SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model
Paper
•
2502.02737
•
Published
•
153
private beta for deeplinks
non-profit
AI & ML interests
None defined yet.
Recent Activity
View all activity
pbdeeplinks's activity
reach-vb
authored
a
paper
5 days ago
Post
33714
Switching to
I posted a poll on twitter, and others have mentioned the interest in me using the convention of including the author name in the model path when I upload.
It has a couple advantages, first and foremost of course is ensuring clarity of who uploaded the original model (did Qwen upload Qwen2.6? Or did someone fine tune Qwen2.5 and named it 2.6 for fun?)
The second thing is that it avoids collisions, so if multiple people upload the same model and I try to quant them both, I would normally end up colliding and being unable to upload both
I'll be implementing the change next week, there are just two final details I'm unsure about:
First, should the files also inherit the author's name?
Second, what to do in the case that the author name + model name pushes us past the character limit?
Haven't yet decided how to handle either case, so feedback is welcome, but also just providing this as a "heads up"
author_model-nameI posted a poll on twitter, and others have mentioned the interest in me using the convention of including the author name in the model path when I upload.
It has a couple advantages, first and foremost of course is ensuring clarity of who uploaded the original model (did Qwen upload Qwen2.6? Or did someone fine tune Qwen2.5 and named it 2.6 for fun?)
The second thing is that it avoids collisions, so if multiple people upload the same model and I try to quant them both, I would normally end up colliding and being unable to upload both
I'll be implementing the change next week, there are just two final details I'm unsure about:
First, should the files also inherit the author's name?
Second, what to do in the case that the author name + model name pushes us past the character limit?
Haven't yet decided how to handle either case, so feedback is welcome, but also just providing this as a "heads up"
- 3 replies
Post
9462
After some heated discussion 🔥, we clarify our intent re. storage limits on the Hub
TL;DR:
- public storage is free, and (unless blatant abuse) unlimited. We do ask that you consider upgrading to PRO and/or Enterprise Hub if possible
- private storage is paid above a significant free tier (1TB if you have a paid account, 100GB otherwise)
docs: https://huggingface.co/docs/hub/storage-limits
We optimize our infrastructure continuously to scale our storage for the coming years of growth in Machine learning, to the benefit of the community 🔥
cc: @reach-vb @pierric @victor and the HF team
TL;DR:
- public storage is free, and (unless blatant abuse) unlimited. We do ask that you consider upgrading to PRO and/or Enterprise Hub if possible
- private storage is paid above a significant free tier (1TB if you have a paid account, 100GB otherwise)
docs: https://huggingface.co/docs/hub/storage-limits
We optimize our infrastructure continuously to scale our storage for the coming years of growth in Machine learning, to the benefit of the community 🔥
cc: @reach-vb @pierric @victor and the HF team
-
-
- 28 replies
Post
44697
Looks like Q4_0_N_M file types are going away
Before you panic, there's a new "preferred" method which is online (I prefer the term on-the-fly) repacking, so if you download Q4_0 and your setup can benefit from repacking the weights into interleaved rows (what Q4_0_4_4 was doing), it will do that automatically and give you similar performance (minor losses I think due to using intrinsics instead of assembly, but intrinsics are more maintainable)
You can see the reference PR here:
https://github.com/ggerganov/llama.cpp/pull/10446
So if you update your llama.cpp past that point, you won't be able to run Q4_0_4_4 (unless they add backwards compatibility back), but Q4_0 should be the same speeds (though it may currently be bugged on some platforms)
As such, I'll stop making those newer model formats soon, probably end of this week unless something changes, but you should be safe to download and Q4_0 quants and use those !
Also IQ4_NL supports repacking though not in as many shapes yet, but should get a respectable speed up on ARM chips, PR for that can be found here: https://github.com/ggerganov/llama.cpp/pull/10541
Remember, these are not meant for Apple silicon since those use the GPU and don't benefit from the repacking of weights
Before you panic, there's a new "preferred" method which is online (I prefer the term on-the-fly) repacking, so if you download Q4_0 and your setup can benefit from repacking the weights into interleaved rows (what Q4_0_4_4 was doing), it will do that automatically and give you similar performance (minor losses I think due to using intrinsics instead of assembly, but intrinsics are more maintainable)
You can see the reference PR here:
https://github.com/ggerganov/llama.cpp/pull/10446
So if you update your llama.cpp past that point, you won't be able to run Q4_0_4_4 (unless they add backwards compatibility back), but Q4_0 should be the same speeds (though it may currently be bugged on some platforms)
As such, I'll stop making those newer model formats soon, probably end of this week unless something changes, but you should be safe to download and Q4_0 quants and use those !
Also IQ4_NL supports repacking though not in as many shapes yet, but should get a respectable speed up on ARM chips, PR for that can be found here: https://github.com/ggerganov/llama.cpp/pull/10541
Remember, these are not meant for Apple silicon since those use the GPU and don't benefit from the repacking of weights
-
- 16 replies
Post
4876
VLMs are going through quite an open revolution AND on-device friendly sizes:
1. Google DeepMind w/ PaliGemma2 - 3B, 10B & 28B: google/paligemma-2-release-67500e1e1dbfdd4dee27ba48
2. OpenGVLabs w/ InternVL 2.5 - 1B, 2B, 4B, 8B, 26B, 38B & 78B: https://huggingface.co/collections/OpenGVLab/internvl-25-673e1019b66e2218f68d7c1c
3. Qwen w/ Qwen 2 VL - 2B, 7B & 72B: Qwen/qwen2-vl-66cee7455501d7126940800d
4. Microsoft w/ FlorenceVL - 3B & 8B: https://huggingface.co/jiuhai
5. Moondream2 w/ 0.5B: https://huggingface.co/vikhyatk/
What a time to be alive! 🔥
1. Google DeepMind w/ PaliGemma2 - 3B, 10B & 28B: google/paligemma-2-release-67500e1e1dbfdd4dee27ba48
2. OpenGVLabs w/ InternVL 2.5 - 1B, 2B, 4B, 8B, 26B, 38B & 78B: https://huggingface.co/collections/OpenGVLab/internvl-25-673e1019b66e2218f68d7c1c
3. Qwen w/ Qwen 2 VL - 2B, 7B & 72B: Qwen/qwen2-vl-66cee7455501d7126940800d
4. Microsoft w/ FlorenceVL - 3B & 8B: https://huggingface.co/jiuhai
5. Moondream2 w/ 0.5B: https://huggingface.co/vikhyatk/
What a time to be alive! 🔥
Post
16249
Old mixtral model quants may be broken!
Recently Slaren over on llama.cpp refactored the model loader - in a way that's super awesome and very powerful - but with it came breaking of support for "split tensor MoE models", which applies to older mixtral models
You may have seen my upload of one such older mixtral model, ondurbin/bagel-dpo-8x7b-v0.2, and with the newest changes it seems to be able to run without issue
If you happen to run into issues with any other old mixtral models, drop a link here and I'll try to remake them with the new changes so that we can continue enjoying them :)
Recently Slaren over on llama.cpp refactored the model loader - in a way that's super awesome and very powerful - but with it came breaking of support for "split tensor MoE models", which applies to older mixtral models
You may have seen my upload of one such older mixtral model, ondurbin/bagel-dpo-8x7b-v0.2, and with the newest changes it seems to be able to run without issue
If you happen to run into issues with any other old mixtral models, drop a link here and I'll try to remake them with the new changes so that we can continue enjoying them :)
-
- 2 replies
Post
3170
wow 😮
INTELLECT-1 is the first collaboratively trained 10 billion parameter language model trained from scratch on 1 trillion tokens of English text and code.
PrimeIntellect/INTELLECT-1-Instruct
INTELLECT-1 is the first collaboratively trained 10 billion parameter language model trained from scratch on 1 trillion tokens of English text and code.
PrimeIntellect/INTELLECT-1-Instruct
Post
4595
Massive week for Open AI/ ML:
Mistral Pixtral & Instruct Large - ~123B, 128K context, multilingual, json + function calling & open weights
mistralai/Pixtral-Large-Instruct-2411
mistralai/Mistral-Large-Instruct-2411
Allen AI Tülu 70B & 8B - competive with claude 3.5 haiku, beats all major open models like llama 3.1 70B, qwen 2.5 and nemotron
allenai/tulu-3-models-673b8e0dc3512e30e7dc54f5
allenai/tulu-3-datasets-673b8df14442393f7213f372
Llava o1 - vlm capable of spontaneous, systematic reasoning, similar to GPT-o1, 11B model outperforms gemini-1.5-pro, gpt-4o-mini, and llama-3.2-90B-vision
Xkev/Llama-3.2V-11B-cot
Black Forest Labs Flux.1 tools - four new state of the art model checkpoints & 2 adapters for fill, depth, canny & redux, open weights
reach-vb/black-forest-labs-flux1-6743847bde9997dd26609817
Jina AI Jina CLIP v2 - general purpose multilingual and multimodal (text & image) embedding model, 900M params, 512 x 512 resolution, matroyoshka representations (1024 to 64)
jinaai/jina-clip-v2
Apple AIM v2 & CoreML MobileCLIP - large scale vision encoders outperform CLIP and SigLIP. CoreML optimised MobileCLIP models
apple/aimv2-6720fe1558d94c7805f7688c
apple/coreml-mobileclip
A lot more got released like, OpenScholar (https://huggingface.co/collections/OpenScholar/openscholar-v1-67376a89f6a80f448da411a6), smoltalk ( HuggingFaceTB/smoltalk), Hymba ( nvidia/hymba-673c35516c12c4b98b5e845f), Open ASR Leaderboard ( hf-audio/open_asr_leaderboard) and much more..
Can't wait for the next week! 🤗
Mistral Pixtral & Instruct Large - ~123B, 128K context, multilingual, json + function calling & open weights
mistralai/Pixtral-Large-Instruct-2411
mistralai/Mistral-Large-Instruct-2411
Allen AI Tülu 70B & 8B - competive with claude 3.5 haiku, beats all major open models like llama 3.1 70B, qwen 2.5 and nemotron
allenai/tulu-3-models-673b8e0dc3512e30e7dc54f5
allenai/tulu-3-datasets-673b8df14442393f7213f372
Llava o1 - vlm capable of spontaneous, systematic reasoning, similar to GPT-o1, 11B model outperforms gemini-1.5-pro, gpt-4o-mini, and llama-3.2-90B-vision
Xkev/Llama-3.2V-11B-cot
Black Forest Labs Flux.1 tools - four new state of the art model checkpoints & 2 adapters for fill, depth, canny & redux, open weights
reach-vb/black-forest-labs-flux1-6743847bde9997dd26609817
Jina AI Jina CLIP v2 - general purpose multilingual and multimodal (text & image) embedding model, 900M params, 512 x 512 resolution, matroyoshka representations (1024 to 64)
jinaai/jina-clip-v2
Apple AIM v2 & CoreML MobileCLIP - large scale vision encoders outperform CLIP and SigLIP. CoreML optimised MobileCLIP models
apple/aimv2-6720fe1558d94c7805f7688c
apple/coreml-mobileclip
A lot more got released like, OpenScholar (https://huggingface.co/collections/OpenScholar/openscholar-v1-67376a89f6a80f448da411a6), smoltalk ( HuggingFaceTB/smoltalk), Hymba ( nvidia/hymba-673c35516c12c4b98b5e845f), Open ASR Leaderboard ( hf-audio/open_asr_leaderboard) and much more..
Can't wait for the next week! 🤗
Post
4441
What a brilliant week for Open Source AI!
Qwen 2.5 Coder by Alibaba - 0.5B / 1.5B / 3B / 7B / 14B/ 32B (Base + Instruct) Code generation LLMs, with 32B tackling giants like Gemnini 1.5 Pro, Claude Sonnet
Qwen/qwen25-coder-66eaa22e6f99801bf65b0c2f
LLM2CLIP from Microsoft - Leverage LLMs to train ultra-powerful CLIP models! Boosts performance over the previous SOTA by ~17%
microsoft/llm2clip-672323a266173cfa40b32d4c
Athene v2 Chat & Agent by NexusFlow - SoTA general LLM fine-tuned from Qwen 2.5 72B excels at Chat + Function Calling/ JSON/ Agents
Nexusflow/athene-v2-6735b85e505981a794fb02cc
Orca Agent Instruct by Microsoft - 1 million instruct pairs covering text editing, creative writing, coding, reading comprehension, etc - permissively licensed
microsoft/orca-agentinstruct-1M-v1
Ultravox by FixieAI - 70B/ 8B model approaching GPT4o level, pick any LLM, train an adapter with Whisper as Audio Encoder
reach-vb/ultravox-audio-language-model-release-67373b602af0a52b2a88ae71
JanusFlow 1.3 by DeepSeek - Next iteration of their Unified MultiModal LLM Janus with RectifiedFlow
deepseek-ai/JanusFlow-1.3B
Common Corpus by Pleais - 2,003,039,184,047 multilingual, commercially permissive and high quality tokens!
PleIAs/common_corpus
I'm sure I missed a lot, can't wait for the next week!
Put down in comments what I missed! 🤗
Qwen 2.5 Coder by Alibaba - 0.5B / 1.5B / 3B / 7B / 14B/ 32B (Base + Instruct) Code generation LLMs, with 32B tackling giants like Gemnini 1.5 Pro, Claude Sonnet
Qwen/qwen25-coder-66eaa22e6f99801bf65b0c2f
LLM2CLIP from Microsoft - Leverage LLMs to train ultra-powerful CLIP models! Boosts performance over the previous SOTA by ~17%
microsoft/llm2clip-672323a266173cfa40b32d4c
Athene v2 Chat & Agent by NexusFlow - SoTA general LLM fine-tuned from Qwen 2.5 72B excels at Chat + Function Calling/ JSON/ Agents
Nexusflow/athene-v2-6735b85e505981a794fb02cc
Orca Agent Instruct by Microsoft - 1 million instruct pairs covering text editing, creative writing, coding, reading comprehension, etc - permissively licensed
microsoft/orca-agentinstruct-1M-v1
Ultravox by FixieAI - 70B/ 8B model approaching GPT4o level, pick any LLM, train an adapter with Whisper as Audio Encoder
reach-vb/ultravox-audio-language-model-release-67373b602af0a52b2a88ae71
JanusFlow 1.3 by DeepSeek - Next iteration of their Unified MultiModal LLM Janus with RectifiedFlow
deepseek-ai/JanusFlow-1.3B
Common Corpus by Pleais - 2,003,039,184,047 multilingual, commercially permissive and high quality tokens!
PleIAs/common_corpus
I'm sure I missed a lot, can't wait for the next week!
Put down in comments what I missed! 🤗
Post
3230
🫵👾 LM Studio is hiring engineers who know the ins and outs of NodeJS & want to work on local LLMs.
Know anyone who might be interested?
Apply here: https://docs.google.com/forms/d/e/1FAIpQLSc_786-_i_q4fo5ESqYnNyjIH0B5Rs45QIwejd_NV5AjNDZ7A/viewform
Know anyone who might be interested?
Apply here: https://docs.google.com/forms/d/e/1FAIpQLSc_786-_i_q4fo5ESqYnNyjIH0B5Rs45QIwejd_NV5AjNDZ7A/viewform
Post
1701
Smol TTS models are here! OuteTTS-0.1-350M - Zero shot voice cloning, built on LLaMa architecture, CC-BY license! 🔥
> Pure language modeling approach to TTS
> Zero-shot voice cloning
> LLaMa architecture w/ Audio tokens (WavTokenizer)
> BONUS: Works on-device w/ llama.cpp ⚡
Three-step approach to TTS:
> Audio tokenization using WavTokenizer (75 tok per second)
> CTC forced alignment for word-to-audio token mapping
> Structured prompt creation w/ transcription, duration, audio tokens
The model is extremely impressive for 350M parameters! Kudos to the
OuteAI team on such a brilliant feat - I'd love to see this be applied on larger data and smarter backbones like SmolLM 🤗
Check out the models here: OuteAI/outetts-6728aa71a53a076e4ba4817c
> Pure language modeling approach to TTS
> Zero-shot voice cloning
> LLaMa architecture w/ Audio tokens (WavTokenizer)
> BONUS: Works on-device w/ llama.cpp ⚡
Three-step approach to TTS:
> Audio tokenization using WavTokenizer (75 tok per second)
> CTC forced alignment for word-to-audio token mapping
> Structured prompt creation w/ transcription, duration, audio tokens
The model is extremely impressive for 350M parameters! Kudos to the
OuteAI team on such a brilliant feat - I'd love to see this be applied on larger data and smarter backbones like SmolLM 🤗
Check out the models here: OuteAI/outetts-6728aa71a53a076e4ba4817c
reach-vb
authored
a
paper
3 months ago
Post
3030
Smol models ftw! AMD released AMD OLMo 1B - beats OpenELM, tiny llama on MT Bench, Alpaca Eval - Apache 2.0 licensed 🔥
> Trained with 1.3 trillion (dolma 1.7) tokens on 16 nodes, each with 4 MI250 GPUs
> Three checkpoints:
- AMD OLMo 1B: Pre-trained model
- AMD OLMo 1B SFT: Supervised fine-tuned on Tulu V2, OpenHermes-2.5, WebInstructSub, and Code-Feedback datasets
- AMD OLMo 1B SFT DPO: Aligned with human preferences using Direct Preference Optimization (DPO) on UltraFeedback dataset
Key Insights:
> Pre-trained with less than half the tokens of OLMo-1B
> Post-training steps include two-phase SFT and DPO alignment
> Data for SFT:
- Phase 1: Tulu V2
- Phase 2: OpenHermes-2.5, WebInstructSub, and Code-Feedback
> Model checkpoints on the Hub & Integrated with Transformers ⚡️
Congratulations & kudos to AMD on a brilliant smol model release! 🤗
amd/amd-olmo-6723e7d04a49116d8ec95070
> Trained with 1.3 trillion (dolma 1.7) tokens on 16 nodes, each with 4 MI250 GPUs
> Three checkpoints:
- AMD OLMo 1B: Pre-trained model
- AMD OLMo 1B SFT: Supervised fine-tuned on Tulu V2, OpenHermes-2.5, WebInstructSub, and Code-Feedback datasets
- AMD OLMo 1B SFT DPO: Aligned with human preferences using Direct Preference Optimization (DPO) on UltraFeedback dataset
Key Insights:
> Pre-trained with less than half the tokens of OLMo-1B
> Post-training steps include two-phase SFT and DPO alignment
> Data for SFT:
- Phase 1: Tulu V2
- Phase 2: OpenHermes-2.5, WebInstructSub, and Code-Feedback
> Model checkpoints on the Hub & Integrated with Transformers ⚡️
Congratulations & kudos to AMD on a brilliant smol model release! 🤗
amd/amd-olmo-6723e7d04a49116d8ec95070
Post
2499
What a great day for Open Science!
@AIatMeta
released models, datasets, and code for many of its research artefacts! 🔥
1. Meta Segment Anything Model 2.1: An updated checkpoint with improved results on visually similar objects, small objects and occlusion handling. A new developer suite will be added to make it easier for developers to build with SAM 2.
Model checkpoints: reach-vb/sam-21-6702d40defe7611a8bafa881
2. Layer Skip: Inference code and fine-tuned checkpoints demonstrating a new method for enhancing LLM performance.
Model checkpoints: facebook/layerskip-666b25c50c8ae90e1965727a
3. SALSA: New code enables researchers to benchmark AI-based attacks to validate security for post-quantum cryptography.
Repo: https://github.com/facebookresearch/LWE-benchmarking
4. Meta Lingua: A lightweight and self-contained codebase designed to train language models at scale.
Repo: https://github.com/facebookresearch/lingua
5. Meta Open Materials: New open source models and the largest dataset to accelerate AI-driven discovery of new inorganic materials.
Model checkpoints: fairchem/OMAT24
6. MEXMA: A new research paper and code for our novel pre-trained cross-lingual sentence encoder covering 80 languages.
Model checkpoint: facebook/MEXMA
7. Self-Taught Evaluator: a new method for generating synthetic preference data to train reward models without relying on human annotations.
Model checkpoint: facebook/Self-taught-evaluator-llama3.1-70B
8. Meta Spirit LM: An open-source language model for seamless speech and text integration.
Repo: https://github.com/facebookresearch/spiritlm
1. Meta Segment Anything Model 2.1: An updated checkpoint with improved results on visually similar objects, small objects and occlusion handling. A new developer suite will be added to make it easier for developers to build with SAM 2.
Model checkpoints: reach-vb/sam-21-6702d40defe7611a8bafa881
2. Layer Skip: Inference code and fine-tuned checkpoints demonstrating a new method for enhancing LLM performance.
Model checkpoints: facebook/layerskip-666b25c50c8ae90e1965727a
3. SALSA: New code enables researchers to benchmark AI-based attacks to validate security for post-quantum cryptography.
Repo: https://github.com/facebookresearch/LWE-benchmarking
4. Meta Lingua: A lightweight and self-contained codebase designed to train language models at scale.
Repo: https://github.com/facebookresearch/lingua
5. Meta Open Materials: New open source models and the largest dataset to accelerate AI-driven discovery of new inorganic materials.
Model checkpoints: fairchem/OMAT24
6. MEXMA: A new research paper and code for our novel pre-trained cross-lingual sentence encoder covering 80 languages.
Model checkpoint: facebook/MEXMA
7. Self-Taught Evaluator: a new method for generating synthetic preference data to train reward models without relying on human annotations.
Model checkpoint: facebook/Self-taught-evaluator-llama3.1-70B
8. Meta Spirit LM: An open-source language model for seamless speech and text integration.
Repo: https://github.com/facebookresearch/spiritlm
- 3 replies
Post
23860
In regards to the latest mistral model and GGUFs for it:
Yes, they may be subpar and may require changes to llama.cpp to support the interleaved sliding window
Yes, I got excited when a conversion worked and released them ASAP
That said, generation seems to work right now and seems to mimic the output from spaces that are running the original model
I have appended -TEST to the model names in an attempt to indicate that they are not final or perfect, but if people still feel mislead and that it's not the right thing to do, please post (civilly) below your thoughts, I will highly consider pulling the conversions if that's what people think is best. After all, that's what I'm here for, in service to you all !
Yes, they may be subpar and may require changes to llama.cpp to support the interleaved sliding window
Yes, I got excited when a conversion worked and released them ASAP
That said, generation seems to work right now and seems to mimic the output from spaces that are running the original model
I have appended -TEST to the model names in an attempt to indicate that they are not final or perfect, but if people still feel mislead and that it's not the right thing to do, please post (civilly) below your thoughts, I will highly consider pulling the conversions if that's what people think is best. After all, that's what I'm here for, in service to you all !
-
- 6 replies
Post
5510
Multimodal Ichigo Llama 3.1 - Real Time Voice AI 🔥
> WhisperSpeech X Llama 3.1 8B
> Trained on 50K hours of speech (7 languages)
> Continually trained on 45hrs 10x A1000s
> MLS -> WhisperVQ tokens -> Llama 3.1
> Instruction tuned on 1.89M samples
> 70% speech, 20% transcription, 10% text
> Apache 2.0 licensed ⚡
Architecture:
> WhisperSpeech/ VQ for Semantic Tokens
> Llama 3.1 8B Instruct for Text backbone
> Early fusion (Chameleon)
I'm super bullish on HomeBrew/ Jan and early fusion, audio and text, multimodal models!
(P.S. Play with the demo on Hugging Face: jan-hq/Ichigo-llama3.1-s-instruct)
> WhisperSpeech X Llama 3.1 8B
> Trained on 50K hours of speech (7 languages)
> Continually trained on 45hrs 10x A1000s
> MLS -> WhisperVQ tokens -> Llama 3.1
> Instruction tuned on 1.89M samples
> 70% speech, 20% transcription, 10% text
> Apache 2.0 licensed ⚡
Architecture:
> WhisperSpeech/ VQ for Semantic Tokens
> Llama 3.1 8B Instruct for Text backbone
> Early fusion (Chameleon)
I'm super bullish on HomeBrew/ Jan and early fusion, audio and text, multimodal models!
(P.S. Play with the demo on Hugging Face: jan-hq/Ichigo-llama3.1-s-instruct)
Post
3180
NEW: Open Source Text/ Image to video model is out - MIT licensed - Rivals Gen-3, Pika & Kling 🔥
> Pyramid Flow: Training-efficient Autoregressive Video Generation method
> Utilizes Flow Matching
> Trains on open-source datasets
> Generates high-quality 10-second videos
> Video resolution: 768p
> Frame rate: 24 FPS
> Supports image-to-video generation
> Model checkpoints available on the hub 🤗: rain1011/pyramid-flow-sd3
> Pyramid Flow: Training-efficient Autoregressive Video Generation method
> Utilizes Flow Matching
> Trains on open-source datasets
> Generates high-quality 10-second videos
> Video resolution: 768p
> Frame rate: 24 FPS
> Supports image-to-video generation
> Model checkpoints available on the hub 🤗: rain1011/pyramid-flow-sd3
Post
2176
On-device AI framework ecosystem is blooming these days:
1. llama.cpp - All things Whisper, LLMs & VLMs - run across Metal, CUDA and other backends (AMD/ NPU etc)
https://github.com/ggerganov/llama.cpp
2. MLC - Deploy LLMs across platforms especially WebGPU (fastest WebGPU LLM implementation out there)
https://github.com/mlc-ai/web-llm
3. MLX - Arguably the fastest general purpose framework (Mac only) - Supports all major Image Generation (Flux, SDXL, etc), Transcription (Whisper), LLMs
https://github.com/ml-explore/mlx-examples
4. Candle - Cross-platform general purpose framework written in Rust - wide coverage across model categories
https://github.com/huggingface/candle
Honorable mentions:
1. Transformers.js - Javascript (WebGPU) implementation built on top of ONNXruntimeweb
https://github.com/xenova/transformers.js
2. Mistral rs - Rust implementation for LLMs & VLMs, built on top of Candle
https://github.com/EricLBuehler/mistral.rs
3. Ratchet - Cross platform, rust based WebGPU framework built for battle-tested deployments
https://github.com/huggingface/ratchet
4. Zml - Cross platform, Zig based ML framework
https://github.com/zml/zml
Looking forward to how the ecosystem would look 1 year from now - Quite bullish on the top 4 atm - but open source ecosystem changes quite a bit! 🤗
Also, which frameworks did I miss?
1. llama.cpp - All things Whisper, LLMs & VLMs - run across Metal, CUDA and other backends (AMD/ NPU etc)
https://github.com/ggerganov/llama.cpp
2. MLC - Deploy LLMs across platforms especially WebGPU (fastest WebGPU LLM implementation out there)
https://github.com/mlc-ai/web-llm
3. MLX - Arguably the fastest general purpose framework (Mac only) - Supports all major Image Generation (Flux, SDXL, etc), Transcription (Whisper), LLMs
https://github.com/ml-explore/mlx-examples
4. Candle - Cross-platform general purpose framework written in Rust - wide coverage across model categories
https://github.com/huggingface/candle
Honorable mentions:
1. Transformers.js - Javascript (WebGPU) implementation built on top of ONNXruntimeweb
https://github.com/xenova/transformers.js
2. Mistral rs - Rust implementation for LLMs & VLMs, built on top of Candle
https://github.com/EricLBuehler/mistral.rs
3. Ratchet - Cross platform, rust based WebGPU framework built for battle-tested deployments
https://github.com/huggingface/ratchet
4. Zml - Cross platform, Zig based ML framework
https://github.com/zml/zml
Looking forward to how the ecosystem would look 1 year from now - Quite bullish on the top 4 atm - but open source ecosystem changes quite a bit! 🤗
Also, which frameworks did I miss?
- 1 reply
Post
3003
What a great milestone to celebrate! The huggingface_hub library is slowly becoming a cornerstone of the Python ML ecosystem when it comes to interacting with the
@huggingface
Hub. It wouldn't be there without the hundreds of community contributions and feedback! No matter if you are loading a model, sharing a dataset, running remote inference or starting jobs on our infra, you are for sure using it! And this is only the beginning so give a star if you wanna follow the project 👉 https://github.com/huggingface/huggingface_hub
- 1 reply
Post
2881
Less than two days ago Kyutai Labs open sourced Moshi - an ~7.6B on-device Speech to Speech foundation model and Mimi - SoTA streaming speech codec! 🔥
The release includes:
1. Moshiko & Moshika - Moshi finetuned on synthetic data (CC-BY license) ( kyutai/moshi-v01-release-66eaeaf3302bef6bd9ad7acd)
2. Mimi - Streaiming Audio Codec, processes 24 kHz audio, down to a 12.5 Hz representation with a bandwidth of 1.1 kbps (CC-BY license) ( kyutai/mimi)
3. Model checkpoints & Inference codebase written in Rust (Candle), PyTorch & MLX (Apache license) (https://github.com/kyutai-labs/moshi)
How does Moshi work?
1. Moshi processes two audio streams: one for itself and one for the user, with the user's stream coming from audio input and Moshi's stream generated by the model.
2. Along with these audio streams, Moshi predicts text tokens for its speech, enhancing its generation quality.
3. The model uses a small Depth Transformer for codebook dependencies and a large 7B parameter Temporal Transformer for temporal dependencies.
4. The theoretical latency is 160ms, with a practical latency of around 200ms on an L4 GPU.
Model size & inference:
Moshiko/ka are 7.69B param models
bf16 ~16GB VRAM
8-bit ~8GB VRAM
4-bit ~4GB VRAM
You can run inference via Candle 🦀, PyTorch and MLX - based on your hardware.
The Kyutai team, @adefossez @lmz and team are cracked AF, they're bringing some serious firepower to the open source/ science AI scene, looking forward to what's next! 🐐
The release includes:
1. Moshiko & Moshika - Moshi finetuned on synthetic data (CC-BY license) ( kyutai/moshi-v01-release-66eaeaf3302bef6bd9ad7acd)
2. Mimi - Streaiming Audio Codec, processes 24 kHz audio, down to a 12.5 Hz representation with a bandwidth of 1.1 kbps (CC-BY license) ( kyutai/mimi)
3. Model checkpoints & Inference codebase written in Rust (Candle), PyTorch & MLX (Apache license) (https://github.com/kyutai-labs/moshi)
How does Moshi work?
1. Moshi processes two audio streams: one for itself and one for the user, with the user's stream coming from audio input and Moshi's stream generated by the model.
2. Along with these audio streams, Moshi predicts text tokens for its speech, enhancing its generation quality.
3. The model uses a small Depth Transformer for codebook dependencies and a large 7B parameter Temporal Transformer for temporal dependencies.
4. The theoretical latency is 160ms, with a practical latency of around 200ms on an L4 GPU.
Model size & inference:
Moshiko/ka are 7.69B param models
bf16 ~16GB VRAM
8-bit ~8GB VRAM
4-bit ~4GB VRAM
You can run inference via Candle 🦀, PyTorch and MLX - based on your hardware.
The Kyutai team, @adefossez @lmz and team are cracked AF, they're bringing some serious firepower to the open source/ science AI scene, looking forward to what's next! 🐐
- 1 reply