N
Most of sparks in this community run headless in multiuser mode, no gui, x11. Most keep it lean af and only run inference on it full stop and just one large model no hot swapping, at least multi spark users.
N
I noticed that a contributor in the NVIDIA 620 release feedback thread has published a patch addressing a potential deadlock in the NVIDIA DRM event-handling path: Commit gulafaran/open-gpu-kernel-modules nvidia-drm: handle display events off the NVKMS event thread committed 04:28AM - 19 Nov 26 UTC VeilSilence +463 -69 The NVKMS event thread also delivers the flip completions that a blocking atomic… commit waits for, while holding connection_mutex. The display change, content protection and…
N
I think for a single user without huge autonomous agents pool 2 sparks is a golden bibi now. With latest speed upgrades in dsv41 and glm53f on tensorfold I do have more compute and speed than I really need. It get things done and then I am the bottle neck to field test, add my bits, etc. So compute is idling. Ram Altman is right to a degree lol
N
Nah, maybe in us of a, you get epic clown show going on. World is bigger. There will be places. You cans stop open source. Learned from pgp story. The question though about data, training. Arch advances are super solid.
N
Max sounds not promising, it’s 2.8 T model that doesn’t look better than ds v41 flash really much. hy4 flash is out but qwhen waiting made people ignore it. Metal community is delivering quant and engine mods already similar ngram arch like qwen4 and dsv41
N
Adding measurements from two Founders Edition units on the same versions as reported above, in case it helps NVIDIA reproduce. Setup (both units identical) DGX Spark FE, board P4242, BIOS 5.36_0ACUM028 EC 0x03000d07 (3.13.7), SoC 0x02009e0f (2.158.15), DGX OS 7.6.0, kernel 7.0.0-1024-nvidia Headless, Secure Boot on, GPU clock cap nvidia-smi -lgc 300,2200 A third-party external fan base was attached and running during all measurements below. It starts at 75 °C and reaches full speed at 85 °C.…
N
Use 0.6 temp. Top_p 0.95 or 1 makes no diff. And you can send sampling in DSH, just requires bit customization like everything in DSH. I have same issue in hermes, it forces temp 1.0 for all custom providers so all my hermes instances sit behind MediaLLMProxy, that overrides sampling, adds antiloop and other stuff
N
A fix is add 2 lines in /usr/src/jetson_sipl_api/sipl/CMakeLists.txt, sudo rm -rf /hololink, then rebuild. It then installs *.hpp to: /usr/local/include/hololink/core 17 LANGUAGES CXX 18 ) 19 20 +include(GNUInstallDirs) 21 + 22 set(PDK_HSL_DIR "${CMAKE_CURRENT_SOURCE_DIR}/hsl") 23 set(PDK_UDDF_DIR "${CMAKE_CURRENT_SOURCE_DIR}/uddf") 24
N
Has anyone else with an RTX 4080 Super experienced similar symptoms and checked their 12VHPWR cable for signs of overheating?
N
I’m really looking forward to your release))
N
hi: 我们有一款相机分辨率是1920x1536,在dts里面配置了embedded_metadata_height = “1”; active_w = "1920"; active_h = "1537"; embedded_metadata_height = "1"; 这样配置之后,在thor侧如何获取这一行的嵌入式数据呢?
N
Assuming you plan to use JP 7.2.1, here’s how USB boot applies to your setup. UEFI boot. Yes, provided the reader presents the card as a USB mass-storage device using Bulk-Only Transport and UEFI recognizes the prepared card on your custom carrier. Insert the card and connect the reader to a USB host port before powering on. USB boot requirements Root filesystem. Yes, under the same conditions. The boot firmware can reside in QSPI-NOR and the Linux root filesystem can reside on the card through…
N
Hi NVIDIA Team, I am a developer based in Pakistan and am trying to generate an API key on build.nvidia.com to test NVIDIA NIM models. During phone number verification, Pakistan (+92) isn’t listed in the country code dropdown, and I don’t have an international phone number. Could a moderator please manually verify/whitelist my developer account so I can access the API catalog? Registered Email: zain03020300@gmail.com Country: Pakistan (+92) Phone Number (if needed for record): +92 310 7532674…
N
Is it intentional for Jetson_SIPL_API_R39.2.2_aarch64.tbz2 build to install *.hpp to root /hololink? scott@chithor:/usr/src/jetson_sipl_api/sipl/build$ grep hololink $PWD/install_manifest.txt /usr/local/lib/libhololink_core.a /hololink/core/arp_wrapper.hpp /hololink/core/csi_controller.hpp /hololink/core/named_lock.hpp /hololink/core/csi_formats.hpp /hololink/core/data_channel.hpp /hololink/core/deserializer.hpp /hololink/core/enumerator.hpp /hololink/core/hololink.hpp /hololink/core/jesd.hpp…
N
I am receiving a “410 status code (no body)” error when making requests to the /v1/chat/completions endpoint using my personal developer account, although /v1/models works fine. It seems my organization lacks the “Public API Endpoints” entitlement. Could an administrator please enable this permission for my account? Account Email: [obsidianberserker1@gmail.com] API Key Prefix: nvapi-uV8tKLP_qn I have been at this for months now emails posting here and nothing is done
N
Requesting you to try with latest released driver 620.73.04 and share the feedback.
N
Requesting you to try with latest released driver 620.73.04 and share the feedback.
N
Hi All, Requesting you to try with latest released driver 620.73.04 and share the feedback.
N
Same here not gonna “stack-hop” until Qwen 4 releases next week.
N
Correct. That’s why i’m using the Nvidia fiber splitter cable precisely to achieve the 200Gb speed by the optic (2 fiber on split ends creates 200G and reduces power)
N
@james.meadlock great setup! What models are you running on the GB300 and how fast is it compared to the Sparks in terms of tok/s. Wondering specifically if you are running DeepSeek.
N
Hello NVIDIA team, I generated a valid nvapi- key at build.nvidia.com (valid for 1 year) but cannot use any NIM model (e.g. z-ai/glm-5.3-flash or others). Account is Personal Organization. Please enable “Public API Endpoints” permission for my account: Email: amcod951@gmail.com API Key prefix: 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 Sample model I want to use: z-ai/glm-5.3-flash (or any NIM model) Thank you! amirabbas amini 11/16/2026
N
Прошу вручную подтвердить аккаунт. Из России это сделать нет возможности
N
You probably already know this but the internal wiring of the ConnectX-7 is maxing out at ~200 Gbps. ┌──────────── GB10 package (TSMC 3 nm, ~140 W) ─────────────┐ │ MediaTek CPU die (20× Arm) ◄═ NVLink-C2C ═► Blackwell GPU │ │ memory controllers + all PCIe ~2 400 Gbps/direction* │ │ │ (~4 800 aggregate) 24MB L2 │ └────────┼────────────────────────────────────────────────────┘ │ LPDDR5x 256-bit @8533: 2 184 Gbps ◄── LLM-decode limit │ (hot chips spec @9400: ~2 400 Gbps) PCIe root ports on CPU…
N
Yeah, same I am not commenting mostly on just 100% everything working right now. Eargerly awaiting some model releases, but all ok right now.
N
TP2 will be possible in the next release
N
QFN and GLM5.3 but not because of model - it just corrupt the formats and incomaptible profiles after updates.
N
Ah well. Out of curiosity, what model did you run with it?