XIO

xWire

Posts from every feed QUEST has found across the web — blogs, news, podcasts, channels and more.

NVIDIA Developer Forums - Latest posts

Latest posts

Course Completed 100% but Grade Shows 0% – Certificate Not Issued

Hello NVIDIA Support Team, I completed the “Securing Agents with NemoClaw and OpenShell” course, and all the exercises are marked complete with 100% progress in the interactive course. However, my NVIDIA My Learning page still shows 0% completion and a 0% grade, so I am unable to receive my certificate. I have attached screenshots showing the completed exercises and the 0% progress displayed on my NVIDIA account. Could you please review my course completion status and help me obtain my…

DeepSeek-V4.1-Flash-1017 (EXL3 2.9bpw) on 2x DGX Spark with TensorFold: 57 tok/s text, 92 tok/s code, 2.1-2.3x the vLLM kit

Anyone else feel free to chime in if seeing better numbers or config - What I’m running right now: jayleaton’s deepseek-v41-tensorfold-spark with its own 2.8bpw pack. Main reasons: Structured output works (response_format with strict schemas, tool_choice required/named). The MiaAI-Lab Zig build returns 400 on response_format. Upstream ships real tests (structured-output suite, grammar goldens) and publishes SHA256SUMS for the pack. Long-context prefill is ~11% faster than MiaAI-Lab at a matched…

OptiX 9.1: SER + OptiX-IR/PTX crashes with HAGS enabled (CUDA 719);

@lspano I was doing some research and saw Blender had some issues with system lagging/input issues and one of the recommendations was to turn off HAGS, which of course is not a proper solution. Turning off HAGS also solved my crashing almost entirely, apart from a very, very long run but repeated runs and HAGS off + SER was generally solid. They closed that thread recently, saying drivers 620+ should fix it. I have now installed Studio Driver 621.24 and so far I have not had a crash on multiple…

[PyPI/nvidia-ml-py] Issue Reports for `nvidia-ml-py`

nvidia-ml-py 13.615.71, driver 615.71.09, GeForce RTX 5060 Ti, Ubuntu. These wrappers in pynvml.py drop the NVML return code and always return NVML_SUCCESS, so a failed call looks like success: nvmlDeviceGetClockOffsets nvmlDeviceSetClockOffsets nvmlDeviceSetPowerManagementLimit_v2 nvmlDeviceGetPowerMizerMode_v1 and nvmlDeviceSetPowerMizerMode_v1 return the raw code instead of raising NVMLError. On top of that, Get/SetClockOffsets and Get/SetPowerMizerMode_v1 pass the struct as is. Without…

Manual Account Verification Request - Pakistan (+92)

Hello NVIDIA Support Team, I’m trying to generate an API key on build.nvidia.com for the NIM APIs, but I’m unable to complete phone number verification because Pakistan is not listed in the country dropdown for phone verification at all. Account email: adeel.khan7400@gmail.com Country: Pakistan Phone country code: +92 Issue: Pakistan does not appear in the country selection dropdown, so I cannot even attempt to enter a phone number for verification. Could you please manually verify my account,…

Manual Account Verification Request - Pakistan (+92)

Hello NVIDIA Support Team, I’m trying to generate an API key on build.nvidia.com for the NIM APIs, but I’m unable to complete phone number verification because Pakistan is not listed in the country dropdown for phone verification at all. Account email: adeel.khan7400@gmail.com Country: Pakistan Phone country code: +92 Issue: Pakistan does not appear in the country selection dropdown, so I cannot even attempt to enter a phone number for verification. Could you please manually verify my account,…

620 release feedback & discussion

Note: this was diagnosed with the help of an LLM Regression in 620.70.06 (works in 615.78.08): Hardware cursor refresh rate becomes locked to game FPS in Borderless Fullscreen XWayland games when using an in-game FPS limiter Environment: RTX 2070 (nvidia-open 620.70.06), Linux 6.18, KDE Plasma 6.8.2 Wayland, 120Hz DisplayPort monitor, VRR/Adaptive Sync disabled. Issue: Following the 620.70.06 Wayland explicit-sync changes, running a DX12/Vulkan game under XWayland/Proton (e.g., World of…

Access to Public API Endpoints

Could you please enable Public API Endpoints access for my NVIDIA Developer / NGC account? I am trying to use hosted NIM models via the OpenAI-compatible API (://nvidia.com), and my requests are failing with the following 404 error: Function '23d4f03a-b8a6-4adb-a183-7daa083a09cc': Not found for account 'yx36G0iwCWtzdp2K9RaUmC_sTCvP7sWxi-WtmykFNZg' Account Details: Account Hash / ID: yx36G0iwCWtzdp2K9RaUmC_sTCvP7sWxi-WtmykFNZg Use Case: Personal development and testing with open-source LLM…

Reproducible Xid 109 (CTX SWITCH TIMEOUT) / VK_ERROR_DEVICE_LOST in Mount & Blade II: Bannerlord (DXVK) on RTX 5070 Ti — 595.84 and 620.70.06

Summary Mount & Blade II: Bannerlord (Steam, Windows build through Proton/DXVK) hangs the GPU within 1–3.5 minutes of starting a large naval battle on one particular scene. The kernel logs Xid 109 … CTX SWITCH TIMEOUT for the game process, DXVK reports Timeout occured, GPU likely hung followed by VK_ERROR_DEVICE_LOST, and the game shows “Direct3D 11 device has been removed or reset” (0x887A0007). It is reproducible on demand: 9 hangs in 9 battles at the game’s default large battle size, on both…

DGX Spark USB-C PD firmware inventory mismatch after Dashboard update: fwupd/ESRT versus SMBIOS

NVIDIA Customer Care recommended a separate DGX Spark / GB10 forum discussion, referencing case 261109-000302. We are seeking a supported verification procedure, not reporting a confirmed hardware defect or claiming that the active PD firmware bank has been established. Fresh November 15, 2026 inventory We collected a new read-only inventory across all six systems on November 15 at approximately 18:08 UTC. This included fwupdmgr get-devices --json, SMBIOS Type 45 (dmidecode -t 45), and ESRT…

404 Error / Missing API Endpoint Code for cosmos3-nano on build.nvidia.com

We’re trying to test Cosmos3 Nano for image-to-video generation and have encountered two issues. 1. Image upload in the browser playground On the Cosmos3 Nano Experience page, we selected Image to Video, but could not find a control to upload an image. We also tried another browser and pasting the image into the prompt area; neither worked. The Send button remained disabled. NVIDIA’s built-in image-to-video example did work: selecting it inserted the example image, enabled Send, and generated a…

Bug: Segfault when closing a multi-window Vulkan app under Wayland

Platform: Fedora 45, kernel 7.3.3-300.fc45.x86_64 with KDE 6.8.3 on Wayland with drivers 620.70.06 When closing a Vulkan app that uses multiple windows, with each window having its own swapchain, the final call to vkDestroySwapchainKHR causes a segfault somewhere in the Nvidia driver. The easiest way to reproduce this is with the SDL test suite, which can be built by building SDL with the -DSDL_TESTS=ON CMake parameter set. After building, run SDL_VIDEO_DRIVER=wayland SDL_RENDER_DRIVER=vulkan…

[Guide] Cosmos Reason2-2B for safety checks: four fixes that took off-list answers from 59 to 0 and false PPE alarms from 11 to 0

I moved the same safety checker to Cosmos 3 Edge on the Thor. With one server setting added, it found the person in 44 of 45 labeled pictures against 29 for Reason 2, with no false proximity alarms, and read the protective gear less well. The launch setting and the full comparison are here: [Guide] Cosmos 3 Edge on Jetson AGX Thor: the launch setting that made it usable, and what changed from Reason 2

620 release feedback & discussion

Using cachyOS with driver 620.70.06 i am unable to create 32-bit context. Using 4070Ti Super card, and I did not have these problems with the 615 driver branch. Small test program: #include <stdio.h> #include <dlfcn.h> typedef int CUresult; typedef int CUdevice; typedef void\* CUcontext; typedef CUresult (\*pfn_cuInit)(unsigned int flags); typedef CUresult (\*pfn_cuDeviceGetCount)(int \*count); typedef CUresult (\*pfn_cuDeviceGet)(CUdevice \*device, int ordinal); typedef CUresult…

[Guide] Cosmos 3 Edge on Jetson AGX Thor: the launch setting that made it usable, and what changed from Reason 2

I run a safety checker for a robot workcell on a Jetson AGX Thor. It looks at camera frames and reports each person and whether they wear a hard hat and a vest, and rules turn those answers into safety events. It ran on nvidia/Cosmos-Reason2-2B. When Cosmos 3 came out, I put nvidia/Cosmos3-Edge behind the same request to see what changed. The first run was not usable. One server setting fixed it, and after that Cosmos 3 Edge found people Reason 2 missed. Setup: Cosmos3-Edge (BF16, revision…

DeepSeek-V4.1-Flash-1017 (EXL3 2.9bpw) on 2x DGX Spark with TensorFold: 57 tok/s text, 92 tok/s code, 2.1-2.3x the vLLM kit

DeepSeek-V4.1 on TensorFold: MiaAI-Lab 2× (2.9bpw) vs jayleaton (2.8bpw) Follow-up . Both nodes on stock watermarks, as in the earlier post. Credit to jayleaton and ashhart for the engine work behind this. What changed. Structured output now works. Verified with one nested strict schema (required nested object, integer, enums at both levels, additionalProperties: false), which returned valid JSON with the correct value. Upstream’s structured suite also passes direct: 12 schema cases and 10 tool…

Quiet MikroTik CRS804-4DDQ: one Noctua NF-A12x25 instead of the 2 × 40 mm Delta fans

The CRS804-4DDQ (4 × 400G, popular for clustering NVIDIA DGX Spark / GB10 boxes) cools its switch chip with two 40 × 40 × 28 mm Delta FFB0412UHN fans (18,500 rpm, 58.5 dB(A) at full speed). They are loud, and RouterOS can only bend the fan curve a little (fan-full-speed-temp is capped at 65 °C). This mod replaces both fan modules with 3D-printed plugs that slide into the fan bays like the original hot-swap carriers, ducting both bays into one 120 mm Noctua NF-A12x25 PWM tilted 45° behind the…

Is it worth buying a second DGX Spark now?

Not obscure and not micro. Every inference stack I’ve looked at keeps a few CPU cores busy-polling the GPU millions of times a second while GPU generates. On the Spark the CPU and GPU share one power and thermal budget, so every spinning core is heat and watts taken straight out of the GPU. For us that adds up to 10W per box and a peak temperature 15C higher for zero extra speed. If your agent wakes up once an hour, does its thing and goes back to sleep, you’ll never notice. Run 3 agents in…

Request to enable "Public API Endpoints" permission for my personal organization

Hi NVIDIA team, I registered an NVIDIA account and completed email verification via build.nvidia.com, and a personal NGC organization was created automatically. However, when generating a Personal API Key, the “Services Included” list only shows “Secrets Manager” and “NGC Catalog” — the “Public API Endpoints” scope is missing, so every call to https://integrate.api.nvidia.com/v1/chat/completions returns 403 {“detail”:“Authorization failed”}. Also, the on-site key generation flow is blocked by…

Is it worth buying a second DGX Spark now?

I won’t argue and I’ll agree with you. A lot depends on the stack, and TensorFold is great on a number of models and on Spark too, delivering several times higher speeds (though not without problems, of course). Personally, the choice for me now comes down to the following: one Mac and in fact only one active model in the agentic loop four Sparks and the ability to run anywhere from four to one models independently on that stack given the figures published above, I personally lean toward the…

Is it worth buying a second DGX Spark now?

Are you talking about that random thread you made about some obscure, extremely low level performance issue that you probably used Claude to find? I don’t want to downplay performance penalties here, but I think you’re kind of proving my point for me. We’re at the stage of mostly optimizations. There are just tons of recipes out there now for the DGX Spark, many of them are just “download this repo, run this script”, and suddenly you just have a working and well optimized high end inference…

NVIDIA, THIS IS UNACCEPTABLE! RTX 4080 SUPER — BLACK SCREEN ON DRIVER 621.14, 5120×2160 @ 120 Hz BROKEN, WinDbg EVIDENCE INCLUDED

Update. It’s looking like my issue was caused by an overheating 12VHPWR cable. Three of the pin sheathes were showing signs of deformation, probably causing these pins to lose conductivity during periods of high use. Replacing the cable seems to have solved my problems.