Lequn Chen

@tskaerobot @Yuchenj_UW Upload all tax documents. Prompt "prepare my 2025 tax" and your information (like location, single or married, ...). Same as what you would send to CPA. (If you don't know which docs are needed, just ask it)

@abcdabcd987 @Yuchenj_UW Wow. Can you recommend a tutorial. I paid a cpa $2000 and I think he didn’t do a great job f

Anthropic killed this, Anthropic killed that, why cant Anthropic kill TurboTax

@iamup @AravSrinivas I uploaded all tax documents and also equity contracts. Same as what I sent to my CPA previously.

@AravSrinivas Does one have to upload full tax documents showing SSN or just add the income etc and get the tax return prepared? @abcdabcd987

Perplexity Computer is more reliable than a CPA for filing taxes.

Check out my talk at Ray Summit 2025 on RDMA Point-to-Point Communication for LLM Systems youtube.com/watch?v=Nl8iqY…

zhihu: zhuanlan.zhihu.com/p/197123241073…

Wrote a blog post on why collective communication feels awkward for newer LLM workloads (disaggregated inference, RL weight update, MoE), why people don’t just use raw RDMA, how we approached it, and some behind-the-scenes stories. le.qun.ch/en/blog/2025/1…

Faster than DeepEP for Decode on ConnectX-7. First viable kernel on EFA. SM-Free RDMA transfer. Support prefill. (Maybe portable to other hardware as well)

Read more in the blog post! research.perplexity.ai/articles/weigh…

We divide the weight transfer process into pipeline stages to enable overlapped execution over different hardware resources (CPU->GPU memcpy, GPU computation, RDMA, Ethernet).

We recently achieved 1.3-second cross-machine parameter update for Kimi-K2 (1T parameters), as opposed to a few minutes in popular frameworks.

How does @deepseek_ai Sparse Attention (DSA) work? It has 2 components: the Lightning Indexer and Sparse Multi-Latent Attention (MLA). The indexer keeps a small key cache of 128 per token (vs. 512 for MLA). It scores incoming queries. The top-2048 tokens to pass to Sparse MLA.

Introducing Perplexity Search API We've built a search index of billions of webpages to provide real-time, quality information from the web. Now developers have access to the full power of our index, providing the most accurate results in milliseconds. perplexity.ai/hub/blog/intro…

Just got a sneak peek of the breakout sessions lineup for #RaySummit2025 – and it’s 🔥 Sessions from: 🔹 @character_ai on Scaling LLM Post-Training 🔹 The State of @vllm_project in 2025 🔹 @Roblox on Training 3D Foundation Models with Ray 🔹 @xai on Scaling Image + Video Processing 🔹 @zoox on Reliable, Multimodal LLM Serving 🔹 @perplexity_ai on RDMA P2P for KvCache + MoE Looking forward to learning from the teams actually building these systems. Come join us. Save 25% with code ANYJOIN25 →anyscale.com/ray-summit/202…

@LigengZhu Glad that you enjoyed it! To be precise, it's EP64 on the inference side, around 30GB per inference GPU. So it's around 30GB / 1.3s = 23 GB/s.

Every RL infra resesarchers should read's @abcdabcd987 blog, 1T / 1.3s / 16 nodes = 49GB/s. Nearly fully reach the peak of the IB bandwidth! For Kimi-K2 (1T params), with 256 GPUs in BF16 training and 128 GPUs in FP8 inference, weight updates take less than 1.3 seconds.

@LiyuanLucas Glad that you enjoyed it!

This is really impressive! Enjoyed the Chinese blog a lot, didn't know it has an English version as well :)

@vwxyzjn Haha. Glad that you enjoy it :)

What a great blog by Lequn. Love the writing, too "So in my head, shipping weights should be about as simple as putting an elephant into a fridge."

@drisspg Glad that you like it 😀

@abcdabcd987 Great read!

1.5 seconds is long enough to transfer model weights from training nodes to RL rollout nodes (as opposed to 100s). Here's the full story of how I made it (not just presenting the solution): le.qun.ch/en/blog/2025/0…