Unable to Access DeepSeek V4.1 Flash and Kimi Models Through NVIDIA NIM API
Hello NVIDIA Support Team, I am requesting assistance with accessing NVIDIA NIM hosted APIs. Affected models: DeepSeek V4.1 Flash Kimi K2.6 Kimi K3 I use NVIDIA NIM for personal development, AI experimentation, and coding assistance. Could you please check whether my personal organization has the required “Public API Endpoints” permission? If my account requires additional verification or approval, please let me know how to complete the process. I would also appreciate confirmation that these…
英伟达团队您好,最近遇到iperf3打流问题,请教下能否提供一些建议
免责声明: 本内容经 AI 辅助编写。相关技术细节及操作步骤请在您的环境中测试,并以 NVIDIA 官方指南为准。 您好 @1516775545, 从您附带的截图中可以看到,第二次打流时 TCP 拥塞窗口(Cwnd)直接从正常的 ~3.47 MBytes 骤降到了 1.38 KBytes,导致大量的重传(Retr)与传输暂停(0 bits/sec)。 这在 Jetson Thor / Orin 的 10GbE (MGBE) 网卡测试中较为常见,主要与网卡电源管理 (EEE/LPI)、TCP 环形缓冲区未及时释放以及硬件 offload 状态有关。您可以按照以下步骤排查和优化: 1. 禁用节能模式 (EEE / LPI) MGBE 网卡在服务端停止打流后可能会迅速进入低功耗空闲状态(LPI),导致再次发起打流时丢包。请在客户端和服务端同时执行: # 禁用网卡 EEE 节能 sudo ethtool --set-eee mgbe0_0 eee off (请将 mgbe0_0 替换为您实际使用的 10G 网口名称,如 mgbe2_0) 2. 避免套接字 TIME_WAIT 影响…
Optimal Model Deployment, but priotize on max concurrency
Really? 10M KV pool? Is this with eugr’s vLLM receipt or some of the newer inferencing system like TensorFold?
NCCL initialization fails under nsys injection on GB300 aarch64/MNNVL: CUDA error 400 at cuMemImportFromShareableHandle
Thanks for the update and for completing the tests. Glad to hear that all three configurations now capture the workload successfully and you can continue profiling. If the original error returns or you uncover anything new, please update this thread and we’ll pick up the investigation.
DGX Spark vs RTX Spark
One think I didn’t consider is that in the long run RTX might become much more supported and stable. And for most users who are tied with apps on windows you’d be able to get your portable assistant running on the same machine of your other apps, with no api nor internet access needed.
Manual Account Verification Request - Russia (+7)
Hello, I am trying to access the NVIDIA NIM API on build.nvidia.com, but I cannot complete the phone verification. My country (Russia, +7) is not available in the country code list for SMS verification. Additionally, after a few attempts to solve this, my account was temporarily locked with a 12-hour ban due to verification failures. Could you please help me unlock my account and manually verify my NVIDIA Developer status? Account Email: wombat200612@gmail.com Country: Russia (+7) Purpose:…
FreeBSD 595.99.02: SIGSEGV (NULL deref) in libnvidia-glcore.so.1 during glUnmapBuffer path, Qt Quick render thread (plasmashell)
Hi @freebsd_avg Thanks for reporting the issue, I have filed a bug 7006908 for tracking purpose.
Flare-local-inference: GLM-5.3-Flash Guide for 2× DGX Spark — 3.4M-Token KV Cache
That’s right. Flare Local Inference builds the entire environment during the initial run and saves the resulting state. For the model featured in my guide, it uses approximately 173 GB per device.
Introducing Tool Eval Bench CLI
Sorry for not getting back to you earlier. I’m currently on vacation but will add both requests to the list!
Manual Account Verification Request – Region: Russia(+7)
Hello NVIDIA Support Team,I am trying to generate an API key on build.nvidia.com to access NIM microservices for educational purposes. However, I am unable to complete the phone verification step because my country code is missing from the automated SMS dropdown list. Could you please review my account details and grant manual API key access? Registered Email: vladkoc10@gmail.com Country/Region: Russia (+7)
EfficientDetD4 ONNX conversion to TRT engine fails on ORIN NX SUPER 8GB
Thank you for your help. Native conversion worked successfully. However, We are using a Jetson Orin Nano running JetPack 6.2 (L4T r36.4.3). Native model conversion is working correctly with TensorRT 10.3.0.30, which is currently the only version available from the NVIDIA Jetson repository. However, our deployment environment uses the Docker image: nvidia/l4t-tensorrt:r10.3.0-runtime which contains TensorRT 10.3.0.26-1+cuda12.5. Our workflow is: Convert ONNX models to TensorRT engines natively…
Flare-local-inference: GLM-5.3-Flash Guide for 2× DGX Spark — 3.4M-Token KV Cache
Hey, how much disk space is needed for initial launch? I have ~398Gb free after weights are downloaded from HF still keep running out of disk space :(
[Benchmark] Inferact/Muse-Glimmer-30B-NVFP4-W4A4
only difference that muse will send to somewhere in Ohio or Texas :) it is time to fine tune own safeguard.
AODT 1.5.1 Access Issue – NGC Product Activation Invitation Returns Not Found
Thank you for your comment. It looks like your computer/phone are still logging you into NGC using your previous credentials. Please try to logout completely, clear the cache and reattempt.
Kernel panic after static discharge
Hello. We investigated this error and found it’s reason. ESD causes unrecoverable link failure (SDES - Surprise Down Error Status). Device interrupt dereferences a pointer which is null after surprise down. As null pointer dereference occurs in device interrupt, it causes kernel panic. Functions of pci_error_handlers struct were not implemented in the driver of our device so the driver kept running after link failure. We’ve implemented pci_error_handlers.error_detected and added pointer’s…
New Deepseek 4.1 flash recipe
I hear that there are hints added to deepseek harness for v5…
New Deepseek 4.1 flash recipe
This would be really nice if something like this happens. ngram for sparks really does make a big difference. Lets hope we get more models like this. Qwen 4 should be something similar, but who knows the parameter sizes yet.
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
not exactly, I run kv-cache storage from GitHub - kvcache-ai/Mooncake: Mooncake is the serving platform for Kimi, a leading LLM service provided by Moonshot AI. · GitHub , just wired it to tensorfold with help of opus. I needed earlier as I have 16gb gpu with lots of repetitive prompts thru it, which was constantly evicting when traffic was from different agent. on 10gbps network pulling 800mb cache e2e ±1.3s vs re-processing for tens of seconds in best case. you can use same to have…
CachyOS frequent freezes and crashes when watching video and gaming
The freezes are caused by the OS and drivers losing communication with the GPU so it is hardware related. If connections are good and it isn’t because it is overheating, it might be something else like power delivery, like a bad PSU or cable, or could be the GPU is going bad.
Mobile-Gaming+Samsung-DEX+GeForce~NOW+Cloud-Gaming
Cloud gaming performance depends mostly on your internet stability and latency, not just phone settings. Try a 5 GHz Wi-Fi connection, keep your phone cool, and close background apps. Also, avoid changing too many developer or browser flags, as they can sometimes cause issues. For the Arc Raiders ban, only the game’s support team can review and resolve the account restriction.
New Deepseek 4.1 flash recipe
Update, We were able to run 763B-parameter models using TP4. Thanks, ngram—we could potentially run 1526B-parameter models across 8 nodes. Just in case a new version comes out?

New Deepseek 4.1 flash recipe
Ok, here you go - courtesy of my speaking to my agent over the apple car play while driving - will never get over how cool that is. @rob01 - tagging you in so you can have a look too if you’ve still got the knapcio around (FWIW I’m running a 100gb MikroTik CRS504-4XQ switch with 100gb cables in TP4 - 200gb really doesn’t seem necessary for inference) https://github.com/stujmiller/runtime-abliteration
New Deepseek 4.1 flash recipe
Stu, I think your auto correct got its wires crossed recently in your posts, it is SGLang, not SDLang.
CUPTI 13.0.2 late attach: WHILE-node body kernels get one record per graph launch, not per iteration
Hi, @colinpeppler1 Sorry for the late response. Regarding you two questions, we have reproduced both. Thanks for the detailed report. We have reported 2 internal bugs for tracking. We’ll update this thread when a fix is available.
Request to enable “Public API Endpoints” for personal organization
Hello NVIDIA Support Team, I am an individual developer using a personal account organization for local application prototyping and model integration. My API key successfully authorizes GET /v1/models requests. However, whenever I submit a query to the POST /v1/chat/completions endpoint, the upstream server returns a 410 error code across all model entries. Based on community forum resolutions, this issue stems from my personal developer profile lacking the required “Public API Endpoints” role…
Title: NVIDIA Build API returns 403 Forbidden for all models
Hello NVIDIA Support, I am building a personal AI English Learning Simulator using FastAPI. My API key is generated through NVIDIA Build, but every request to the chat completions endpoint returns HTTP 403 Forbidden. I have tried generating a new API key and changing the model, but the issue continues. Could you please check whether my account has the required Public API Endpoints permission for hosted inference? I am using the API for learning and personal development. Thank you.
英伟达团队您好,最近遇到iperf3打流问题,请教下能否提供一些建议
系统版本R38.4.0 操作步骤:一端运行"iperf3 -s"; 客户端运行"iperf3 -c 10.42.1.102 -t 60"; 正常情况下能正常跑完: 如果iperf3 -c 10.42.1.102 -t 60跑完后,此时服务端停掉iperf3 -s然后再运行iperf3 -s;客户端再运行“iperf3 -c 10.42.1.102 -t 60”;就会出问题有0bits/sec的情况,如下图:

RTX PRO 6000 Blackwell + RTX 5090: Xid 79 fallen off the bus, 7 events, no precursor
I can confirm the same result from a different machine. I have a GIGABYTE AORUS GeForce RTX 5090 STEALTH ICE, and it showed me the exact same sequence on two different AMD boards — first an ASUS A620M-A-CSM, then an ASRock X670E Steel Legend (BIOS 4.43), both with driver 615.78.08 (open kernel module), the same driver you tested with. Every time: Xid 79 followed by Xid 154, card gone, black screen, and only a full reboot bringing it back. Since the problem followed the card across two different…