Loading video...

Video Failed to Load

Go Home

Serving GLM5.2 NVFP4 Agentic Workload with SGLang: How We Reached 500 TPS on 8xB300 at bs=1 In this deep dive, we explain how SGLang reaches 500+ tok/s/user at bs=1 on 8xB300, with 18 to 34% higher single-user interactivity within two weeks since day-0, and 6 to 11% better peak...

lmsysorg's profile picture

LMSYS Org

16,266 subscribers

104,832 views • 9 days ago •via X (Twitter)

Science & TechnologyEducationGaming
LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

Timelapse #156 (36 hrs) - Worked with the tiny corp on getting GLM 5.2 running on 8xMI300X (sglang won here) - Launched KernelBench-Mega and updated Kernelbench-Hard with h100 and b200 sweeps - Took care of boring business stuff - Did some training sweeps for specialized technical vocab audio model - Some bugs with putting kernelbench-mega and hard on cloud instances so had to do some reruns. Learned a lot though - Setting up my own local rl infra and profiling concurrency 128 rollouts with vllm. Became clear to me that I need to serve in nvfp4, use MoE only for throughput, reap so training doesn’t OOM, dig into the vllm kernel graph itself to not underutilize my hardware from poor flashinfer/cutlass selections for my rtx pro 6000 sm120 architecture - Might do online distillation from glm 5.2 but for now taking it one step at a time - Slept for a bit then woke up and showered - Fixed an issue with SGLang tensor parallel deadlock on GLM 5.2 architecture with MTP enabled - GLM 5.2 inference is 2-3x faster than coding plans and running on amd boxes - Spent time with family - Hung out with some friends - Recorded some yoctogpt lectures with the revamped notebook (high taste btw) - Setting up dflash training for GLM 5.2
0:30

Timelapse #156 (36 hrs) - Worked with the tiny corp on getting GLM 5.2 running on 8xMI300X (sglang won here) - Launched KernelBench-Mega and updated Kernelbench-Hard with h100 and b200 sweeps - Took care of boring business stuff - Did some training sweeps for specialized technical vocab audio model - Some bugs with putting kernelbench-mega and hard on cloud instances so had to do some reruns. Learned a lot though - Setting up my own local rl infra and profiling concurrency 128 rollouts with vllm. Became clear to me that I need to serve in nvfp4, use MoE only for throughput, reap so training doesn’t OOM, dig into the vllm kernel graph itself to not underutilize my hardware from poor flashinfer/cutlass selections for my rtx pro 6000 sm120 architecture - Might do online distillation from glm 5.2 but for now taking it one step at a time - Slept for a bit then woke up and showered - Fixed an issue with SGLang tensor parallel deadlock on GLM 5.2 architecture with MTP enabled - GLM 5.2 inference is 2-3x faster than coding plans and running on amd boxes - Spent time with family - Hung out with some friends - Recorded some yoctogpt lectures with the revamped notebook (high taste btw) - Setting up dflash training for GLM 5.2

Elliot Arledge

45,447 views • 1 month ago

SGLang now supports DSpark, enabling confidence-driven, variable-length verification for speculative decoding 🎉 DSpark addresses a key bottleneck under load: instead of verifying every draft token, it verifies only where the draft model is confident, so the gains hold even as batch size scales. We heavily optimized variable-length verification in SGLang. Across batch sizes 1 to 256, DSpark gives the best throughput/latency tradeoff on DeepSeek-V4-Flash, ahead of both MTP and non-spec. At high concurrency, dynamic scheduling provides up to ~20% higher throughput compared to a fixed budget, while maintaining high verification quality across workloads. With fused kernels and zero-overhead scheduling, DeepSeek-V4-Pro reaches 383.7 tok/s at B=1 on B300. DSpark is now available in SGLang with support for Qwen3 and DeepSeek-V4. Thanks DeepSeek for open-sourcing! Blog with full technical details and commands to run below 👇
0:33

SGLang now supports DSpark, enabling confidence-driven, variable-length verification for speculative decoding 🎉 DSpark addresses a key bottleneck under load: instead of verifying every draft token, it verifies only where the draft model is confident, so the gains hold even as batch size scales. We heavily optimized variable-length verification in SGLang. Across batch sizes 1 to 256, DSpark gives the best throughput/latency tradeoff on DeepSeek-V4-Flash, ahead of both MTP and non-spec. At high concurrency, dynamic scheduling provides up to ~20% higher throughput compared to a fixed budget, while maintaining high verification quality across workloads. With fused kernels and zero-overhead scheduling, DeepSeek-V4-Pro reaches 383.7 tok/s at B=1 on B300. DSpark is now available in SGLang with support for Qwen3 and DeepSeek-V4. Thanks DeepSeek for open-sourcing! Blog with full technical details and commands to run below 👇

LMSYS Org

167,269 views • 17 days ago

🎉 Congrats to MiniMax (official) on releasing MiniMax M3! Frontier coding and agentic capabilities, native image and video input, computer use, and a 1M-token context window, all in a single open model. At the heart of M3 is MSA, a new sparse attention architecture: instead of attending densely over the full KV cache, each query scores 128-token KV blocks and runs attention only over the top blocks. That is what makes 1M-token context practical to serve. M3 runs in vLLM with day-0 support, verified on NVIDIA and AMD hardware: ✨ MSA sparse attention with dedicated prefill and decode kernels ✨ 1M-token context serving with prefix caching and chunked prefill ✨ BF16 and MXFP8 checkpoints, with MoE backends for both Hopper and Blackwell ✨ Native multimodal input (image + video) ✨ Tool calling, reasoning parsing, and thinking-mode control for agent workloads Day-0 support like this is a true team effort. Grateful to the teams at MiniMax (official), NVIDIA AI, AI at AMD, and Inferact, and to the vLLM community for making it happen. 🙏 Deep dive into the implementation, kernel work, and deployment recipes: 🔗
0:21

🎉 Congrats to MiniMax (official) on releasing MiniMax M3! Frontier coding and agentic capabilities, native image and video input, computer use, and a 1M-token context window, all in a single open model. At the heart of M3 is MSA, a new sparse attention architecture: instead of attending densely over the full KV cache, each query scores 128-token KV blocks and runs attention only over the top blocks. That is what makes 1M-token context practical to serve. M3 runs in vLLM with day-0 support, verified on NVIDIA and AMD hardware: ✨ MSA sparse attention with dedicated prefill and decode kernels ✨ 1M-token context serving with prefix caching and chunked prefill ✨ BF16 and MXFP8 checkpoints, with MoE backends for both Hopper and Blackwell ✨ Native multimodal input (image + video) ✨ Tool calling, reasoning parsing, and thinking-mode control for agent workloads Day-0 support like this is a true team effort. Grateful to the teams at MiniMax (official), NVIDIA AI, AI at AMD, and Inferact, and to the vLLM community for making it happen. 🙏 Deep dive into the implementation, kernel work, and deployment recipes: 🔗

vLLM

40,306 views • 1 month ago

I love using GLM 5.2 for web app iteration. My workflow: generate 6 variations, then pick the best one and continue iterating on it. I built Recast to make this even easier. Give it a prompt, get 6 variations, download the code for your favorite version, then continue iterating in your own coding agent. GLM 5.2 works really well for this use case since it's fast, cheap, and great at design. On average, you can make 3-6 landing pages with GLM 5.2 for every one you can generate with Opus 4.8. And the design variety of the GLM landing pages helps a lot with trying to think through what style you want to go with. Recast is launching soon (as a web app + skill).
0:34

I love using GLM 5.2 for web app iteration. My workflow: generate 6 variations, then pick the best one and continue iterating on it. I built Recast to make this even easier. Give it a prompt, get 6 variations, download the code for your favorite version, then continue iterating in your own coding agent. GLM 5.2 works really well for this use case since it's fast, cheap, and great at design. On average, you can make 3-6 landing pages with GLM 5.2 for every one you can generate with Opus 4.8. And the design variety of the GLM landing pages helps a lot with trying to think through what style you want to go with. Recast is launching soon (as a web app + skill).

Hassan

20,164 views • 28 days ago

Episode 10 of How I AI is my personal favorite so far: we talk to John Blackman, 91 years old, about using AI to code a web app for the first time in his life. John and his grandson Brandon show us: - Claude for writing user stories - Replit agents for coding a HUGE volunteer support app - the challenges of going from dev to production - user feedback for Anthropic and Replit ⠕ as a beginner If you’ve been sitting on the sidelines, this is your call to action! No excuses, at any age. As always, thanks to our fab sponsors! 💼 WorkOS: make your app enterprise ready today 🔀 Orkes: the platform for reliable apps and agentic workflows
1:17

Episode 10 of How I AI is my personal favorite so far: we talk to John Blackman, 91 years old, about using AI to code a web app for the first time in his life. John and his grandson Brandon show us: - Claude for writing user stories - Replit agents for coding a HUGE volunteer support app - the challenges of going from dev to production - user feedback for Anthropic and Replit ⠕ as a beginner If you’ve been sitting on the sidelines, this is your call to action! No excuses, at any age. As always, thanks to our fab sponsors! 💼 WorkOS: make your app enterprise ready today 🔀 Orkes: the platform for reliable apps and agentic workflows

claire vo 🖤

138,420 views • 1 year ago

Learn how to build an optimized LLM inference system from the ground up in our new short course, Efficiently Serving LLMs, built in collaboration with Predibase by Rubrik and taught by Travis Addair. Whether you're serving your own LLM or using a model hosting service, this course will give you a deep understanding of the optimizations required to efficiently serve many users at once. - Learn how LLMs generate text one token at a time, and how techniques like KV caching, continuous batching, and quantization speed things up and optimize memory usage for serving multiple users. - Benchmark the performance of these LLM optimizations to explore the trade-offs between quickly responding to an individual user’s request vs. serving many users at once. - Use techniques like low-rank adaptation (LoRA) to efficiently serve hundreds of unique, custom fine-tuned models on a single device, without sacrificing throughput. - Use Predibase's LoRAX framework to see optimization techniques in action on a real LLM server. Sign up here:
2:57

Learn how to build an optimized LLM inference system from the ground up in our new short course, Efficiently Serving LLMs, built in collaboration with Predibase by Rubrik and taught by Travis Addair. Whether you're serving your own LLM or using a model hosting service, this course will give you a deep understanding of the optimizations required to efficiently serve many users at once. - Learn how LLMs generate text one token at a time, and how techniques like KV caching, continuous batching, and quantization speed things up and optimize memory usage for serving multiple users. - Benchmark the performance of these LLM optimizations to explore the trade-offs between quickly responding to an individual user’s request vs. serving many users at once. - Use techniques like low-rank adaptation (LoRA) to efficiently serve hundreds of unique, custom fine-tuned models on a single device, without sacrificing throughput. - Use Predibase's LoRAX framework to see optimization techniques in action on a real LLM server. Sign up here:

Andrew Ng

104,727 views • 2 years ago

“This is a problem that begins and ends with men.” In the last part of our special series on violence against women, we hear from the men who are leading the way and calling on others to stand up. Watch #9News at 6pm to find out how you can help keep women safe.
0:54

“This is a problem that begins and ends with men.” In the last part of our special series on violence against women, we hear from the men who are leading the way and calling on others to stand up. Watch #9News at 6pm to find out how you can help keep women safe.

9News Adelaide

11,187 views • 2 years ago

Inkling, Thinking Machines' first open model, dropped today: 975B total / 41B active MoE, up to 1M context, reasoning natively over text, images, and audio. Serving and RL support are already live: you can run and shape it on an open stack, starting now. Day 0 support on SGLang SGLang and Miles RadixArk👇 - Inkling's new architecture (ShortConv, attention with relative positional embedding, shared expert sink MoE) is natively implemented and deeply optimized, with prefill full CUDA graph and MXFP8 KV cache - Full parameter and LoRA RL in a customized Megatron backend, train inference consistency via customized kernels, routing replay, and cross-runtime parameter synchronization - DFlash speculative decoding from Modal for low-latency serving Launching now, blog and cookbook in the comments ⬇️
1:01

Inkling, Thinking Machines' first open model, dropped today: 975B total / 41B active MoE, up to 1M context, reasoning natively over text, images, and audio. Serving and RL support are already live: you can run and shape it on an open stack, starting now. Day 0 support on SGLang SGLang and Miles RadixArk👇 - Inkling's new architecture (ShortConv, attention with relative positional embedding, shared expert sink MoE) is natively implemented and deeply optimized, with prefill full CUDA graph and MXFP8 KV cache - Full parameter and LoRA RL in a customized Megatron backend, train inference consistency via customized kernels, routing replay, and cross-runtime parameter synchronization - DFlash speculative decoding from Modal for low-latency serving Launching now, blog and cookbook in the comments ⬇️

LMSYS Org

139,053 views • 8 days ago

WE'RE BACK! 🇺🇸 BRAVE BOOKS and I are excited to announce our third annual See You at the Library Day on August 16th! This is an opportunity for YOU to join the fight! This is your chance to stand up, speak out, and take part in a nationwide movement to revive our culture, reawaken the heart of our nation, and strengthen the conviction of the next generation. We’re calling on BRAVE individuals like YOU to host a story hour at your local library. As a host, you’ll receive: •BRAVE resources to make your event a success •Support with promoting your story hour in your community •Virtual meeting with me and the BRAVE staff two weeks before the event to answer your questions Learn more about it or sign up to be a host for this exciting, one day only, nationwide event at We are excited to See YOU at the library August 16! 👏🏻👏🏻👏🏻
0:49

WE'RE BACK! 🇺🇸 BRAVE BOOKS and I are excited to announce our third annual See You at the Library Day on August 16th! This is an opportunity for YOU to join the fight! This is your chance to stand up, speak out, and take part in a nationwide movement to revive our culture, reawaken the heart of our nation, and strengthen the conviction of the next generation. We’re calling on BRAVE individuals like YOU to host a story hour at your local library. As a host, you’ll receive: •BRAVE resources to make your event a success •Support with promoting your story hour in your community •Virtual meeting with me and the BRAVE staff two weeks before the event to answer your questions Learn more about it or sign up to be a host for this exciting, one day only, nationwide event at We are excited to See YOU at the library August 16! 👏🏻👏🏻👏🏻

Kirk Cameron

52,134 views • 1 year ago

GLM 4.7 is one of the top open-source models on LM Arena—and it's going toe-to-toe with Claude Opus 4.5 and Gemini Pro. We sat down with Anastasios Nikolas Angelopoulos, co-founder and CEO of Arena.ai, to break down 8,000+ developer votes: → Within 30 points of Gemini Pro in math & coding → Frontier-level multi-turn & instruction following → The open-source model devs are actually switching to The best part? You can run it at 1,500+ tokens/sec on Cerebras—for free.
4:14

GLM 4.7 is one of the top open-source models on LM Arena—and it's going toe-to-toe with Claude Opus 4.5 and Gemini Pro. We sat down with Anastasios Nikolas Angelopoulos, co-founder and CEO of Arena.ai, to break down 8,000+ developer votes: → Within 30 points of Gemini Pro in math & coding → Frontier-level multi-turn & instruction following → The open-source model devs are actually switching to The best part? You can run it at 1,500+ tokens/sec on Cerebras—for free.

Cerebras

28,515 views • 5 months ago

“This (players holding out) is how non-guarantees are gonna come into our league. For shyt like this. We f***in fought our asses off for 5-6 lockouts to get to a point where we protect the integrity of the talent” The league going down a slippery slope with players holding out. I get there may be friction with the higher ups but at the end of the day you a professional under contract to do a certain job. P and I had moments where we didn’t see eye to eye with the leadership of our teams but at the end of the day we still came in and did our job that we agreed to do. Hoping a resolution comes soon or we may be heading down non-guaranteed contracts lord 💯
1:02

“This (players holding out) is how non-guarantees are gonna come into our league. For shyt like this. We f***in fought our asses off for 5-6 lockouts to get to a point where we protect the integrity of the talent” The league going down a slippery slope with players holding out. I get there may be friction with the higher ups but at the end of the day you a professional under contract to do a certain job. P and I had moments where we didn’t see eye to eye with the leadership of our teams but at the end of the day we still came in and did our job that we agreed to do. Hoping a resolution comes soon or we may be heading down non-guaranteed contracts lord 💯

Kevin Garnett

50,064 views • 1 year ago

We've just launched beta for Solana Vault Standard! The first step of Superteam Brasil roadmap of a common, OS, tooling suite to help out the day to day of every type of Solana developer 💻 Our goal is to enable builders to fork and rely on audited code already, similar to how evm devs rely on OpenZeppelin packages - but ready for the agentic world and with more expansive UX 🚀 With SVS, you can create, configure, manage, monitor, audit and power up your yield bearing, compliant and/or private tokenized vault in a single sdk. Check it out at
1:14

We've just launched beta for Solana Vault Standard! The first step of Superteam Brasil roadmap of a common, OS, tooling suite to help out the day to day of every type of Solana developer 💻 Our goal is to enable builders to fork and rely on audited code already, similar to how evm devs rely on OpenZeppelin packages - but ready for the agentic world and with more expansive UX 🚀 With SVS, you can create, configure, manage, monitor, audit and power up your yield bearing, compliant and/or private tokenized vault in a single sdk. Check it out at

kaue

12,014 views • 4 months ago

Robert Saleh was asked on a s scale of 1-10, how much does he want to be a head coach again: “We all want to get to the top of our profession. But honestly, I’m just in this space right now where I’m solely focused on trying to do my best for the guys here. It’s not a secret, we all want to at the top of our profession and show that we can achieve greatness at the top of our profession, but I say this with great humility, my focus is here and just trying our best to get to the playoffs and see if we can make some damage and host a Super Bowl here in our building.” via: JosinaAnderson
0:42

Robert Saleh was asked on a s scale of 1-10, how much does he want to be a head coach again: “We all want to get to the top of our profession. But honestly, I’m just in this space right now where I’m solely focused on trying to do my best for the guys here. It’s not a secret, we all want to at the top of our profession and show that we can achieve greatness at the top of our profession, but I say this with great humility, my focus is here and just trying our best to get to the playoffs and see if we can make some damage and host a Super Bowl here in our building.” via: JosinaAnderson

Coach Yac 🗣

160,001 views • 7 months ago

| $SGT Latest Update is LIVE! At we're dedicated to building an all-in-one DeFi ecosystem that covers everything from the hottest crypto trends and breaking news to top performers and in-depth analysis of call groups and influencer signals. Our latest update is designed to enhance your user experience, making it easier to navigate through our utilities and access critical insights. Here’s What’s New: 🔍 Search Bar: for tokens mentioned on our platform by name or contract hash, streamlining your access to key data. 📊 Token Profile Pages: with essential metrics such as Market Cap, Volume, Price, and Performance Charts. A detailed table at the bottom shows which influencers mentioned the token and how it performed afterward. 📈 Top 5 Visited Tokens: Check out the most viewed tokens on our platform. This new table highlights the tokens generating the most user interest, helping you stay ahead of the trends. 💻 Telegram App & Bots: With our focus on Telegram, we've added dedicated pages for our app and bots, making it easier to explore their features and utilities as we continue to expand our offerings. ⚡️ UI & Speed Enhancements: The platform has been optimized for a faster, smoother experience, with an improved user interface for better navigation. 📺 Watch the full tutorial: Join our community for more news and updates! 🌐 Visit us at :
3:51

| $SGT Latest Update is LIVE! At we're dedicated to building an all-in-one DeFi ecosystem that covers everything from the hottest crypto trends and breaking news to top performers and in-depth analysis of call groups and influencer signals. Our latest update is designed to enhance your user experience, making it easier to navigate through our utilities and access critical insights. Here’s What’s New: 🔍 Search Bar: for tokens mentioned on our platform by name or contract hash, streamlining your access to key data. 📊 Token Profile Pages: with essential metrics such as Market Cap, Volume, Price, and Performance Charts. A detailed table at the bottom shows which influencers mentioned the token and how it performed afterward. 📈 Top 5 Visited Tokens: Check out the most viewed tokens on our platform. This new table highlights the tokens generating the most user interest, helping you stay ahead of the trends. 💻 Telegram App & Bots: With our focus on Telegram, we've added dedicated pages for our app and bots, making it easier to explore their features and utilities as we continue to expand our offerings. ⚡️ UI & Speed Enhancements: The platform has been optimized for a faster, smoother experience, with an improved user interface for better navigation. 📺 Watch the full tutorial: Join our community for more news and updates! 🌐 Visit us at :

ShillGuard $SGT

16,239 views • 1 year ago

Excited to announce that PaalX v0.1, your AI-powered crypto research & trading terminal, is here! 🤖 To kick things off, join us for our NFT mint event, where lucky holders will: 🔹 Enjoy fee-free trading on the PaalX launch 🔹 Unlock advanced trading tools and research features 🔹 Earn passive $ETH from NFT transactions 🔹 Gain voting rights on new features and many more surprises There will be 1,000 NFTs up for grabs, with the first special set of 500 exclusive to whitelisted wallets. 5 of the first 500 NFTs will also come with a special gift package from our partners at CoinGecko! To be whitelisted: 💼 Stake a minimum of 2,000 $PAAL at for a chance at the first 500 NFTs. 🔁 Simply quote this tweet and repost it to enter; winners will be announced before Saturday, March 2nd. The mint event at will open on March 2nd between 12-3pm EST. Huge thanks to our partners CoinGecko for data support and Web3Auth for security measures. Dive deeper into PaalX in the docs link in our bio! #PaalX
0:58

Excited to announce that PaalX v0.1, your AI-powered crypto research & trading terminal, is here! 🤖 To kick things off, join us for our NFT mint event, where lucky holders will: 🔹 Enjoy fee-free trading on the PaalX launch 🔹 Unlock advanced trading tools and research features 🔹 Earn passive $ETH from NFT transactions 🔹 Gain voting rights on new features and many more surprises There will be 1,000 NFTs up for grabs, with the first special set of 500 exclusive to whitelisted wallets. 5 of the first 500 NFTs will also come with a special gift package from our partners at CoinGecko! To be whitelisted: 💼 Stake a minimum of 2,000 $PAAL at for a chance at the first 500 NFTs. 🔁 Simply quote this tweet and repost it to enter; winners will be announced before Saturday, March 2nd. The mint event at will open on March 2nd between 12-3pm EST. Huge thanks to our partners CoinGecko for data support and Web3Auth for security measures. Dive deeper into PaalX in the docs link in our bio! #PaalX

PAAL AI 🧠🤖

756,758 views • 2 years ago

It was another good Christmas for M&S, building on a strong performance in the prior year. Sales records were broken across the business - with Food recording its biggest day and Clothing, Home & Beauty online its biggest week - but we’re not complacent. As a growth business it’s our job to break records. The external environment remains challenging, with cost and economic headwinds to navigate, but there is much within our control. At M&S, we stay close to our customers and their needs, and with that in mind our investment in trusted value, along with great quality, style and innovation remains our priority. Transforming M&S is a marathon, not a sprint, and we go into 2025 shifting up a gear and raring to go as we accelerate the scale and pace of change. Thank you to all of our customers for shopping with M&S this Christmas and to all of our colleagues for their extraordinary effort delivering M&S’s busiest ever peak🙏
3:11

It was another good Christmas for M&S, building on a strong performance in the prior year. Sales records were broken across the business - with Food recording its biggest day and Clothing, Home & Beauty online its biggest week - but we’re not complacent. As a growth business it’s our job to break records. The external environment remains challenging, with cost and economic headwinds to navigate, but there is much within our control. At M&S, we stay close to our customers and their needs, and with that in mind our investment in trusted value, along with great quality, style and innovation remains our priority. Transforming M&S is a marathon, not a sprint, and we go into 2025 shifting up a gear and raring to go as we accelerate the scale and pace of change. Thank you to all of our customers for shopping with M&S this Christmas and to all of our colleagues for their extraordinary effort delivering M&S’s busiest ever peak🙏

Stuart Machin

15,320 views • 1 year ago

Today is Thank A First Responder Day. Here’s a little throwback to spending the day with Rural Metro Fire in Tucson, Arizona and trying on some of their gear. Today we share our gratitude for first responders and their families. Their hard work and sacrifice keeps us all safe. We are proud to be hosting a great day at Australia Zoo for all first responders with Fortem Australia. And Thanks again to rural metro, they shared knowledge of how they fight fires and keep the community safe, and we shared some tips on the best ways to rescue rattlesnakes out of homes!
0:32

Today is Thank A First Responder Day. Here’s a little throwback to spending the day with Rural Metro Fire in Tucson, Arizona and trying on some of their gear. Today we share our gratitude for first responders and their families. Their hard work and sacrifice keeps us all safe. We are proud to be hosting a great day at Australia Zoo for all first responders with Fortem Australia. And Thanks again to rural metro, they shared knowledge of how they fight fires and keep the community safe, and we shared some tips on the best ways to rescue rattlesnakes out of homes!

Robert Irwin

45,923 views • 1 year ago