Tamer

HUJI • AI21 Labs

[Tamer @TamerGhattas911]

By the agents for the agents: web based lidar point cloud viewer based on Rust and WASM, tried it with 4GB files with no issues. Provides http hosted and local MCP server bridges for agent use.

app: https://t.co/AEmBWo9UKD
repo: https://t.co/BtrYchoOeQ

#pointcloud #lidar #E57

[vLLM @vllm_project]

Thanks to @AI21Labs for tracking down a silent uint32 overflow in vLLM's Mamba-1 CUDA kernel and contributing the fix.
Root cause: `uint32_t` stride × cache_index overflows silently at scale. Fix merged in #35275. The debugging story is worth a read.
🔗 https://t.co/S4XBnEn1uv

[AI21 Labs @AI21Labs]

4/4 The fix upstream? Two characters, basically:uint32_t → size_t

Weeks of debugging for one tiny type-width bug. Great reminder that in RL infra, the hardest part is often finding the right boundary around the failure.

👉Blog with the full investigation + the upstream vLLM PR: https://t.co/fDsQICAb1j

[AI21 Labs @AI21Labs]

3/4 Root cause: a silent 32-bit integer overflow inside a vLLM CUDA kernel for Mamba-1 selective scan forward kernel.

cache_index * ssm_states_batch_stride overflowed once cache slots got large enough, corrupting writes with no crash and no warning.

[AI21 Labs @AI21Labs]

2/4 The weird part: the spikes were periodic. When we increased rollouts per prompt, the spike pattern moved with them.

That was the clue that this wasn’t “training instability”, it was a structured rollout-path bug.

[AI21 Labs @AI21Labs]

1/4 We hit a strange logprob mismatch while training Jamba 3B with GRPO.

Rollout logprobs and training-side recompute should match before any weight update. Ours didn’t.

That was the canary. 🧵

[Bo Wang @BoWang87]

Prof. Donald Knuth opened his new paper with "Shock! Shock!"

Claude Opus 4.6 had just solved an open problem he'd been working on for weeks — a graph decomposition conjecture from The Art of Computer Programming.

He named the paper "Claude's Cycles."

31 explorations. ~1 hour. Knuth read the output, wrote the formal proof, and closed with: "It seems I'll have to revise my opinions about generative AI one of these days."

The man who wrote the bible of computer science just said that. In a paper named after an AI.

Paper: https://t.co/juSOmK9vOt

[Ethan Mollick @emollick]

GPT-5, Claude, Kimi, and Gemini: "I can travel back in time to any time before 1500 and change only one thing, what is the single thing you would change, nothing obvious."

[Tamer @TamerGhattas911]

#EMNLP2025

[Tamer @TamerGhattas911]

Excited to share: our paper “On Pruning State-Space LLMs” was accepted to EMNLP 2025! 🎉

Preprint: https://t.co/8TD56aroDc
Code: https://t.co/ofi9ZxPDOT
Model: Smol-Mamba-1.9B → https://t.co/AIq2XOn0dS
w/ @MichaelHassid & @royschwartzNLP (HUJI)
#Mamba #ModelCompression

[Tamer @TamerGhattas911]

🚀 New Paper Drop! 🚀

“On Pruning SSM LLMs” – We check the prunability of MAMBA🐍 based LLMs.

We also release Smol2-Mamba-1.9B, a MAMBA based LLM distilled from Smol2-1.7B on 🤗: [https://t.co/AIq2XOny3q]

📖 Read more: [https://t.co/8TD56arWsK]

@royschwartzNLP @MichaelHassid