Romil Bhardwaj

218 posts

Romil Bhardwaj

Romil Bhardwaj

@bromil101

Building SkyPilot @skypilot_org | PhD AI+Systems @Berkeley_EECS @ucbrise

Berkeley, CA Katılım Haziran 2020
195 Takip Edilen460 Takipçiler
Romil Bhardwaj
Romil Bhardwaj@bromil101·
My favorite talk at the Meetup @lbz____
Romil Bhardwaj tweet media
SkyPilot@skypilot_org

AI Infra Meetup with SkyPilot and CoreWeave - what a night! Packed room, great conversations, and solid talks on scaling AI infra across K8s, Slurm, and neocloud. 🌟 Highlights: @lbz____ (@AIatMeta) shared how Meta unified 100k+ GPUs across Slurm clusters with SkyPilot. @Michaelvll1 + @kevin_mingtarja (@skypilot_org) walked through SkyPilot's new features and ran a live multi-cloud demo. @deok_filho (@CoreWeave) broke down SUNK and the SkyPilot x CoreWeave integration. Huge shoutout to all speakers! Thanks to @CoreWeave @wandb for being amazing partners in making this happen. Excited to keep building this AI infra community. More events coming soon.

English
1
1
32
2.8K
Romil Bhardwaj
Romil Bhardwaj@bromil101·
Online RL is one of the harder infra problems in frontier model training - your trainer needs to interact with live inference servers for generation in a tight loop. Slurm is not built for this. Read how @hcompany_ai used @skypilot_org to scale RL across 2,000+ GPUs👇
H@hcompany_ai

🚀 The H Company Tech Stack: Part 1 We are excited to launch a new series of technical deep dives into the AI Tech Stack powering H Company. Over the coming weeks, we’ll be sharing how we build, scale, and optimize the infrastructure behind our Holo frontier models. First up: Unlocking Online RL and AI Workflows on K8s using SkyPilot. (1/5🧵)

English
0
0
5
272
Romil Bhardwaj
Romil Bhardwaj@bromil101·
We gave Karpathy's autoresearch agent a GPU cluster. What we didn't expect: it started exploiting hardware heterogeneity on its own. This used to be weeks of sweep configs and pssh. Now it's "here's my train script, here's my infra, go." 🤯 blog.skypilot.co/scaling-autore…
SkyPilot@skypilot_org

Karpathy's Autoresearch is bottlenecked by a single GPU. We removed the bottleneck. We gave the agent access to our K8s cluster with H100s and H200s and let it provision its own GPUs. Over 8 hours: • ~910 experiments instead of ~96 sequentially • Discovered that scaling model width mattered more than all hparam tuning • Taught itself to exploit heterogenous hardware: use H200s for validation, screen ideas on H100s Full setup and results: blog.skypilot.co/scaling-autore… @karpathy

English
0
0
3
246
David Bar
David Bar@observie·
Running three mjlab trainings concurrently on 3 separate aws machines I didn't spend a minute setting up myself, and I have a feeling this is just the beginning. all thanks to @skypilot_org sky acceleration tips: - use .skyignore to prevent large & unnecessary folders from being sync'ed (think: .gitignore for sky) - autostop after 5 idle minutes to control cost, but keep the disk (down: false) so next launch starts in seconds instead of minutes
David Bar tweet media
English
2
0
20
1K
Romil Bhardwaj
Romil Bhardwaj@bromil101·
@observie @skypilot_org neat! as you scale out, `sky jobs launch` lets you run 100s of jobs in parallel. SkyPilot handles execution and cleanup :) for rank in $(seq 1 100); do sky jobs launch --async -y train.yaml --env RANK=$rank done
English
2
0
2
78
Romil Bhardwaj
Romil Bhardwaj@bromil101·
Awesome work by @kevin_zakka and the Berkeley gang 🎉 Uses SkyPilot to transparently provision GPUs and handle infra, letting researchers focus on the fun bits. mjlab: github.com/mujocolab/mjlab
David Bar@observie

mjlab is RL with fairy dust oh my, it's so much fun It takes no time to setup the training machine, because you literally don't need to setup a training machine. It assigns you an ad-hoc GPU via SkyPilot No containers, and the software is slim: a single git clone and you can start Post-training data is immediately available on wandb, long after the cloud machine was ditched to save cost. checkpoints, onnx and training logs are automatically uploaded Visualization of results is available locally via the viser web viewer if no GPU. no need to remote desktop anything Effectively, it feels like I am training on my MacBook (below is Q1 Hoper learning to fly)

English
1
0
10
2.6K
Romil Bhardwaj retweetledi
Kevin Zakka
Kevin Zakka@kevin_zakka·
mjlab now supports cloud training via SkyPilot. One command launches a GPU instance, syncs your code, trains, and tears down when done. We support 2 modes: direct uv install and Docker. Multi-GPU and hyperparameter sweeps work out of the box. mujocolab.github.io/mjlab/main/sou…
English
3
12
86
12.4K
Romil Bhardwaj retweetledi
SkyPilot
SkyPilot@skypilot_org·
RL post-training needs heterogeneous hardware - beefy GPUs for the trainer, cheap GPUs for rollouts, and high-memory CPU instances for replay buffers. Running it all on top-tier GPUs is wasteful. SkyPilot Job Groups simplifies workloads with heterogeneous requirements: • One YAML to run RL workloads on heterogeneous hardware. • Automatic service discovery • Coordinated creation and shutdown Define each component with the right resources and launch as one unit. Read the blog now: 🔗blog.skypilot.co/job-groups/
SkyPilot tweet media
English
2
5
31
2.3K
Romil Bhardwaj retweetledi
SkyPilot
SkyPilot@skypilot_org·
🚫🦞 DON'T run OpenClaw on your main machine. ✅🦞 DO run it in an isolated environment. But isolation isn't easy: provisioning, SSH, security groups, cost management. This post breaks down your options (Docker, dedicated hardware, cloud VM) and shows you how to launch a secure OpenClaw setup on any cloud with one command. 🔗 blog.skypilot.co/openclaw-on-sk…
SkyPilot tweet media
English
0
2
10
732
Romil Bhardwaj retweetledi
marimo
marimo@marimo_io·
marimo notebooks are reproducible Python programs. @skypilot_org runs Python anywhere. Together, they take notebooks from laptop to cloud with zero rewrites. Learn more at our latest blog: marimo.io/blog/skypilot
English
0
5
20
2.6K
Romil Bhardwaj retweetledi
SkyPilot
SkyPilot@skypilot_org·
SkyPilot now auto-detects and displays external links in your Dashboard, making it easier to quickly access cloud consoles and experiment trackers like @wandb. The dashboard shows links to: - Your instance's cloud console links (AWS, GCP, Azure) - External dashboards from your cluster logs
English
1
4
8
1.1K
Romil Bhardwaj retweetledi
SkyPilot
SkyPilot@skypilot_org·
Moving from Slurm to Kubernetes shouldn't mean rewriting every job script and learning container orchestration. @skypilot_org brings Slurm-like simplicity to K8s: 🔄 6-line task YAML, not 60-line manifests 💻 SSH into GPU nodes just like salloc 🚀 Unified interface for all your Slurm and K8s clusters Same workflow. Modern infra. 🔗 blog.skypilot.co/slurm-to-k8s-m…
SkyPilot tweet media
English
2
3
13
982
Romil Bhardwaj retweetledi
SkyPilot
SkyPilot@skypilot_org·
Congrats to @hcompany_ai on the Holo2 release - the new gold standard in UI localization! 🎉 Proud to power the infra for this SOTA model — H Company migrated their training stack from Slurm to Kubernetes with SkyPilot as the unified interface across their clusters 🚀
H@hcompany_ai

Holo2-235B-A22B: #1 on ScreenSpot-Pro, #1 on OSWorldG 🎯 🚀 Today, we are releasing Holo2-235B-A22B 🤗, our most capable GUI localization model yet! Holo2 now leads on all major GUI localization benchmarks: 78.5% on ScreenSpot-Pro and 79.0% on OSWorld-G!

English
0
2
7
772
Romil Bhardwaj retweetledi
Mikhail Parakhin
Mikhail Parakhin@MParakhin·
Running Training and Inference at scale gets tricky in a hurry. Heterogeneous workloads: large-scale HSTU training, tiny L4 models and the development - all need to be supported seamlessly. Here is a new blogpost on how we approach it at Shopify: shopify.engineering/skypilot
English
5
8
42
10.5K
Romil Bhardwaj
Romil Bhardwaj@bromil101·
Looking to build a multi-cloud AI platform? Read how @Shopify did it: • SkyPilot running on K8s clusters on Nebius and GCP • Custom SkyPilot policy plugin for routing, infiniband and cost-tracking • Kueue for fair-sharing + priority classes (emergency > interactive > batch)
Shopify Engineering@ShopifyEng

Multi-cloud GPU scheduling without losing your mind. Inside our @skypilot_org setup: how we route GPU training workloads across providers, keep scheduling fair, and let engineers stay close to the metal instead of buried in cloud consoles.

English
0
0
2
315
Romil Bhardwaj retweetledi
SkyPilot
SkyPilot@skypilot_org·
Shopify now runs all AI training on SkyPilot. • H200s on @nebiusai, L4s on @googlecloud - multi-cloud AI enabled by SkyPilot • One unified interface for engineers • Cost tracking, Inifiniband support, fair scheduling Excited to support @Shopify's journey to build the future of e-commerce with AI 💪🏻
SkyPilot tweet media
Shopify Engineering@ShopifyEng

Multi-cloud GPU scheduling without losing your mind. Inside our @skypilot_org setup: how we route GPU training workloads across providers, keep scheduling fair, and let engineers stay close to the metal instead of buried in cloud consoles.

English
1
3
12
1.2K
Romil Bhardwaj retweetledi
SkyPilot
SkyPilot@skypilot_org·
AI workloads need fast storage. Object stores aren't always the best fit. SkyPilot Volumes let you access high-performance storage designed for AI. 💾 Fast access to data & checkpoints ⚡ Easy to use - attach w/ 1 line ☸️ Works with any Kubernetes PVC blog.skypilot.co/skypilot-volum…
English
0
1
6
393
Romil Bhardwaj retweetledi
SkyPilot
SkyPilot@skypilot_org·
Scaling batch workloads: run SAM3 video segmentation across K8s and AWS clusters with one command. SkyPilot Pools turns scattered GPU capacity into a unified batch queue. 🌎 Unified GPU pool across multiple K8s + clouds 📈 Fast starts w/ linear scaling blog.skypilot.co/skypilot-pools…
SkyPilot tweet media
English
0
3
5
392
Romil Bhardwaj retweetledi
SkyPilot
SkyPilot@skypilot_org·
🚀 SkyPilot now ships precooked YAML templates for launching clusters with popular frameworks and patterns. Templates are automatically available on all new SkyPilot clusters. With a single command you can launch fully configured environments without having to write any boilerplate code. Here's how to launch a Ray cluster on your infra with a single line of code. blog.skypilot.co/skypilot-templ…
English
0
2
6
402

Keşfet

@lbz____ @hcompany_ai @skypilot_org @observie @kevin_zakka @wandb @Shopify @nebiusai