George Hotz | Programming | Mistral mixtral on a tinybox | AMD P2P multi-GPU mixtral-8x7b-32kseqlen
YouTube Viewers YouTube Viewers
194K subscribers
47,507 views
0

 Published On Dec 10, 2023

Date of the stream 9 Dec 2023.
from $1250 buy https://comma.ai/shop/comma-3x & best ADAS system in the world https://openpilot.comma.ai
Live-stream chat added as Subtitles/CC - English (Twitch Chat) - at the bottom - Show Transcript

Sources:
- https://github.com/tinygrad/tinygrad
- https://arxiv.org/pdf/2101.03961.pdf
- https://news.ycombinator.com/item?id=...
- https://github.com/tinygrad/gpuctypes
Follow for notifications:
-   / georgehotz  
Support George:
-   / georgehotz  
Pre-order tinybox:
- https://buy.stripe.com/5kAaGL6lk9uX9n... (https://tinygrad.org/)

Chapters:
00:00:00 muted intro
00:01:12 un-muted intro
00:01:25 tiny keyboards, feud with oneplus
00:02:38 bought oneplus open via oneplus store
00:03:16 bricked phone, fastboot set active slot, qdl mode, qdl data not available
00:05:14 no images to restore, not honoring return policy
00:06:50 oneplus destroyed their brand
00:07:47 chargeback, buy oneplus buy from amazon
00:08:20 comma return policy complete and true
00:09:10 chargeback, apple late delivery
00:11:40 processing emotions, do not buy it from oneplus store
00:13:00 comma does not want you to have a bad time
00:13:27 sending it back
00:13:55 edl, unofficial qualcomm firehose
00:16:50 phone market, why no images
00:17:48 companies don't want you to own anything
00:18:15 Apple consumer focused
00:18:35 jailbreaking is not a solution
00:19:05 forks of the software and hardware
00:20:35 possible to make phones
00:21:28 don't expect anything from oneplus
00:21:38 oppositional defiance, great regrets in life
00:23:05 oneplus release qdl files
00:23:40 suing someone different
00:24:55 2% of unreasonable people
00:25:18 reading all the emails
00:25:33 reading youtube comments
00:25:50 oppressive powers, free market, apple web browser lock in
00:27:48 anarcho capitalism
00:28:27 interns pay $1400-$2000 week
00:28:40 trans people, J. K. Rowling, transhumanism
00:30:10 ssh to tinybox 8
00:30:20 mixtral.py
00:32:08 python mmap
00:33:35 only 6 GPUs
00:34:15 bigger font, glasses, astigmatism
00:35:10 torch_load dynamically
00:37:25 Scaling to Trillion Parameter Models
00:37:41 HN Mixtral
00:41:30 how GPT4 works
00:42:00 StripedHyena-7B, mixtral-7b-8expert
00:44:05 MixtureFeedForward
00:58:25 AT&T feud, billy naming
01:01:35 underlying buffer is not writable
01:04:35 read only memoryview
01:05:40 ctypes const pointer
01:07:05 ctypes from buffer
01:08:20 asking Quentin
01:11:19 memoryview readonly
01:14:10 get address from memoryview
01:15:33 using a read-only buffer with ctypes from_buffer
01:16:30 b partition in honor in oneplus phone
01:16:44 worst customer experience
01:19:05 HIP=1 DEBUG=2, realize() struggle
01:24:50 output of a gate
01:29:21 easy multigpu use with tinygrad
01:34:20 you need to run all the experts
01:34:50 stupid questions about buying tinybox
01:36:20 emotional regulation, don't be evil, us government
01:38:20 only batch size 1
01:42:30 print top, norm
01:45:20 rocm-smi, htop
01:48:50 bug in AMD gpu
01:49:00 multiple $1k bounties in tinygrad
01:51:00 libhsa-runtime64
01:53:20 rude to blame people
01:55:00 HIP runtime rewrite
01:57:25 first token 330
01:58:55 what is Mistral going to say
02:00:00 GPU usage explained, letting it speak
02:03:10 tinybox can run the model, mac can't
02:04:00 pushing to github
02:06:30 letting non-subs talk
02:06:44 AMD bug
02:07:15 real use of tinybox
02:07:40 mistral getting attention
02:08:00 how did we split them
02:08:10 quantization performance
02:08:20 Cmulator 50 gifted subs
02:10:10 bug with errors with the micro engine scheduler
02:10:32 bug hip compilation inexplicably takes 5x longer then CL compilation because of different code path
02:10:40 open source, submiting pull requests and CI
02:10:55 tinygrad beating cuda, mmap nvme copy in to gpu bug
02:12:45 opinion on triton openai, delivering tinyboxes
02:13:50 best Mixtral content so far
02:17:55 cross gpu overhead, copies between gpu
02:22:45 working on synchronize things across gpus
02:23:00 GO, C speed, fixing the algorithms for speed
02:23:55 non-sub asking for moderator, tinygrad CEO
02:26:18 gpuctypes, GPT2 below 2ms per token
02:28:10 three new $1k bounties
02:30:30 never buy anything on oneplus store
02:31:40 Elon thoughts, famous, opinions, Taylor Swift, Sky Bri
02:34:08 small tasks needed in tinygrad
02:34:30 conferences, networking
02:36:16 oneplus not open anymore, verified image programming
02:37:00 contribute to tinygrad, where to start

Official George Hotz communication channels:
- https://geohot.com
-   / realgeorgehotz  
-   / georgehotz  
- https://tinygrad.org
- https://geohot.github.io/blog
- https://github.com/geohot

We archive George Hotz and comma.ai videos for fun.
Follow for notifications:
-   / geohotarchive  

Thank you for reading and using the SHOW MORE button.
We hope you enjoy watching George's videos as much as we do.
See you at the next video.

show more

Share/Embed