@Rob_Mulla

4x Kaggle GM | ML Fameworks @Google - Views are my own.

 Washington, DC
bit.ly/3CzriFa
Joined April 2008
  • Tweets1.2K
  • Following721
  • Followers6.1K
326Media
This is big.
Google Cloud
@googlecloudOct 21
Native PyTorch on TPUs is here—and TorchTPU is officially open source. Train and serve models using the tools you already know: @vllm_project, @sgl_project, @_llm_d_, TorchTitan, @huggingface, and @raydistributed. Learn more from #PyTorchConf2026 → goo.gle/4hWxEVH
1
19
887
TorchTPU is live today! In this video I explain how it works! youtu.be/u4n558Ul6w0?si=uNwG…

Introducing TorchTPU: Native PyTorch on Google Cloud TPUs

TorchTPU is an open-source, native PyTorch backend for Google Cloud...

youtube.com
1
3
520
Rob retweeted
PyTorch
@PyTorchOct 20
Replying to @Google
@Google & @Meta launch TorchTPU for native PyTorch support on Google TPUs! 🔥True eager execution & torch.compile backend ⚡Fast inference via vLLM & SGLang 🏋️Near-linear scaling for Llama 3 405B & Qwen 3 MoE via TorchTitan Read more here: bit.ly/4hV19Ho
24
42
20
279
16,872
Had a blast giving the demo. Stop by the booth to learn more about TorchTPU!
Google Open Source@GoogleOSSOct 20
Join Rob Mulla at 10:40 AM for our sponsored demo: Unifying & Accelerating Open-Source LLM Serving on Google Cloud TPUs with TorchTPU! See how it provides a unified backend for vLLM & SGLang, deploying SOTA models with minimal code changes.
7
285
Join Rob Mulla at 10:40 AM for our sponsored demo: Unifying & Accelerating Open-Source LLM Serving on Google Cloud TPUs with TorchTPU! See how it provides a unified backend for vLLM & SGLang, deploying SOTA models with minimal code changes.
1
1
1
4
1,632
Going to be in the Bay Area for PyTorchCon next week!!! Lots of cool things dropping and I’ll be presenting at the Demo Theater! Stop by the Google booth and hang out.
2
11
605
New kaggle logo is pretty dope.
2
3
1
59
2,647
Great work by the team! TPUs + vLLM is a beautiful combination.
SemiAnalysis
@SemiAnalysis_Oct 15
BREAKING NEWS🚨: Google has submitted its FIRST TPUv7 FP8 vLLM results to our Agentic Inference Benchmark and has demonstrated 1.67x better perf per dollar than even GB300 NVL72 FP4. 🚀 It is amazing to see Google go full steam ahead with externalizing the power of TPUs outside of Google.👇️ (1/4)🧵
8
563
New Kaggle logo WHAAAAT?
1
1
5
1,251
🔥
1
295
Remember when everyone was predicting 2025 to be the year of the AI agent? Kinda delivered I’m not going to lie.
306
🔥
Google
@GoogleSep 30
Replying to @Google
Gemini 4 Argon is our next era of frontier intelligence. It shows significant improvements across benchmarks, setting a new state of the art for real-world long-horizon software engineering tasks.
4
622
The biggest plays of the 1pm games #1: C.Ward intercepted by J.Holland (Titans at Giants). Giants went from 32% to 100% to win. Which one hurt most? #NFL
Ranked list of the plays that moved win probability most: C.Ward intercepted by J.Holland, Giants +68 points; J.Burrow sacked, fumbles, recovered by S.Joseph, Steelers +37 points; D.Watson pass to H.Fannin for 14 yards, TOUCHDOWN, Browns +36 points; S.Darnold intercepted by K.Medrano, Commanders +34 points; S.Shrader 53 yard field goal is GOOD, Colts +33 points.

ALT Ranked list of the plays that moved win probability most: C.Ward intercepted by J.Holland, Giants +68 points; J.Burrow sacked, fumbles, recovered by S.Joseph, Steelers +37 points; D.Watson pass to H.Fannin for 14 yards, TOUCHDOWN, Browns +36 points; S.Darnold intercepted by K.Medrano, Commanders +34 points; S.Shrader 53 yard field goal is GOOD, Colts +33 points.

1
1
110
Virginia Tech beats Boston College 21–14 Final: Virginia Tech 21, Boston College 14 Total yards: Virginia Tech 370, Boston College 261 Rate the performance.
Game recap: Final: Virginia Tech 21, Boston College 14. A waveform shows every play by quarter as a bar in team color, taller for more yards, with touchdowns, field goals and turnovers marked. Total yards: Virginia Tech 370, Boston College 261.

ALT Game recap: Final: Virginia Tech 21, Boston College 14. A waveform shows every play by quarter as a bar in team color, taller for more yards, with touchdowns, field goals and turnovers marked. Total yards: Virginia Tech 370, Boston College 261.

1
1
116
Whoa, cool viz. Whoever is behind this must be supercool.
Squib
@squibchartsSep 26
Michigan won from <1%. Then lost from 99%. Michigan was 99% to win with 0:28 left in Q4. Final: Iowa 20, Michigan 19. Sep 5 vs. Western Michigan: <1% with 0:19 left in Q4, won 13–12. Win probability, every play 👇
Line chart of Michigan's win probability after every play, by quarter. Games: Sep 26 vs. Iowa, L 19–20; Sep 5 vs. Western Michigan, W 13–12. Michigan was 99% to win with 0:28 left in Q4. Final: Iowa 20, Michigan 19. Highest point: 99% with 0:28 left in Q4.

ALT Line chart of Michigan's win probability after every play, by quarter. Games: Sep 26 vs. Iowa, L 19–20; Sep 5 vs. Western Michigan, W 13–12. Michigan was 99% to win with 0:28 left in Q4. Final: Iowa 20, Michigan 19. Highest point: 99% with 0:28 left in Q4.

5
935
Keep pandas name out of your mouth!
Mark Tenenholtz
@marktenenholtzSep 25
Using pandas over polars? In 2026?
1
8
1,919
🚀
Wall St Engine
@wallstengineSep 24
Google will launch an experimental AI satellite on Oct. 1 as part of Project Suncatcher, its effort to eventually build AI data centers in space. The satellite will carry 4 $GOOGL TPUs and enough compute to answer simple Gemini queries from orbit while testing whether AI chips can reliably operate in space. Google plans two more satellite launches in 2027 and is already designing potential fleets of 80+ satellites that could work together as an orbital AI data center. Google estimates space-based data centers could become cost-competitive with terrestrial facilities by the mid-2030s. Source: NYT
6
671
ALERT ALERT ALERT 🚨 🚨 🚨 VLLM MAINTAINERS HAVE JUST SHOWN THAT TPUv7 CAN GET 700 tok/s/user,  56% BETTER PERFORMANCE THAN NVIDIA GB200 NVL72 THROUGH MEGAKERNEL OPTIMIZATION ON KIMI K3. As we said awhile ago, the TPU externalization of software is full steam ahead. This is ultra important to follow the progress of this.
66
112
38
1,579
297,403
One year at Google today! Doesn't seem like its been that long- but looking back I was able to speak at Google Next and I/O, support a growing community of TPU Builders, and work closely on scaling TPU inference with vLLM. Lots of cool new things coming around the corner too.
1
18
1,059
Replying to @googlecloud
@googlecloud and Inferact are announcing today a partnership to make TPU a first-class citizen in @vllm_project. This partnership puts both teams on one engineering roadmap to bring TPU to the broader open model ecosystem, optimizing vLLM as the agentic production serving engine for TPU: • Production serving features and optimized kernels • A native PyTorch path via TorchTPU • Moving towards day-0 support for frontier model releases We're also launching a community program: shared TPU capacity for open-source contributors, plus dedicated review and design help from the core vLLM maintainers at Inferact. Everything this collaboration produces is open source. Read the full announcement: inferact.ai/news/google-tpu-…
10
38
9
192
105,626
Load more