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

Manual Phone Verification Request - Russia (+7)

Hello! I am trying to access the NVIDIA NIM API on build.nvidia.com, but I cannot verify my phone number because Russia (+7) is not available in the country list for SMS verification. Could you please manually verify my NVIDIA developer account? Purpose: Testing NVIDIA NIM API for personal and educational development. Thank you in advance!

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

Interesting. Will check it out next week to see if it worth upgrading to. Have you guys heard all these heavy rumors: DeepSeek v5 (max, open,next versions) - surfaced in DS Harness code GLM 5.5 GPT 7 10T paramers And its confirmed - Qwen 4 releasing next week, maybe even Monday, tomorrow. Confirmed at the conf in Amsterdam last week

Native DirectX 12 on Linux

Thanks for the tracer, it caught the issue on the second run. I added it to my repo under MIT as you agreed, unchanged from the zip, plus a separate commit that silences the bpftrace 0.26 warnings (kptr() retagging of the syscall args, two-argument delete()); the recorded values are the same. Your tracer: tools: add Bober's RM object and GPU VA lifecycle tracer · kakra/ogkm-experimental@45bee32 · GitHub bpftrace 0.26 fix: tools: silence the bpftrace 0.26 warnings of nv-va-trace.bt ·…

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

40 is slow, i think mlx-serve does faster. If you only have a notebook - fine, but it makes not much sense if you have much stronger AND FASTER model on cluster. For me main reason to use MLX model is superior speed for documentation, exploration, summaries, reviews, minority reports etc ╭───────────────────────────────────────────────────────────────────────────────────────────────────────── 🏆 Benchmark Complete ──────────────────────────────────────────…

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

there is iq3xxs - bpw3 and iq2 quants - that certainly will BTW found optimum for medium - 0.6 temp + medium = 91/100. And the e2e crazy fast - 350s all 104 tests in 4 parallels. dragonscale in record 8 minutes (ds41 - 60, glm - 30). And game/autopilot/git/docs - all fine. 3.5 bpw. For low/medium complex tasks - ideal model to pair alongside ds41 for hard ones

Cannot use ChatNVIDIA()

Hi Support, I created my account today and I am receiving an HTTP 404 error using the raw OpenAI python SDK client targeting the ‘/v1’ path. Account Hash: ‘PUQRAoqNDpG7y031yR3r2DQZQYdQmYLS0htjsLMXCQY’ Please sync my account to the public endpoints group. Thanks!

GB10: cudaMalloc memory is not charged to the process or its cgroup, so the OOM killer and container limits can't stop a unified-memory freeze (580.17

Same here on a GB10 (ASUS Ascent GX10, driver 580.178.04). Our 27B lane runs SGLang in a container started with --memory 100g. During a 5,623-token generation at a 0.76 static fraction, the container’s cgroup held 7.15 GiB of its 100 GiB the whole time, while the engine’s pool sat in GPU memory the cgroup never saw. So the cap only bounds host RSS, as your table shows. What keeps the box alive for us is the engine’s own budget: SGLang’s --mem-fraction-static (0.76 on the 27B, 0.80 crashed under…

veloGB10 v0.7.0 RELEASE - Qwen3.8-Flash-Next TP=2 @ 186 tok/sec , SINGLE at 137 tok/sec on code

Hey Christopher! Actually no need to send anything. The issue has been fixed in v0.7.7: File sf-stav/veloGB10 CHANGELOG.md main ## v0.7.7 — fix: TP=2 MTP lockstep crash on first generation (the `--mtp` blocker) Fixes the v0.7.5/v0.7.6 crash reported on the issue tracker: with MTP enabled at TP=2, the first generation dies with `TP pre-verify FAILED` / `TP agree() FAILED at mtp-round` and exit 70. `--mtp off` worked around it; v0.7.4 was unaffected. **Root cause:** the fractional-K kernel work…

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

I re-ran the DeepSeek V4.1 1017 results from this thread on my own 2-node setup. I also ran GLM 5.3 Flash with the identical TEB build and settings, at two reasoning efforts (high and max), so the two can be compared directly. Setup 2× GX10 (GB10), tensor parallel over RoCE DeepSeek: V4.1-Flash-1017 EXL3 2.8bpw (jayleaton) on 0rand’s Zig stack main-0rand-055c4ae, 8 slots GLM: GLM-5.3-Flash EXL3 4bpw (Mia-AiLab) on TensorFold v0.6.0 TEB: tool-eval-bench 3.0.1.dev6+g7b022ddcb…

New Deepseek 4.1 flash recipe

You are right, default is off on the SGLang ones … I also set thinking off in my c1-c6 tests so it’s apples to apples: Here is what my model said: vLLM (10-25, 10-28): flag sent explicitly, and the preflight proves it worked — 2 tokens with the flag vs 12 without. Thinking was live by default; the flag turned it off. Clean, demonstrable control. SGLang (10-27, 10-28): flag sent explicitly, but the preflight returns 2 tokens either way. So the engine was already not thinking. You measured…

A silent bug in cuSolverDx heev when exactly one eigenvalue is exactly zero

Hi, Both heev and htev are designed to handle very large and very small inputs. For heev this isn’t an explicit scaling at the beginning and end, and instead it’s subroutines individually support large ranges. Do you have matrices where it generates an incorrect result? And thanks for sharing your use case, that’s a helpful datapoint.

Subject: Manual Account Verification Request

Hello NVIDIA Support Team, I am receiving the error message “Please contact support to verify your account” when trying to access build.nvidia.com / generate an API key. I have already sent multiple emails to help@build.nvidia.com but have not received any response yet. Therefore, I am posting here to request manual verification. **Account Details:** - Registered Email: [qqq851619132@gmail.com] - Country: China - Platform: build.nvidia.com **Intended Use:** Personal development / learning /…

New Deepseek 4.1 flash recipe

So down the rabbit hole we go. Using the Mikrotik 804 with 200Gps connections I tried all the top DS4.1F 1017 repos that have been mentioned in these forums: Knapcio, Mia, @stephen.douglas.hawkins DGPP , and @christopher_owen’s. They all have pros/cons, along with things that everyone can learn from each other. Knapcio and Mia are solid, fast on tool eval bench, but slower in my workload c1-c6 tests compared to Christopher’s. DGPP has a nice clean design but it was the slowest of them all.…

AODT 1.5.1 Access Issue – NGC Product Activation Invitation Returns Not Found

Hi eobiodu, Thank you for your suggestion. I have now checked my NGC account on a completely different computer that has never been used to log in to NVIDIA before. My account still shows only one NGC organization, “DGXSparkAODT”, with organization type “INDIVIDUAL”. I also checked the Subscriptions page. It lists three active “NVIDIA Developer Program” subscriptions, but there is no NVIDIA AI Aerial subscription. My 6G Developer Program application was approved on October 23, 2026, and I…

Manual Account Verification Request

Hello NVIDIA Team, I can’t generate an API key on build.nvidia.com. The API Keys page shows: “API Access Unavailable – Please contact support to verify your account access.” This persists even after email verification. I registered from Spain using the same email address associated with this forum account. I plan to use the API keys for personal, non-commercial development, specifically for testing NVIDIA NIM APIs for AI-assisted code review and integration tooling. Could you please review and…

DGX Spark fan-control kernel module — running on dual Sparks since mid-August

Update and a correction: EC 3.5.8 works with the October SoC Hey Christopher, an update, and a correction to what I posted earlier. I wrote that the October SoC (2.158.15) locks owners out of EC 3.5.8. That was wrong. On Oct 25 we only tried the EC mailbox: our own module using nvfancontrol’s protocol (OEM1 command 17 and the shared page, through the SoC firmware’s FF-A eSPI service). Its very first read-only capability query timed out with eSPI status 5, and we jumped from “this path doesn’t…

Nvfortran fails to accept `target` argument to `pointer` dummy argument when in generic interface

module mymod implicit none interface in_generic module procedure with_ptr end interface contains subroutine with_ptr(a) real, pointer, intent(in) :: a(:) end subroutine end module mymod program main use mymod, only : in_generic, with_ptr implicit none real, target :: a(20) integer :: i a = [( real(i) , i = 1, size(a))] ! Does not compile call in_generic(a) ! Compiles with no issue call with_ptr(a) end program main Also here Compiler Explorer In the code snippet above, which I believe is a valid…

New DGX Spark: fans barely speed up at high temps and can't be adjusted. Critical. Please open fan control

Update: fan control works on the October firmware, no kernel module needed Hey everyone, we have an update: the fans work properly now. And it turned out much simpler than anything we tried before. No kernel module, no signing, nothing to compile. Just an older EC firmware from NVIDIA’s own update server (LVFS) and a sysfs file Linux already has. Downgrade only the EC to 3.5.8 (0x03000508), leave the SoC (2.158.15), USB-C PD and BIOS as they are, and the fans listen again: cur_state = 100: both…