Untied Ulysses: Memory-Efficient Context Parallelism via Headwise Chunking
Paper โข 2602.21196 โข Published โข 7
AI & ML interests
None defined yet.
mryabย
submitted a
paper to Daily Papers 8 months ago
Post
8,022
๐ What happened in AI in 2025? ๐
We prepared the 2025 version of the HF AI Timeline Grid, highlighting open vs API-based model releases, and allowing you to browse and filter by access, modality, and release type!
Play with it here:
2025-ai-timeline/2025-ai-timeline
Here's my personal quarterly TL;DR:
1๏ธโฃ Q1 โ Learning to Reason
Deepseek not only releases a top-notch reasoning model, but shows how to train them and compete with closed frontier models. OpenAI debuts Deep Research.
Significant milestones: DeepSeek R1 & R1-Zero, Qwen 2.5 VL, OpenAI Deep Research, Gemini 2.5 Pro (experimental)
2๏ธโฃ Q2 โ Multimodality and Coding
More LLMs embrace multimodality by default, and there's a surge in coding agents. Strong vision, audio, and generative models emerge.
Significant milestones: Llama 4, Qwen 3, Imagen 4, OpenAI Codex, Google Jules, Claude 4
3๏ธโฃ Q3 โ "Gold" rush, OpenAI opens up, the community goes bananas
Flagship models get gold in Math olympiads and hard benchmarks. OpenAI releases strong open source models and Google releases the much anticipated nano-banana for image generation and editing. Agentic workflows become commonplace.
Significant milestones: Gemini and OpenAI IMO Gold, gpt-oss, Gemini 2.5 Flash Image, Grok 4, Claude Sonnet 4.5
4๏ธโฃ Q4 โ Mistral returns, leaderboard hill-climbing
Mistral is back with updated model families. All labs release impressive models to wrap up the year!
Significant milestones: Claude Opus 4.5, DeepSeek Math V2, FLUX 2, GPT 5.1, Kimi K2 Thinking, Nano Banana Pro, GLM 4.7, Gemini 3, Mistral 3, MiniMax M2.1 ๐คฏ
Credits
๐ NHLOCAL for the source data https://github.com/NHLOCAL/AiTimeline
๐ซก @reach-vb for the original idea, design and recipe
๐ @ariG23498 and yours truly for compiling and verifying the 2025 edition
๐ฅณ Here's to 2026, wishing it becomes the best year ever for open releases and on-device-first use-cases! ๐ฅ
We prepared the 2025 version of the HF AI Timeline Grid, highlighting open vs API-based model releases, and allowing you to browse and filter by access, modality, and release type!
Play with it here:
2025-ai-timeline/2025-ai-timeline
Here's my personal quarterly TL;DR:
1๏ธโฃ Q1 โ Learning to Reason
Deepseek not only releases a top-notch reasoning model, but shows how to train them and compete with closed frontier models. OpenAI debuts Deep Research.
Significant milestones: DeepSeek R1 & R1-Zero, Qwen 2.5 VL, OpenAI Deep Research, Gemini 2.5 Pro (experimental)
2๏ธโฃ Q2 โ Multimodality and Coding
More LLMs embrace multimodality by default, and there's a surge in coding agents. Strong vision, audio, and generative models emerge.
Significant milestones: Llama 4, Qwen 3, Imagen 4, OpenAI Codex, Google Jules, Claude 4
3๏ธโฃ Q3 โ "Gold" rush, OpenAI opens up, the community goes bananas
Flagship models get gold in Math olympiads and hard benchmarks. OpenAI releases strong open source models and Google releases the much anticipated nano-banana for image generation and editing. Agentic workflows become commonplace.
Significant milestones: Gemini and OpenAI IMO Gold, gpt-oss, Gemini 2.5 Flash Image, Grok 4, Claude Sonnet 4.5
4๏ธโฃ Q4 โ Mistral returns, leaderboard hill-climbing
Mistral is back with updated model families. All labs release impressive models to wrap up the year!
Significant milestones: Claude Opus 4.5, DeepSeek Math V2, FLUX 2, GPT 5.1, Kimi K2 Thinking, Nano Banana Pro, GLM 4.7, Gemini 3, Mistral 3, MiniMax M2.1 ๐คฏ
Credits
๐ NHLOCAL for the source data https://github.com/NHLOCAL/AiTimeline
๐ซก @reach-vb for the original idea, design and recipe
๐ @ariG23498 and yours truly for compiling and verifying the 2025 edition
๐ฅณ Here's to 2026, wishing it becomes the best year ever for open releases and on-device-first use-cases! ๐ฅ
- 3 replies
alozowskiย
authored a
paper 11 months ago
nouamanetaziย
posted an update 12 months ago
Post
5,300
After training ๐๐ฆ๐จ๐ฅ๐๐๐ on ๐๐๐ ๐๐๐๐๐ฌ for nearly a month, I've come to realize something most people overlook: ๐ข๐ง๐๐ซ๐๐ฌ๐ญ๐ซ๐ฎ๐๐ญ๐ฎ๐ซ๐ ๐ข๐ฌ ๐ญ๐ก๐ ๐ฆ๐๐ค๐-๐จ๐ซ-๐๐ซ๐๐๐ค ๐๐๐๐ญ๐จ๐ซ ๐ข๐ง ๐๐๐ ๐ญ๐ซ๐๐ข๐ง๐ข๐ง๐ . ๐ฅ
Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious ๐๐๐๐ ๐๐ซ๐ซ๐จ๐ซ๐ฌ, or when your expensive GPU cluster is running at ๐๐% ๐๐๐๐ข๐๐ข๐๐ง๐๐ฒ, the problem isn't your model. It's most probably a ๐ฆ๐ข๐ฌ๐ฎ๐ฌ๐ ๐จ๐ ๐ญ๐ก๐ ๐ก๐๐ซ๐๐ฐ๐๐ซ๐. ๐ ๏ธ
Questions that seemed simple but had no clear answers: Why is ๐๐จ๐ ๐ญ๐ซ๐๐ข๐ง๐ข๐ง๐ ๐ฌ๐ฅ๐จ๐ฐ๐๐ซ ๐ญ๐ก๐๐ง ๐๐๐ง๐ฌ๐ ๐ฆ๐จ๐๐๐ฅ๐ฌ? Which ๐๐๐๐ ๐๐ฅ๐๐ ๐ฌ should we actually set? How often should we checkpoint without killing throughput?
That's why we built ๐๐ก๐ ๐๐ฆ๐จ๐ฅ ๐๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ฅ๐๐ฒ๐๐จ๐จ๐ค ๐: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the ๐ข๐ง๐๐ซ๐๐ฌ๐ญ๐ซ๐ฎ๐๐ญ๐ฎ๐ซ๐ ๐ฅ๐๐ฒ๐๐ซ that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: ๐๐๐๐ ๐ก๐ข๐ญ๐ญ๐ข๐ง๐ ๐ ๐๐/๐ฌ, ๐๐๐๐ข๐ง๐ค ๐.๐ ๐ซ๐๐๐๐ก๐ข๐ง๐ ๐๐๐ ๐๐/๐ฌ, ๐๐๐๐ ๐๐๐ง๐ ๐๐ญ ๐๐.๐ ๐๐/๐ฌ. Then we ran collective operations across ๐๐๐ ๐๐๐๐ฌ (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from ๐๐๐ ๐๐/๐ฌ on a single node to ๐๐๐-๐๐๐ ๐๐/๐ฌ across 16 nodes.
If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging.
๐๐ก๐ ๐๐ฆ๐จ๐ฅ ๐๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ฅ๐๐ฒ๐๐จ๐จ๐ค: https://lnkd.in/e5MKXUHS
Shared with โค๏ธ by the HuggingFace team
Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious ๐๐๐๐ ๐๐ซ๐ซ๐จ๐ซ๐ฌ, or when your expensive GPU cluster is running at ๐๐% ๐๐๐๐ข๐๐ข๐๐ง๐๐ฒ, the problem isn't your model. It's most probably a ๐ฆ๐ข๐ฌ๐ฎ๐ฌ๐ ๐จ๐ ๐ญ๐ก๐ ๐ก๐๐ซ๐๐ฐ๐๐ซ๐. ๐ ๏ธ
Questions that seemed simple but had no clear answers: Why is ๐๐จ๐ ๐ญ๐ซ๐๐ข๐ง๐ข๐ง๐ ๐ฌ๐ฅ๐จ๐ฐ๐๐ซ ๐ญ๐ก๐๐ง ๐๐๐ง๐ฌ๐ ๐ฆ๐จ๐๐๐ฅ๐ฌ? Which ๐๐๐๐ ๐๐ฅ๐๐ ๐ฌ should we actually set? How often should we checkpoint without killing throughput?
That's why we built ๐๐ก๐ ๐๐ฆ๐จ๐ฅ ๐๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ฅ๐๐ฒ๐๐จ๐จ๐ค ๐: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the ๐ข๐ง๐๐ซ๐๐ฌ๐ญ๐ซ๐ฎ๐๐ญ๐ฎ๐ซ๐ ๐ฅ๐๐ฒ๐๐ซ that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: ๐๐๐๐ ๐ก๐ข๐ญ๐ญ๐ข๐ง๐ ๐ ๐๐/๐ฌ, ๐๐๐๐ข๐ง๐ค ๐.๐ ๐ซ๐๐๐๐ก๐ข๐ง๐ ๐๐๐ ๐๐/๐ฌ, ๐๐๐๐ ๐๐๐ง๐ ๐๐ญ ๐๐.๐ ๐๐/๐ฌ. Then we ran collective operations across ๐๐๐ ๐๐๐๐ฌ (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from ๐๐๐ ๐๐/๐ฌ on a single node to ๐๐๐-๐๐๐ ๐๐/๐ฌ across 16 nodes.
If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging.
๐๐ก๐ ๐๐ฆ๐จ๐ฅ ๐๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ฅ๐๐ฒ๐๐จ๐จ๐ค: https://lnkd.in/e5MKXUHS
Shared with โค๏ธ by the HuggingFace team
ariG23498ย
authored a
paper about 1 year ago
Post
16,559
deepseek-ai/DeepSeek-OCR is out! ๐ฅ my take โคต๏ธ
> pretty insane it can parse and re-render charts in HTML
> it uses CLIP and SAM features concatenated, so better grounding
> very efficient per vision tokens/performance ratio
> covers 100 languages
> pretty insane it can parse and re-render charts in HTML
> it uses CLIP and SAM features concatenated, so better grounding
> very efficient per vision tokens/performance ratio
> covers 100 languages
- 5 replies
Post
4,982
๐ New blog: Maintain the unmaintainable โ 1M+ Python LOC, 400+ models
How do you stop a million-line library built by thousands of contributors from collapsing under its own weight?
At ๐ค Transformers, we do it with explicit software-engineering tenets, principles that make the codebase hackable at scale.
๐ Inside the post:
โ One Model, One File: readability first โ you can still open a modeling file and see the full logic, top to bottom.
โ Modular Transformers: visible inheritance that cuts maintenance cost by ~15ร while keeping models readable.
โ Config-Driven Performance: FlashAttention, tensor parallelism, and attention scheduling are config-level features, not rewrites.
Written with @lysandre ,@pcuenq and @yonigozlan , this is a deep dive into how Transformers stays fast, open, and maintainable.
Read it here โ transformers-community/Transformers-tenets
How do you stop a million-line library built by thousands of contributors from collapsing under its own weight?
At ๐ค Transformers, we do it with explicit software-engineering tenets, principles that make the codebase hackable at scale.
๐ Inside the post:
โ One Model, One File: readability first โ you can still open a modeling file and see the full logic, top to bottom.
โ Modular Transformers: visible inheritance that cuts maintenance cost by ~15ร while keeping models readable.
โ Config-Driven Performance: FlashAttention, tensor parallelism, and attention scheduling are config-level features, not rewrites.
Written with @lysandre ,@pcuenq and @yonigozlan , this is a deep dive into how Transformers stays fast, open, and maintainable.
Read it here โ transformers-community/Transformers-tenets
Post
7,363
large AI labs open-sourced a ton of models last week ๐ฅ
here's few picks, find even more here merve/sep-16-releases-68d13ea4c547f02f95842f05 ๐ค
> IBM released a new Docling model with 258M params based on Granite (A2.0) ๐ ibm-granite/granite-docling-258M
> Xiaomi released 7B audio LM with base and instruct variants (MIT) XiaomiMiMo/mimo-audio-68cc7202692c27dae881cce0
> DecartAI released Lucy Edit, open Nano Banana ๐ (NC) decart-ai/Lucy-Edit-Dev
> OpenGVLab released a family of agentic computer use models (3B/7B/32B) with the dataset ๐ป OpenGVLab/scalecua-68c912cf56f7ff4c8e034003
> Meituan Longcat released thinking version of LongCat-Flash ๐ญ meituan-longcat/LongCat-Flash-Thinking
here's few picks, find even more here merve/sep-16-releases-68d13ea4c547f02f95842f05 ๐ค
> IBM released a new Docling model with 258M params based on Granite (A2.0) ๐ ibm-granite/granite-docling-258M
> Xiaomi released 7B audio LM with base and instruct variants (MIT) XiaomiMiMo/mimo-audio-68cc7202692c27dae881cce0
> DecartAI released Lucy Edit, open Nano Banana ๐ (NC) decart-ai/Lucy-Edit-Dev
> OpenGVLab released a family of agentic computer use models (3B/7B/32B) with the dataset ๐ป OpenGVLab/scalecua-68c912cf56f7ff4c8e034003
> Meituan Longcat released thinking version of LongCat-Flash ๐ญ meituan-longcat/LongCat-Flash-Thinking
- 2 replies
Post
3,785
IBM just released small swiss army knife for the document models: granite-docling-258M on Hugging Face ๐ฅ
> not only a document converter but also can do document question answering, understand multiple languages ๐คฏ
> best part: released with Apache 2.0 license ๐ use it with your commercial projects!
> it supports transformers, vLLM and MLX from the get-go! ๐ค
> built on SigLIP2 & granite-165M
model: ibm-granite/granite-docling-258M
demo: ibm-granite/granite-docling-258m-demo ๐
> not only a document converter but also can do document question answering, understand multiple languages ๐คฏ
> best part: released with Apache 2.0 license ๐ use it with your commercial projects!
> it supports transformers, vLLM and MLX from the get-go! ๐ค
> built on SigLIP2 & granite-165M
model: ibm-granite/granite-docling-258M
demo: ibm-granite/granite-docling-258m-demo ๐
Post
10,540
We're kick-starting the process of Transformers v5, with @ArthurZ and @cyrilvallez !
v5 should be significant: we're using it as a milestone for performance optimizations, saner defaults, and a much cleaner code base worthy of 2025.
Fun fact: v4.0.0-rc-1 came out on Nov 19, 2020, nearly five years ago!
v5 should be significant: we're using it as a milestone for performance optimizations, saner defaults, and a much cleaner code base worthy of 2025.
Fun fact: v4.0.0-rc-1 came out on Nov 19, 2020, nearly five years ago!
- 6 replies
Post
1,479
a ton of image/video generation models and LLMs from big labs ๐ฅ
> Meta released facebook/mobilellm-r1-68c4597b104fac45f28f448e, smol LLMs for on-device use ๐ฌ
> Tencent released tencent/SRPO, high res image generation model and tencent/POINTS-Reader, cutting edge OCR ๐
> ByteDance released bytedance-research/HuMo, video generation from any input โฏ๏ธ
find more models, datasets, demos here merve/sep-11-releases-68c7dbfa26bea8cd921fa0ac
> Meta released facebook/mobilellm-r1-68c4597b104fac45f28f448e, smol LLMs for on-device use ๐ฌ
> Tencent released tencent/SRPO, high res image generation model and tencent/POINTS-Reader, cutting edge OCR ๐
> ByteDance released bytedance-research/HuMo, video generation from any input โฏ๏ธ
find more models, datasets, demos here merve/sep-11-releases-68c7dbfa26bea8cd921fa0ac
Post
1,235
fan-favorite vision LM Florence-2 is now officially supported in transformers ๐ค
find all the models in
florence-community org ๐ซก
find all the models in
Post
3,376
New post is live!
This time we cover some major updates to
๐ค
This time we cover some major updates to
transformers.๐ค
- 2 replies
Post
1,998
past week was great for open LLMs ๐ฅ merve/sep-1-releases-68bede0e729c12597eefd050
> Google released google/embeddinggemma-300m, new embedding model with 300M params
> new update to Kimi-K2 just landed moonshotai/Kimi-K2-Instruct-0905 ๐
> OpenBMB released a new version to MiniCPM with 8B params openbmb/MiniCPM4.1-8B
also soooo many Qwen-Image & Kontext LoRAs dropped!
> Google released google/embeddinggemma-300m, new embedding model with 300M params
> new update to Kimi-K2 just landed moonshotai/Kimi-K2-Instruct-0905 ๐
> OpenBMB released a new version to MiniCPM with 8B params openbmb/MiniCPM4.1-8B
also soooo many Qwen-Image & Kontext LoRAs dropped!
Post
3,867
upgrade your transformers ๐ฅ
it comes with insanely capable models like merve/sam2-66ac9deac6fca3bc5482fe30, microsoft/kosmos-2.5, and more ๐ซก
I built a notebook you can run with free Colab T4 to walk through the API for new models ๐๐ปโโ๏ธ merve/smol-vision
fine-tuning will follow-up soon!
it comes with insanely capable models like merve/sam2-66ac9deac6fca3bc5482fe30, microsoft/kosmos-2.5, and more ๐ซก
I built a notebook you can run with free Colab T4 to walk through the API for new models ๐๐ปโโ๏ธ merve/smol-vision
fine-tuning will follow-up soon!
Post
6,414
large AI labs have dropped so many open models last week ๐ฅ don't miss out on them
โ Apple released on-device vision LMs apple/fastvlm-68ac97b9cd5cacefdd04872e & apple/mobileclip2-68ac947dcb035c54bcd20c47
โ OpenGVLab released InternVL3.5, 32 new vision LMs with one based on gpt-oss! (OS) OpenGVLab/internvl35-68ac87bd52ebe953485927fb
โ MSFT released a killer small TTS model (OS) microsoft/VibeVoice-1.5B
find more herehttps://huggingface.co/collections/merve/august-29-releases-68b5a3754cfb8abf59e2b486
โ Apple released on-device vision LMs apple/fastvlm-68ac97b9cd5cacefdd04872e & apple/mobileclip2-68ac947dcb035c54bcd20c47
โ OpenGVLab released InternVL3.5, 32 new vision LMs with one based on gpt-oss! (OS) OpenGVLab/internvl35-68ac87bd52ebe953485927fb
โ MSFT released a killer small TTS model (OS) microsoft/VibeVoice-1.5B
find more herehttps://huggingface.co/collections/merve/august-29-releases-68b5a3754cfb8abf59e2b486
- 1 reply
Post
6,185
first vision language model built off openai/gpt-oss-20b just dropped! ๐ฅ
InternVL3.5 comes with 32 models ๐คฏ pre-trained, fine-tuned, aligned in various sizes OpenGVLab/internvl35-68ac87bd52ebe953485927fb
comes with gpt-oss or Qwen3 for LLM part โคต๏ธ
InternVL3.5 comes with 32 models ๐คฏ pre-trained, fine-tuned, aligned in various sizes OpenGVLab/internvl35-68ac87bd52ebe953485927fb
comes with gpt-oss or Qwen3 for LLM part โคต๏ธ
- 1 reply
Post
3,400
GPT-4.1-mini level model right in your iPhone ๐คฏ
openbmb/MiniCPM-V-4 is only 4B while surpassing GPT-4.1-mini in vision benchmarks ๐ฅ
allows commercial use as well!
openbmb/MiniCPM-V-4 is only 4B while surpassing GPT-4.1-mini in vision benchmarks ๐ฅ
allows commercial use as well!
Post
1,250
we're all sleeping on this OCR model https://huggingface.co/rednote-hilab/dots.ocr ๐ฅ
dots.ocr is a new 3B model with sota performance, support for 100 languages & allowing commercial use! ๐คฏ
single e2e model to extract image, convert tables, formula, and more into markdown ๐
try it MohamedRashad/Dots-OCR
dots.ocr is a new 3B model with sota performance, support for 100 languages & allowing commercial use! ๐คฏ
single e2e model to extract image, convert tables, formula, and more into markdown ๐
try it MohamedRashad/Dots-OCR
Post
757
massive releases and tons of Flux 1. Krea LoRas past week!
here's some of the picks, find more models in collection ๐ซก merve/releases-august-2-6890c14248203522b7d0267f
LLMs ๐ฌ
> Tencent dropped tencent/Hunyuan-7B-Instruct
> Qwen released Qwen/Qwen3-Coder-30B-A3B-Instruct, 30B MoE with 3B params for coding (OS)
vision/multimodal
> RedNote released https://huggingface.co/rednote-hilab/dots.ocr - 3B OCR model (OS)
> Cohere released CohereLabs/command-a-vision-07-2025 - 112B (dense!) VLM for 6 languages
> StepFun-AI shipped stepfun-ai/step3 - 321B MoE VLM (OS)
> Skywork shipped Skywork/Skywork-UniPic-1.5B - new any-to-any model (image+text โ image+text) (OS)
here's some of the picks, find more models in collection ๐ซก merve/releases-august-2-6890c14248203522b7d0267f
LLMs ๐ฌ
> Tencent dropped tencent/Hunyuan-7B-Instruct
> Qwen released Qwen/Qwen3-Coder-30B-A3B-Instruct, 30B MoE with 3B params for coding (OS)
vision/multimodal
> RedNote released https://huggingface.co/rednote-hilab/dots.ocr - 3B OCR model (OS)
> Cohere released CohereLabs/command-a-vision-07-2025 - 112B (dense!) VLM for 6 languages
> StepFun-AI shipped stepfun-ai/step3 - 321B MoE VLM (OS)
> Skywork shipped Skywork/Skywork-UniPic-1.5B - new any-to-any model (image+text โ image+text) (OS)