CTO, Co-Founder, @tryArcade | @redisinc @HPE_Cray | AI, LLM, Python, HPC, ⚾️ fan, roll tide | Blog: partee.io | Opinions are my own.
San Francisco, CA
- Tweets 1,385
- Following 463
- Followers 1,649
- Likes 3,972
Ok @langchain peeps, I bring you 🥁
ArXiv ChatGuru!
Put in a topic, number of papers, and chat with a guru who knows all about them.
Features:
▶️ Langchain-based QnA
▶️ ArXiv API loader
▶️ Semantic caching
▶️ @Redisinc vector db stats
▶️ LLM and retrieval controls
8
51
340
95,695
📣 Life update!! 📣
I’ve decided to step down from my position at Redis to start my own company.
If you love building software and enjoy working harder than anyone you know, give me a shout.
Time to build 🚀
39
4
218
39,131
With great, "new" programming languages like #Python, @ChapelLanguage, and @JuliaLanguage - Why are the #Climate models that help us confront generational challenges like climate change written in a language from 1950?? 👇
partee.io/2021/02/21/climate…
16
38
141
I know a ton of amazing engineers that would not perfectly explain these terms.
Lots of people love to gatekeep being a software engineer.
If you’re thinking about it, or new to it and struggling, don’t let posts like this get you down.
You’re a great engineer if you know the definition of:
- idempotent
- monoid
- decoupled
- dependency injection
- unit
- functional programming
- asynchronous vs parallel programming
- thread locking
- eventual consistency
- exactly-once semantics
- lambda vs kappa architecture
- push vs pull architectures
- write-audit-publish pattern
What else would you add?
#dataengineering
#softwareengineering
1
4
59
10,416
My new way of looking for Arxiv papers in my field?
@langchain + @Redisinc
Here's the full, simplified, example:
4
13
61
10,032
When I was in HPC, I thought, why isn't there anything better than MPI?
Then, I saw 50+ ML frameworks,"invent" the same thing. I realized..
It's not the framework, it's the definition of the collective action and how much you can optimize
few will get this and I don't care.
6
4
53
12,278
The number of people using @Redisinc and @langchain has absolutely exploded 🚀
Spent last couple weeks on improvements to our integration that just landed in main.
github.com/langchain-ai/lang…
Described in thread 👇
3
18
46
32,725
Big news 🚀
We at #Redis have put out a client library specifically focused on vector database use cases.
Introducing 🥁 Redis Vector Library ("RedisVL")
The client library for using Redis as a vector database.
docs: redisvl.com/docs/html/index.…
github.com/RedisVentures/red…
1
8
39
5,402
If you want to know more about how vector databases are used for context/information retrieval with LLMs come to the @mlopscommunity event!
⏲️11am on April 13th
I'll cover @Redisinc's new @langchain (@hwchase17) integration, our upcoming work with @NVIDIAAI, and more! :)
2
5
34
5,958
Latest with @langchain
Any topic on Arxiv, pulls the papers, creates an index, and then you can chat with them. If you use HF models and @Redisinc docker it's free.
Fun way to explore the effects of context and similarity in RAG apps.
Yes, I'm reposting my own tweet. Merging this in later this week.
@hwchase17 DEMOOOO
It's still highly prone to over utilize context, but it's not that bad. You need to help me with my prompts next time I come over.
@lizziepika should we put this behind a @twilio API so people can text it?
4
36
15,652
#OpenSource your project they said.
Users will help you develop it, they said.
They will read the docs.. they said.
Many humbling hours debugging on each snowflake #HPC system and O(100) code-less feature requests later..
I am very #tired
4
4
28
Anyone who has ever setup a github action has at least 100 commits along the lines of
git commit -m "fix ci"
8
3
32
3,202
Still very early and lots of features to build in like conversational memory and more LLMs, but it's pretty fun!
Spin it up with Docker too! Heres the codebase:
github.com/RedisVentures/ArX…
2
8
30
1,839
#HPC obsession with speedups. Cloud will win out if HPC keeps focusing on raw speed instead of usability and productivity. People sleep on the fact that both reduce time to solution. One just doesn’t sound as good in a paper.
1
5
26
Has anyone else noticed the wave of academics in #climate/#ai space going to private companies??
@nvidia with @SciPritchard and @NoahBrenowitz
@Google with @raspstephan and others who’ve been there.
It’s like everyone from my last lit review now works at an AI powerhouse. 🤔
2
27
Only in SF can I walk 5 blocks and listen to great speakers talk about my field while drinking free beer.
Thanks @bobvanluijt, @weaviate_io, @UnstructuredIO and @deepset_ai
Fun stuff
1
1
29
2,709
SmartSim is now open source!
github.com/CrayLabs/SmartSim
What is SmartSim? 🧵
2
7
24
I always just “dumb it down”
Embedding = List of numbers
Each number means something about the subject of the embedding (I.e picture of a cat)
They are useful because this representation is very compact (space) but full of information (what color is it?)
24
2,771
Little late on the ball, but I’m excited to announce that my paper on Model Parameter Optimization has been accepted to @NeurIPSConf at the machine learning for physical science workshop!
1
1
27
Finally we unveiled our work with @NCAR_Science today!
The first ever #AI augmented Ocean model (MOM6) used for climate modeling with our library, SmartSim. This is what I mean by “next-gen” #HPC and #AI workloads.
Watch it run in the 🧵 below!
So excited about this @HPE_Cray system for @NCAR_Science. The heterogeneous mix of @AMD EPYC and @NVIDIA A100 is perfect for next-gen #AI and #HPC workloads. Even more excited to talk about our teams collaboration with @NCAR_Science this Tuesday.
👇
hpe.com/us/en/newsroom/press…
1
7
25
So this is something I think about alot being that I develop on #AMD and #Nvidia. Being a developer in #HPC means you often have access to machines that have both. The fact that the market believes in #Nvidia isn't saying anything about x86 imo.
a thread ->
Nvidia is now worth more than Intel + AMD combined.
That's the market's way of saying: x86 is a legacy architecture.
The workload of the future is SIMULATION.
Light -> raytracing
Physics -> n-body, CFD
AI -> neural nets
..and these will run on GPUs.
1
7
24
My talk for @NVIDIAGTC was accepted!
We took a recommendation system pipeline from ~30 infer/sec to greater than ~500 infer/sec by optimizing feature storage and vector search with @Redisinc. All the code/models will be open sourced too :)
link-> tinyurl.com/2swpbn9u
4
22
2,740
Love to see all these awesome people hacking away on LLM apps with
@Redisinc
@RelevanceAI_
Great job @mlopscommunity
1
22
1,442
So excited about this @HPE_Cray system for @NCAR_Science. The heterogeneous mix of @AMD EPYC and @NVIDIA A100 is perfect for next-gen #AI and #HPC workloads. Even more excited to talk about our teams collaboration with @NCAR_Science this Tuesday.
👇
hpe.com/us/en/newsroom/press…
2
4
21
YESS!
So happy to have this work get out there, let alone be mentioned by Jensen in a GTC keynote. 😳
Working with @NVIDIA has been a blast and now we get to work together on vector search in Redis. Just awesome.
Big things to come :)
Thank you @NVIDIAAI and Jensen Huang for including #Redis as part of the #GTC23 keynote. We’re excited to join the NVIDIA RAPIDS™ integrations for #vectorsimilarity search. Find out more about Redis as a vector database and our work with NVIDIA here ➡️ bit.ly/3JTAskj
1
20
1,260
New paper on arXiv! Our work with @NCAR_Science
Using Machine Learning at Scale in HPC Simulations with SmartSim: an Application to Ocean Climate Modeling
Paper: arxiv.org/abs/2104.09355
Code: github.com/CrayLabs/NCAR_ML_…
More to come.
11
16
personally, I think Lumi is by far the coolest looking supercomputer I’ve ever seen.
#HPC
❄️ Welcome LUMI ❄️
The fastest & most energy-efficient #supercomputer in Europe, 3rd fastest in the world, will be inaugurated today in Kajaani 🇫🇮.
Part of @EuroHPC_JU, thanks to its massive computing capacity Lumi will be a protagonist in the EU #DigitalTransformation ⚡️🚀
1
1
18
Wrote a blog with @nvidia!
Been working on this for a long time, and it's the foundation of my GTC talk around feature storage and vector search for recommendation systems!
Let me know what you think! Repo for assets below :)
developer.nvidia.com/blog/of…
3
1
18
1,605
Came into the office and found two #fortran textbooks in a wooden box on my desk.
I think I have a secret ... mentor?
2
1
16
Our BoF about at #SC21 was accepted!
@NCAR_Science, @argonne_lcf, @ORNL and @HPE_Cray will lead a discussion about the convergence of #AI and #HPC workloads.
also... it's going to be virtual :)
1
15
Ok so @boborado got me on Lambda with an H100 and damn that setup was so easy. I haven't done anything yet, but that setup felt like if apple had a cloud.
What should I do with this? @vboykis I'm thinking I use this for that live-code stream on vector search I've been putting off. local, parallel embed gen?
Tagging people with ideas: @thedanvass @tchutch94 @svonava @softboyjimbo
6
1
18
8,818
Credibility isn’t your problem. It’s exposure. Expose people to more opinions. Display all the sides and let people choose what they think is right. I prototyped a site like this using zt blockchain for tracking opinions. You don’t need a bigger lens you need a brighter light.
15
Latest release of @langchain includes a @Redisinc integration for vector search!
I'll share more on this after i'm done with a little demo but check it out! S/o to @hwchase17 for an awesome project. 🚀
1
16
2,344
Do you use @nvidia triton?
You can now cache inference responses in @Redisinc
p.s. I hate cmake.
developer.nvidia.com/blog/ho…
2
3
15
1,877
Ever wanted to use @raydistributed on #Slurm or #PBS based #HPC systems?
SmartSim can now launch #Ray on #HPC systems in a few lines of #Python (#Jupyter too).
#RL guided simulations anyone?
Thanks to @alrigazzi for this contribution.
#AI #opensource
1
8
13
Ditto! Did a hike today 1 hour outside the city. Glacier lake at the top, “snow lake”. Beautiful.
Seattle is awesome for quick access to nature.
1
15
ML feeling magical today.
@tchutch94 and I put up a demo using @huggingface models and #Redis vector search to make a search bar for (mostly) #ML arXiv papers.
Coolest part is how you can use natural language to search like: "can ML help me get skinnier"
check it out👇
1
5
13
Lots being said about OpenAI DevDay.
I’m going. Should I do some live tweeting about what I’m seeing/reactions??
5
14
5,176
@hwchase17 DEMOOOO
It's still highly prone to over utilize context, but it's not that bad. You need to help me with my prompts next time I come over.
@lizziepika should we put this behind a @twilio API so people can text it?
3
1
15
19,952
Sam: describes how AI model works
Lab: so what speedup do we get?
Sam: it increases the accuracy of the simulation significantly compared to observations and much higher resolution simulations!
Lab: but what speedup do we get?
Sam: it increases the.. f*** it nvm.
1
10
Would you watch a live-coding stream on vector search, start to finish, with @OpenAI / @CohereAI embeddings for a cool side project?
Like if you would attend! Comment things you would want to see!
4
14
1,877
Speaking of! People should check out this example recently done by @tchutch94 which shows @Redisinc being used as a vector database AND semantic cache with @langchain and @OpenAI
Wicked cool.
github.com/RedisVentures/red…
1
14
4,267
Looking for an internship in AI? Want to develop and deploy AI on the worlds best supercomputers? Are you an aspiring data scientist who isn't afraid of C/C++ and parallel programming? Do you want to learn about HPC and AI?
If you answered yes to any of these questions, DM me
5
13
After many requests, @Redisinc vector database now has a @gpt_index integration!
More examples and improvements to follow soon!
Kudos to @gpt_index folks and @jerryjliu0 for building an awesome framework.
LlamaIndex + @Redisinc 🛠️
Redis provides in-built vector storage/search capabilities. Now you can easily use it with LlamaIndex ingestion/indexing/querying! 💪
HUGE shoutout to @SamPartee for helping to land this change.
See full guide here: gpt-index.readthedocs.io/en/…
2
13
2,259
This aged well. My guy has been called up to the Yankees!!!! three levels in 1 year.
I’m going to continue to be pumped for him because I know he’s just treating this like another day at work right now.
Go get yours Steve!
Imagine being a freshman catcher and showing up day 1 for fall ball to find out the dude your catching throws 95-98mph. My reaction -> 😳
Steve has unrivaled work ethic. People sleep on his off-speed too.
Pumped to watch @ridings17 in pinstripes.
12
Exciting news!
Next week, I'll be speaking at #RedisConf2021 about SmartSim's usage of Redis + RedisAI
Also excited to present at PyTorch Ecosystem Day at FB about using #PyTorch in #HPC
And the weather in Seattle is finally turning around, which is huge. 😁
3
11
Almost done. But with all arxiv papers.
github.com/RedisVentures/red…
1. Enter topic
2. Select number of papers
3. Hit enter
4. Chat with those papers.
Uses @langchain, @Redisinc, and you can change the number of context pieces, tokens, etc.
Next steps, conversational memory, and semantic caching.
And better yet, no API signup needed, just docker :)
3
2
12
1,187
Takeaways from my first international conference:
1. Don’t wear dress shoes
2. Listen
3. The attitudes of curiousity and humility exponentially increases the amount learned.
4. If you start talking about football, be prepared to get asked about the champions league.
12
SmartSim 0.3.2 is released!
Readme and docs have all new examples!
- #HPC ex on #Slurm, #LSF, #PBS and more
- #ML ex with @PyTorch @TensorFlow #keras #scikitlearn
- Online analysis ex
- Specific ex for @ORNL and @NCAR_Science users
github.com/CrayLabs/SmartSim
1
1
11
Somehow after working at Cray/HPE for over half a decade, I never knew that the original #Cray research was founded on on the same day as my birthday.
#HPC #TIL
Today marks the 50th anniversary of the founding of Cray!
Cray Research was founded in Chippewa Falls, WI, on April 6, 1972.
@HPE_Cray is the latest instantiation of the company founded by the father of #supercomputing, Seymour Cray.
ALT Logo of Cray Research. In Black letters, it says "CRAY," then below it, white letters in a black rectangle say, "RESEARCH, INC."
4
11
Had an awesome time at @QCon talking about LLMs, RAG, and @Redisinc!
Thanks to Hein Liu (@DoorDash) for the invite!
Special shoutout to the @sourcegraph team whose raffle I won for this wicked WASD keyboard 😍
2
12
914
Nvidia has put in the work to make their hardware work well with the most popular software applications. Whether it knows it or not, THATs what the market values. Correctly co-designed hardware and software with useful applications in multiple verticals.
12
So now my distributed training stack looks like
- Data Parallel DL: PyTorch DDP/Horovod/Ray
- Model Parallel DL: Deepspeed
- ML/Tabular: Ray + X
- Data: Ray/Dask/Rapids
It may look different in a year, but given the trend here one thing is almost certain: it'll be easier.
2
11
Awesome to see @svonava and @superlinked get featured as a public use case for @Redisinc vector database.
redis.com/customers/superlin…
Also, look at this adorable picture of us with @chipro and @JackyGKoh at the LLM Stack hackathon!
1
11
1,065
This is how it handles conflicts. Usually means to take a look at deps (requirements) and pin better versions if possible.
Do a ‘pip freeze’ and checkout the versions it lands on. Use those versions to pin requirements to.
1
11
3,113
If you're in SF, you should come to the LLM Stack Hackathon being run with @RelevanceAI_ and @mlopscommunity There are spots for a few more teams.
Check out some of the awesome work people did in a recent hackathon. 🧵
3
3
10
691
Shoutout to @tchutch94 for being a core contributor to this and to people like
@thedanvass
@JackyGKoh
@hwchase17
@jerryjliu0
@simba_khadder
@svonava
for being sources of inspiration and help with this and growing the Redis vector database community.
1
10
260
Very excited for this! I’ve been using @pangeo_data ecosystem tools for years now. Happy that we (@HPE_Cray) can contribute back in some way.
We’ll discuss SmartSim, our latest paper, and our future research directions!
#HPC #AI #OpenSource
✅ Don't forget! Pangeo Showcase tomorrow (6/16) at at 12pm ET with @SamPartee from @HPE_Cray on "SmartSim: Online Analytics and Machine Learning for HPC Simulations" pangeo.io/pangeo-showcase.ht…
1
9
The market values nvidia because of CUDA. It's one thing to make great hardware, but without the correctly co-designed software and application, hardware is useless. nvidia is winning because they got the low level software right, early. CUDA is the backbone of many apps.
1
9
Redis! Database you already know with other features besides just vector search. Depends a lot on the use case though.
Here’s some comparison info done by a third party integrator (jina)
docarray.jina.ai/advanced/do…
9
855
Arxiv Paper search demo update!
Can now switch between @huggingface, @OpenAI, and (soon) @CohereAI embeddings.
Super interesting to see the difference in the search results between the embeddings. > 400k papers now too. Try "more like this" too.
👉docsearch.redisventures.com/
1
9
744
Very true. You basically have to implement dependency handling in the setup scripts.
It’s comes a long way and it’s better now, but conda and @spackpm are much better for mixed language/compiler/platform builds.
I actually just wrote a piece on my blog about cibuildwheel.
2
8
939
Seeing alot of hate on OSS projects today. Libraries like pandas, polars, hell... the whole pydata ecosystem..
they are built largely in free time, for free, by people like @RitchieVink.
Instead of hating on them, how bout y'all *build something better*. Oh wait, you can't.
So stfu.
Polars is optional people. I made it with passion. Saying it should fail is harsh.
Just don't use it then. :)
9
1,147
We're #hiring!
- Supercomputers with SOTA hardware
- Work the the top scientists and labs in the country
- Stacks of A100s on demand
- Fully #OpenSource
- #ML for science
- Remote if want/need
- 6 month parental leave
DM me! More details below! 👇
1
7
8
Question for people who manage #OpenSource organizations or repos:
What is the #1 thing you did that helped build a community of contributors and users? - Examples? integrations? plain old marketing?
People like @tgamblin or @mrocklin might have good answers. Others?
7
2
7
My first #earthquake ... I thought someone was behind me grabbing my chair..
that was fun.
3
8
@Redisinc + @langchain = 🔥
Redis is working with innovative startups like @langchain to drive forward #genAI. Read how #Redis Cloud is serving the persistent storage needs for their #OpenGPTs project to making it easier to build #RAG-powered apps with LangChain Templates. 👉 bit.ly/3sH71wX
1
8
1,898
Imagine being a freshman catcher and showing up day 1 for fall ball to find out the dude your catching throws 95-98mph. My reaction -> 😳
Steve has unrivaled work ethic. People sleep on his off-speed too.
Pumped to watch @ridings17 in pinstripes.
Stephen Ridings, Yankees’ new 6-foot-8 flame thrower, opens eyes in Somerset Patriots’ Opening Night win - nj.com nj.com/yankees/2021/05/steph…
4
7
Shoutout to @spackpm for the ‘spack compiler find’ command.
When brew decides it wants Clang 13 and GCC 11 now and it breaks everything, this command will literally show you everything you need to know about what’s installed. 🙏🏻
Also, HOMEBREW_NO_AUTO_UPDATE=1
1
8
Great to see @NCAR_Science part of new #AI for #climate modeling initiatives with @NSF.
news.ucar.edu/132809/ncar-wi…
8
Had an experienced engineer at a fortune 50 company today say,
“Yea but it’s just a key-value store”
So this was nice to see.
The sorted set in Redis is such a useful data structure – I really wish something like it was built into standard libraries of programming languages.
redis.io/docs/data-types/sor…
1
8
2,026
Speaking strictly about #AI and #HPC, there are a number of tools you can try to use to make AMDs work with popular libraries. ROCm is getting better. They even have a TF 1.15 branch, but it still doesn't rival CUDA.
- ROCm rocmdocs.amd.com/en/latest/i…
- HIP github.com/ROCm-Developer-To…
2
8
The in-memory database, launched by SmartSim, houses ML runtimes that can be called using the SmartRedis clients.
With SmartSim, scientists can write ML models with #TensorFlow, #PyTorch (and more) in #Python and utilize them in HPC simulations by adding ~3 lines to their sim
7