kernel design requires. SM120 uses the older HMMA instruction family (the same register-to-register MMA approach used since Volta/Ampere) and the TMEM hardware is physically absent from the GB202 die. This is not a software lock, not a fuse bit, and not a toolchain oversight — it is a silicon-level architectural difference. FA2 via Triton remains the best available attention kernel for the RTX 5090.


Goal

@solatticus
solatticus / poolside_118B_5090RTX.md
Created July 22, 2026 23:05
# Run Poolside Laguna S 2.1 (118B MoE) on a single RTX 5090

What was done

~19 tok/s decode · ~60 tok/s prefill · ~30 GB VRAM · 256k context

Poolside’s Laguna S 2.1 is a 118B total / ~8B active MoE coding model (256 experts, top-10 + shared, hybrid SWA). Official GGUFs: poolside/Laguna-S-2.1-GGUF.

People have been showing hybrid llama.cpp runs on 24 GB cards. That works — but the default “all experts on CPU” placement leaves a lot on the table. On a 5090 (32 GB) the win is packing full layers (experts included) into VRAM with auto-fit, not pure --cpu-moe.

This is a measured recipe from one box, not marketing.