Kimbo

2.2K posts

Kimbo banner
Kimbo

Kimbo

@kimbochen

Learning how systems run fast

NYC Katılım Ekim 2014
867 Takip Edilen2K Takipçiler
Kimbo
Kimbo@kimbochen·
Every time I read Aleksa’s post, I’m in awe of how much better info can be presented It is in fact like what is in his profile name A skill issue
Aleksa Gordić (水平问题)@gordic_aleksa

New in-depth blog post time: "Inside TPU and GPU Clusters: The Anatomy of Collective Communication". If you want to deeply understand the core primitives behind scaling the training / inference for MoEs and dense transformers, going a level below FSDP, expert parallelism, data parallelism, model/tensor parallelism this might be a fun read. I cover: * TPU cluster topology: (super)pods, slices, DCN, PCIe, ICI * All-Gather: 1D/2D rings, and path algo (lots of visuals so should be crystal clear how these work even if you're not a perf engineer) * Reduce-Scatter (which is the dual of AG) and All-Reduce * All-to-All (used to dispatch tokens to target experts in MoEs) * NVIDIA GPU cluster topology (reference DGX architecture): nodes, scalable units, fat tree * GPU collectives within the node: rings, trees (log2 steps), and SHARP (in network compute unit) * GPU collectives across nodes, hierarchical algorithms over InfiniBand etc. I was heavily inspired to do this deep dive after reading the excellent Scaling book by an excellent group of people @jacobaustin132 @_sholtodouglas @reinerpope and others! What originally started as "let me maybe just make four figures covering All-Gather, Reduce-Scatter, All-Reduce, and All-to-All so I can understand them better, it shouldn't take more than a day, right, right?" somehow turned into this 40 figures later. Along the way, I realized that the collective algorithms only really make sense once you understand the underlying hardware topology. TPUs were a bit easier to reason about, but I couldn't skip GPUs, I love them too much. Rings are cool, but I also wanted to understand tree algorithms. But also SHARP, and fat trees, and hierarchical collectives. :') So the scope slowly expanded, and little by little, this blog post came to fruition. Just a side-quest. Hope you like it! :) --- Also a big thank you to my friends for reviewing the blog and providing feedback: * @ArunDemeure (prev GPU/AI stuff at Magic, GPU architect at Apple and Imagine, my llm.c buddy!) * @axel_s_feldmann (making GPUs go brrr at Jane Street, we met for the first time at @marksaroufim's excellent GPU mode event) * @pranjalssh (ex xAI GPU wizard, one of two people who inspired my original matmul blog!)

English
1
0
11
1.7K
George Grigorev
George Grigorev@iamgrigorev·
Very excited to see Kimi opinion on sparse attention, hyper connections and improvements to MoE recipe in their upcoming model. I find their models a bit more practical for other researchers compared to DS.
English
3
0
25
1.8K
stochasm
stochasm@stochasticchasm·
@iamgrigorev tbh would have said the "practical" thing about deepseek before v4. but very excited to see this as well
English
1
0
2
336
Kimbo
Kimbo@kimbochen·
@_bg_chun Haha yes I did Naver was the one that kept saying I took the wrong train direction
English
0
0
0
13
Kimbo
Kimbo@kimbochen·
wtf is wrong with the gps in Korea All maps tell me I’m going the opposite direction as I actually am Thought I was going crazy Blue is where I am, red is GPS thinks where I am
Kimbo tweet media
English
4
0
7
811
Kimbo retweetledi
Matej Sirovatka
Matej Sirovatka@m_sirovatka·
I can't stress enough how important it is to read this
Aleksa Gordić (水平问题)@gordic_aleksa

New in-depth blog post time: "Inside TPU and GPU Clusters: The Anatomy of Collective Communication". If you want to deeply understand the core primitives behind scaling the training / inference for MoEs and dense transformers, going a level below FSDP, expert parallelism, data parallelism, model/tensor parallelism this might be a fun read. I cover: * TPU cluster topology: (super)pods, slices, DCN, PCIe, ICI * All-Gather: 1D/2D rings, and path algo (lots of visuals so should be crystal clear how these work even if you're not a perf engineer) * Reduce-Scatter (which is the dual of AG) and All-Reduce * All-to-All (used to dispatch tokens to target experts in MoEs) * NVIDIA GPU cluster topology (reference DGX architecture): nodes, scalable units, fat tree * GPU collectives within the node: rings, trees (log2 steps), and SHARP (in network compute unit) * GPU collectives across nodes, hierarchical algorithms over InfiniBand etc. I was heavily inspired to do this deep dive after reading the excellent Scaling book by an excellent group of people @jacobaustin132 @_sholtodouglas @reinerpope and others! What originally started as "let me maybe just make four figures covering All-Gather, Reduce-Scatter, All-Reduce, and All-to-All so I can understand them better, it shouldn't take more than a day, right, right?" somehow turned into this 40 figures later. Along the way, I realized that the collective algorithms only really make sense once you understand the underlying hardware topology. TPUs were a bit easier to reason about, but I couldn't skip GPUs, I love them too much. Rings are cool, but I also wanted to understand tree algorithms. But also SHARP, and fat trees, and hierarchical collectives. :') So the scope slowly expanded, and little by little, this blog post came to fruition. Just a side-quest. Hope you like it! :) --- Also a big thank you to my friends for reviewing the blog and providing feedback: * @ArunDemeure (prev GPU/AI stuff at Magic, GPU architect at Apple and Imagine, my llm.c buddy!) * @axel_s_feldmann (making GPUs go brrr at Jane Street, we met for the first time at @marksaroufim's excellent GPU mode event) * @pranjalssh (ex xAI GPU wizard, one of two people who inspired my original matmul blog!)

English
4
2
68
7.1K
Chaim Eisenberg
Chaim Eisenberg@Chaimeisenberg·
לרגע חשבתי שהאיראנים יורים על קוריאה. Fml
Chaim Eisenberg tweet media
עברית
1
0
1
401
Kimbo retweetledi
Binfeng Xu
Binfeng Xu@billxbf·
While leading RL frameworks like @slime_framework shine at RL scaling through Megatron + SGL, we find Megatron growing increasingly heavy and hard to hack on for researchers. So we built an alternative: Molt 🦋. Molt 🦋 is fully PyTorch-native + vLLM with minimally curated interfaces. It supports fully async, R3, TP + EP + CP, TITO, and multimodal training for MoE RL at scales of up to 1T parameters—all in ~8k lines of code (roughly 1/3 the size of Slime and 1/9 of Verl). Start hacking! 🚀
Jian Hu@hijkzzz

1/ Still looking for a minimalist, high-performance framework for agentic RL research? Meet Molt — an agentic-first, PyTorch-native reinforcement learning framework with roughly 9K lines of RL code for 700B models. ⭐ github.com/NVIDIA-NeMo/la…

English
6
33
300
23.1K
Kimbo
Kimbo@kimbochen·
@ar0cket1 Ah makes sense thanks Where did you see IcePop style objectives, but no old training policy correction? Would love to read more
English
1
0
0
39
ar0cket1
ar0cket1@ar0cket1·
@kimbochen you drop old policy so you avoid recomputing logprobs on old trainers, so you can avoid fetching and recomputing. its the drop old policy part
English
1
0
1
49
Kimbo
Kimbo@kimbochen·
Single-rollout Async Optimization (SAO) shares things in both modern async RL and PPO Async RL side, SAO uses IcePop-style double-clipped IS, token-level masking. They made it a big deal of dropping the old policy, which surprises me because I thought that was the norm. PPO side, SAO modifies the General Advantage Estimation (GAE) and the value model training. GAE skips the tool call result tokens to reduce the noise of model output to tool call output token boundaries. Value model training SAO freezes attn layers and updates the value model more frequently than the policy model to reduce variance. Not sure how they get the gradients though. Anyway, I’m doing this so that I can post this meme.
Kimbo tweet media
English
3
0
15
1.1K
Chinmay
Chinmay@ChinmayKak·
PPO has separate policy and value objectives. The authors simply take two optimizer steps on the value loss for every one optimizer step on the policy loss. During those value-only steps, no policy loss is computed, so there are no policy gradients to discard. This is straightforward if the actor and critic are separate models; if they share a backbone, the implementation has to isolate the critic parameters or accept that value updates also move the shared representation.
English
1
0
1
43
Kimbo retweetledi
Hao AI Lab
Hao AI Lab@haoailab·
🚀 Can Attention-FFN Disaggregation still win on the newest rack-scale GPU systems? We built FastAFD, an open-source AFD runtime for GB200 NVL72: 72 Blackwell GPUs in one NVLink domain. It improves per-GPU decode throughput by 1.35-1.45×. 🧵 Code: github.com/hao-ai-lab/Fas… Blog: haoailab.com/blogs/fastafd/
GIF
English
5
33
154
20.6K
Kimbo
Kimbo@kimbochen·
@ChinmayKak I’m not familiar with PPO Do you mind explaining how they update the value model more frequently than the policy model? Does it mean they drop the gradients for the policy model?
English
1
0
0
76
Chinmay
Chinmay@ChinmayKak·
@kimbochen they just do k optimiser steps to update faster
English
1
0
3
173
ar0cket1
ar0cket1@ar0cket1·
@kimbochen not recomputing icepop has been done before, the savings are pretty big
English
1
0
1
90
Kimbo
Kimbo@kimbochen·
Meet your (Prime Intellect) moots lmao I’m just collecting Prime people atp
Kimbo tweet media
English
2
1
25
889
Kimbo
Kimbo@kimbochen·
Oh also thank you AMD Jesus for taking the photo lol
English
0
0
0
113
Kimbo
Kimbo@kimbochen·
Meet your moots Great seeing moots in person
Kimbo tweet media
English
3
0
22
1.1K