Building GPT-QModel, ModelCloudAI. Contributor to stuff you are probably using.
Joined February 2020
- Tweets 6,514
- Following 4,257
- Followers 1,461
- Likes 3,256
700 Photos and videos
🚀 GPT-QModel v7.0.0 is here!
🔥 Huawei Ascend NPU support
⚡ JIT compilation for CUDA/ROCm kernels on first use
✅ Easier Pip/UV installs with no --no-build-isolation needed
🧠 Models Added: ⚡ GLM 5 / 5.1 👁️ GLM OCR 🎙️ GLM ASR 💎 Gemma 3n 🦅 Falcon Mamba 🖼️ InternVL Chat
6
2
21
1,906
Fix it. mpt3sas and mpi3mr driver have race conditions in the scsi handshake/discovery protocols when there are multiple sas expanders, lots of slots, and drives such as the segate which is exposes 2 drives (luns) per physical devices so a perfect storm of packet ping/pong where some packets gets lost due to the concurrency cause random drives to timeout during discovery and slot disabled in error.
1. Reset the card on linux kernel boot (module load). Why? because it has unstoppable uefi/bios firmware which does pre-boot work. Have no idea what it does (likey scan for bootable flags) but I don't want any part of the state leaking to boot since the bugs may where be broken state within the firmware during pre-kernel boot. The firmware/bios is not open-src so I want no part of it leaking bad card states into driver boot.
2. Make everything serial, as much as possible during scsi handshake and increase the protocol timeouts everywhere especially for expanders.
There is a serious bug in the Linux mpi3mr driver for the following combo: @ASUS Zen 4 host with SAS expander, @Broadcom 9600-24i, and @Seagate Exos 2X 14T drives. I don't have the time, resources, or patience to debug further. The culprit is 99.99% @Broadcom drivers and/or firmware, where it can't even correctly handshake the drives. It will randomly fail SAS drive handshakes on boot/insertion. Key word is random.
Why do I think it's the mpi3mr driver? Because 9500-8i with mpt3sas works perfectly. I also tried like 6 different 9600-24i firmware versions to see where they botched things, and none worked.
Could also be a compatibility issue between the SAS expander firmware and the 9600 series. Again, 9500 with mpt3sas drives has zero issues.
Totally wasted my time. I had to eliminate the drives, power delivery, and firmware versions before finally giving up on the 9600 and replacing it with the old but trusty 9500.
1
118
There is a serious bug in the Linux mpi3mr driver for the following combo: @ASUS Zen 4 host with SAS expander, @Broadcom 9600-24i, and @Seagate Exos 2X 14T drives. I don't have the time, resources, or patience to debug further. The culprit is 99.99% @Broadcom drivers and/or firmware, where it can't even correctly handshake the drives. It will randomly fail SAS drive handshakes on boot/insertion. Key word is random.
Why do I think it's the mpi3mr driver? Because 9500-8i with mpt3sas works perfectly. I also tried like 6 different 9600-24i firmware versions to see where they botched things, and none worked.
Could also be a compatibility issue between the SAS expander firmware and the 9600 series. Again, 9500 with mpt3sas drives has zero issues.
Totally wasted my time. I had to eliminate the drives, power delivery, and firmware versions before finally giving up on the 9600 and replacing it with the old but trusty 9500.
1
1
232
total direct to cpu, no plx, 176 pcie channels if you count the mcio x8 plus x4 connectors.
May 26
NVIDIA "Vera" CPU Benchmarked: Beating Intel Xeon and AMD EPYC in Select Workloads tpu.me/rqs2
2
6
853
Us tax payers not only get shafted once, but twice. First in my ins premiums, and even worse, another is waiting at the ER. I am leaving California.
May 24
California provided full scope Medi-Cal health coverage to 2 million undocumented immigrants between 2020-2024, free. How much do you pay for your health insurance?
1
182
🪣 Bucket list check: In the rain, in the dark, I helped two AC guys pull a 200 lb, 4-ton / 48k BTU/h condenser 10 feet up via a carabiner pulley in a disgusting back-kitchen alley.
Not my plan for today.
But lab servers must be cooled.
My motto: by any means necessary.
1
161
If I am CIA and I have to either control the narrative of AI or cripple the competing nations ability to react, she would be on my top list of recruits.
May 20
On the one-year anniversary of EMPIRE OF AI, I am so, so excited to announce The AI Resist List, a new project that documents examples of resistance to the AI empires around the world.
airesistlist.org
ALT The AI Resist List homepage
1
188
CTranslate is a hidden gem btw.
May 19
I’ve just released CTranslate2 4.7.2 (github.com/OpenNMT/CTranslat…). This new version includes support for Gemma 4 31b dense model and includes several bug fixes and improvements.
3
304
They did nuke H20's FP64 performance to 1 TFLOPS from.......34 TFLOPS. Pretty much banned. 🫶
May 18
Calling GPUs “nuclear bombs” is unserious policy theater. At this rate, they should advocate banning FP64 math too.
1
260
I absolutely agree the 1-8 node non-training compatible H100 instance pricing spike is likely manipulated by someone or entity and not actually used for ai.
May 17
Crazy price action in H200 cloud pricing – up 56% in 3 days.
What is unusual is that the H200 is suddenly trading higher than the B200, a superior GPU.
It’s not crazy to think that a fund could bid up supply in an illiquid tight market at a cost of $50K a day to engineer a short-term move in much more liquid stocks.
1
3
812
Would you believe if I tell you a great percentage of those domestic chips are in machines that are either not online or just turned off?
China is reducing its reliance on foreign AI chips:
China's AI chip self-sufficiency ratio is up to a record 41%.
This measures the proportion of domestic AI chip demand met by locally produced chips, rather than imported ones.
This ratio has QUADRUPLED over the last 5 years.
The AI chip self-sufficiency ratio is now projected to more than DOUBLE to ~85% by 2030, according to Morgan Stanley.
In other words, China could meet nearly all of its own AI chip demand domestically within 5 years.
China's AI chip independence is accelerating.
1
4
676
If you ever used Dell ent, their cabling is masterful. There is perfect length for all connections. Versus vanilla oems which has loose cables and internal parts everywhere. Dells are made like a swiss watch, if you can afford them, get them.
2
2
24
2,869
I had a BeOS as play, school (cs c compiler) machine for like 6 months and it was awesome. Even ran a file sharing service on it which almost led to some disciplinary blowback. My dorm mate unplugged the machine when the school IT came to "check" who was using all their bandwidth.
Apple almost replaced Mac OS with another operating system.
One ex-Apple executive built it from scratch. 🤯
Meet Jean-Louis Gassée 🇫🇷
> Former Apple executive.
> Left Apple after internal power struggles.
> Started Be Inc. in 1990.
> Wanted to build the future of computing.
> Created BeOS from scratch.
> Built for speed and multimedia.
> Real-time audio. Smooth video. Proper multitasking.
> Windows 95 looked ancient beside it.
> Booted insanely fast.
> Lightweight. Stable. Years ahead.
> Mid-90s Apple needed a modern OS desperately.
> Final decision came down to two options:
> BeOS or Steve Jobs’ NeXTSTEP.
> Apple chose Steve Jobs.
> That decision changed tech history forever.
> Jobs returned to Apple.
> NeXTSTEP became the foundation of macOS.
> BeOS slowly started dying.
> Microsoft pressured PC makers away from BeOS machines.
> Palm acquired Be Inc. in 2001.
> Shut the entire OS down.
> One day later, developer Michael Phipps restarted it as open source.
> Rebuilt everything from scratch.
> Called it Haiku. 🌸
> Still developed by volunteers today.
> Still ridiculously fast in 2026.
> Still runs beautifully on old hardware.
Apple chose the future.
But BeOS became one of tech’s greatest “what ifs.” 🔥
2
1
27
3,410
The problem is not the connector or the wire gauge but that almost all 16pin ends (esp the 12pin parts)are semi-loose and can recess, causing uneven connection. Must push both the connector and the cable itself. Hard to describe it but when you see it, you will understand. This an usability design issue.
rtx pro 6000
psu unknown
16pin melted
bilibili 靓女维修佬
bilibili.com/video/BV1h85e6Q…
2
5
1,068