ATTENTION: vLLM is now ~2x faster than vLLM thanks to agentic-generated fused model definitions!
We think this defines a new future for how models get optimized in inference engines and have shared our plans for a sustainable system/project architecture. PTAL :)
75/100 of GPU Grind
i've been interested in the compiler side of GPU recently, so learning about MLIR
as an exercise i translated a simple cuda kernel to MLIR gpu/memref/scf/arith dialects, mostly by hand (it's tedious for sure, but it forces me to understand how MLIR is emitted much more than if just read through something)
and then i can execute the common lowering pipeline '--gpu-lower-to-nvvm-pipeline' to see what passes it includes and what are their roles
it contains quite a lot of passes to review them all in one time, but some things i found interesting:
- convert-gpu-to-nvvm: converts the gpu operations in a gpu module (a kernel), to nvvm operations. most of these seems to be kinda straightforward translations at least in the kernel i wrote:
gpu.thread_id x becomes nvvm.read.ptx.sreg.tid.x : i32,
gpu.barrier becomes nvvm.barrier
etc..
- nvvm-attach-target: simply adds an attribute to the kernel to specify the chip architecture we are targetting in this compiler pipeline
- gpu-module-to-binary: the final pass that converts the GPU code to PTX
74/100 of GPU Grind
watched the gpu mode lecture from last month from joe fioti (luminal) on production megakernels for inference
he makes his point explaining how cuda traditional kernel programming model is limited (global sync between thread blocks for example), and that is
Today we're announcing that Modal has crossed $1B in annualized run rate, as well as raised a $750M Series D at a $15.75B valuation, led by @Accel.
Since our Series C, usage has tripled and we’ve grown more than 15x year-over-year.
We are deeply grateful to the developers and companies building on Modal. There's still so much more to build, and we can't think of a better time to be building software.
some news: i didn't get the job at anthropic :/
i did well at the technicals but they were disappointed that i was unable to commit to inhaling a pack of unfiltered mongolian cigarettes during the interview, citing a lack of confidence that i truly believed in transformative asi
3K Followers 634 FollowingGPU tools @ Apple. Distributed systems before that. Side project: UchenML, an ML framework for hardware that has no business running models.
52K Followers 577 FollowingAssistant prof. @ Stanford; Chief AI Scientist @ MongoDB; Former Co-founder/CEO of Voyage AI
Working on ML, DL, RL, LLMs, and their theory.
53K Followers 221 Followingidiot. cuda og, harvard val, morgan prize, society of fellows, 1 hilbert problem so far, creating friendly, SAFE, delightful, supergenius ..things @anthropicai
22K Followers 1K FollowingAssocProf @CarnegieMellon. Distinguished Eng @NVIDIA. Creator of @XGBoostProject, @ApacheTVM. Member https://t.co/QYyfjQNp4p, @TheASF. Views are on my own
3K Followers 1K FollowingAncient thunder in a modern sound.
Wired to the earth in a savage ballet.
I don't do TL;DR.
Foldb, WILD, Orbit, other software.
4K Followers 220 FollowingLead Technical Recruiter for @microsoftAI Team EMEA - building the teams that are building the next phase of humanist superintelligence
21K Followers 2K FollowingMachine Learning and Agents at @Nvidia, 6x Kaggle Grandmaster CPMP. Arc Prize winner. ML PhD. Ex ENS Ulm, ILOG CPLEX, IBM. Views are my own.