Inference for PROs
β’
47
Victor Mustar PRO
victor
AI & ML interests
Building the UX of this website
Articles
Organizations
victor's activity
reacted to
nroggendorff's
post with π€―
6 days ago
Post
1754
Did you guys know that if you try to link a prepaid card to huggingface it won't work, but then if you press the button again it links anyway? Then you can lock the card (deny any charges), and get resources for free? You're welcome :P
- 4 replies
reacted to
automatedstockminingorg's
post with π
6 days ago
Post
2264
hi everyone,
i have trained a Qwen 14b model on a smaller dataset, but its now very tricky because i have got nowhere to use it via inference (the paid for inference on hf costs quite a lot), does anyone know of anywhere where i can deploy my model and use it via api for a reasonable cost, or ideally none. thanks
i have trained a Qwen 14b model on a smaller dataset, but its now very tricky because i have got nowhere to use it via inference (the paid for inference on hf costs quite a lot), does anyone know of anywhere where i can deploy my model and use it via api for a reasonable cost, or ideally none. thanks
-
- 6 replies
reacted to
yongchanghao's
post with π₯
6 days ago
Post
3607
We just released a paper (NeuZip) that compresses VRAM in a lossless manner to run larger models. This should be particularly useful when VRAM is insufficient during training/inference. Specifically, we look inside each floating number and find that the exponents are highly compressible (as shown in the figure below).
Read more about the work at NeuZip: Memory-Efficient Training and Inference with Dynamic Compression of Neural Networks (2410.20650)
Read more about the work at NeuZip: Memory-Efficient Training and Inference with Dynamic Compression of Neural Networks (2410.20650)
reacted to
qq8933's
post with π
6 days ago
Post
5173
LLaMA-O1: Open Large Reasoning Model Frameworks For Training, Inference and Evaluation With PyTorch and HuggingFace
Large Reasoning Models powered by Monte Carlo Tree Search (MCTS), Self-Play Reinforcement Learning, PPO, AlphaGo Zero's dua policy paradigm and Large Language Models!
https://github.com/SimpleBerry/LLaMA-O1/
What will happen when you compound MCTS β€ LLM β€ Self-Play β€RLHF?
Just a little bite of strawberry!π
Past related works:
LLaMA-Berry: Pairwise Optimization for O1-like Olympiad-Level Mathematical Reasoning (2410.02884)
Accessing GPT-4 level Mathematical Olympiad Solutions via Monte Carlo Tree Self-refine with LLaMa-3 8B (2406.07394)
Large Reasoning Models powered by Monte Carlo Tree Search (MCTS), Self-Play Reinforcement Learning, PPO, AlphaGo Zero's dua policy paradigm and Large Language Models!
https://github.com/SimpleBerry/LLaMA-O1/
What will happen when you compound MCTS β€ LLM β€ Self-Play β€RLHF?
Just a little bite of strawberry!π
Past related works:
LLaMA-Berry: Pairwise Optimization for O1-like Olympiad-Level Mathematical Reasoning (2410.02884)
Accessing GPT-4 level Mathematical Olympiad Solutions via Monte Carlo Tree Self-refine with LLaMa-3 8B (2406.07394)
- 2 replies
reacted to
Muhammadreza's
post with β€οΈ
8 days ago
Post
2519
Hey guys.
This is my first post here on huggingface. I'm glad to be a part of this amazing community!
This is my first post here on huggingface. I'm glad to be a part of this amazing community!
-
- 2 replies
reacted to
TuringsSolutions's
post with π
9 days ago
Post
2856
I have been seeing a specific type of AI hype more and more, I call it, releasing research expecting that no one will ever reproduce your methods, then overhyping your results. I test the methodology of maybe 4-5 research papers per day. That is how I find a lot of my research. Usually, 3-4 of those experiments end up not being reproduceable for some reason. I am starting to think it is not accidental.
So, I am launching a new series where I specifically showcase a research paper by reproducing their methodology and highlighting the blatant flaws that show up when you actually do this. Here is Episode 1!
https://www.youtube.com/watch?v=JLa0cFWm1A4
So, I am launching a new series where I specifically showcase a research paper by reproducing their methodology and highlighting the blatant flaws that show up when you actually do this. Here is Episode 1!
https://www.youtube.com/watch?v=JLa0cFWm1A4
-
- 5 replies
reacted to
Smooke's
post with π
12 days ago
Post
1819
Check out The AI Writing Contest with
@BrightData
! https://www.contests.hackernoon.com/ai-writing-contest Cash prizes for innovative approaches to AI and LLM training. We publish blog posts, research papers, stories of side hustles, you name it.
Any story tagged #AI enters to win. Most recent stories: https://hackernoon.com/tagged/ai and RSS feed https://hackernoon.com/tagged/ai/feed
Couple of favorite recent posts we published:
Why Salesforce and Microsoft Are Battling for the Future of AI Agents https://hackernoon.com/why-salesforce-and-microsoft-are-battling-for-the-future-of-ai-agents
Decentralized AI Summit at MIT Votes OriginTrail As The Best Decentralized AI Project https://hackernoon.com/decentralized-ai-summit-at-mit-votes-origintrail-as-the-best-decentralized-ai-project
Studying is Overrated https://hackernoon.com/studying-is-overrated
Why Canβt AI Count Letters??? https://hackernoon.com/why-cant-ai-count-letters
The Paradox of AI: If It Can't Replace us, Is It Making Us Dumber? https://hackernoon.com/the-paradox-of-ai-if-it-cant-replace-us-is-it-making-us-dumber
How Does Human Memory Work? https://hackernoon.com/how-does-human-memory-work
Is AI Actually Writing Production-Ready Code? https://hackernoon.com/is-ai-actually-writing-production-ready-code
Our AI Coding Tool Went Viral, Then Everything Broke. This is What We Learned. https://hackernoon.com/our-ai-coding-tool-went-viral-then-everything-broke-this-is-what-we-learned
Startups of The Year: Meet the AI Industry https://hackernoon.com/startups-of-the-year-meet-the-ai-industry
Nobel Prize Winner Geoffrey Hinton Explores Two Paths to Intelligence in AI Lecture https://hackernoon.com/nobel-prize-winner-geoffrey-hinton-explores-two-paths-to-intelligence-in-ai-lecture
Comparing AI vs. Blockchain Hype https://hackernoon.com/comparing-ai-vs-blockchain-hype
The SaaS Apocalypse and How aI Will Give Birth to One-person Tech Giants https://hackernoon.com/the-saas-apocalypse-and-how-ai-wi
Any story tagged #AI enters to win. Most recent stories: https://hackernoon.com/tagged/ai and RSS feed https://hackernoon.com/tagged/ai/feed
Couple of favorite recent posts we published:
Why Salesforce and Microsoft Are Battling for the Future of AI Agents https://hackernoon.com/why-salesforce-and-microsoft-are-battling-for-the-future-of-ai-agents
Decentralized AI Summit at MIT Votes OriginTrail As The Best Decentralized AI Project https://hackernoon.com/decentralized-ai-summit-at-mit-votes-origintrail-as-the-best-decentralized-ai-project
Studying is Overrated https://hackernoon.com/studying-is-overrated
Why Canβt AI Count Letters??? https://hackernoon.com/why-cant-ai-count-letters
The Paradox of AI: If It Can't Replace us, Is It Making Us Dumber? https://hackernoon.com/the-paradox-of-ai-if-it-cant-replace-us-is-it-making-us-dumber
How Does Human Memory Work? https://hackernoon.com/how-does-human-memory-work
Is AI Actually Writing Production-Ready Code? https://hackernoon.com/is-ai-actually-writing-production-ready-code
Our AI Coding Tool Went Viral, Then Everything Broke. This is What We Learned. https://hackernoon.com/our-ai-coding-tool-went-viral-then-everything-broke-this-is-what-we-learned
Startups of The Year: Meet the AI Industry https://hackernoon.com/startups-of-the-year-meet-the-ai-industry
Nobel Prize Winner Geoffrey Hinton Explores Two Paths to Intelligence in AI Lecture https://hackernoon.com/nobel-prize-winner-geoffrey-hinton-explores-two-paths-to-intelligence-in-ai-lecture
Comparing AI vs. Blockchain Hype https://hackernoon.com/comparing-ai-vs-blockchain-hype
The SaaS Apocalypse and How aI Will Give Birth to One-person Tech Giants https://hackernoon.com/the-saas-apocalypse-and-how-ai-wi
reacted to
prithivMLmods's
post with π
18 days ago
Post
1895
Crossed more than 100+ of the best and most interesting LoRAs from open-source contributors, based on Flux Dev and Schnell models, which are part of the Flux LoRA DLC Space. If I missed any that are particularly interesting within the Flux group, feel free to contribute to the space!
π₯³Flux LoRA Dlc : prithivMLmods/FLUX-LoRA-DLC
Thankyou!
π₯³Flux LoRA Dlc : prithivMLmods/FLUX-LoRA-DLC
Thankyou!
reacted to
yuexiang96's
post with π
18 days ago
Post
2931
π Iβve always had a dream of making AI accessible to everyone, regardless of location or language. However, current open MLLMs often respond in English, even to non-English queries!
π Introducing Pangea: A Fully Open Multilingual Multimodal LLM supporting 39 languages! πβ¨
https://neulab.github.io/Pangea/
https://arxiv.org/pdf/2410.16153
The Pangea family includes three major components:
π₯ Pangea-7B: A state-of-the-art multilingual multimodal LLM capable of 39 languages! Not only does it excel in multilingual scenarios, but it also matches or surpasses English-centric models like Llama 3.2, Molmo, and LlavaOneVision in English performance.
π PangeaIns: A 6M multilingual multimodal instruction tuning dataset across 39 languages. ποΈ With 40% English instructions and 60% multilingual instructions, it spans various domains, including 1M culturally-relevant images sourced from LAION-Multi. π¨
π PangeaBench: A comprehensive evaluation benchmark featuring 14 datasets in 47 languages. Evaluation can be tricky, so we carefully curated existing benchmarks and introduced two new datasets: xChatBench (human-annotated wild queries with fine-grained evaluation criteria) and xMMMU (a meticulously machine-translated version of MMMU).
Check out more details: https://x.com/xiangyue96/status/1848753709787795679
π Introducing Pangea: A Fully Open Multilingual Multimodal LLM supporting 39 languages! πβ¨
https://neulab.github.io/Pangea/
https://arxiv.org/pdf/2410.16153
The Pangea family includes three major components:
π₯ Pangea-7B: A state-of-the-art multilingual multimodal LLM capable of 39 languages! Not only does it excel in multilingual scenarios, but it also matches or surpasses English-centric models like Llama 3.2, Molmo, and LlavaOneVision in English performance.
π PangeaIns: A 6M multilingual multimodal instruction tuning dataset across 39 languages. ποΈ With 40% English instructions and 60% multilingual instructions, it spans various domains, including 1M culturally-relevant images sourced from LAION-Multi. π¨
π PangeaBench: A comprehensive evaluation benchmark featuring 14 datasets in 47 languages. Evaluation can be tricky, so we carefully curated existing benchmarks and introduced two new datasets: xChatBench (human-annotated wild queries with fine-grained evaluation criteria) and xMMMU (a meticulously machine-translated version of MMMU).
Check out more details: https://x.com/xiangyue96/status/1848753709787795679
reacted to
albertvillanova's
post with π
18 days ago
Post
1887
Finding the Best SmolLM for Your Project
Need an LLM assistant but unsure which hashtag#smolLM to run locally? With so many models available, how can you decide which one suits your needs best? π€
If the model youβre interested in is evaluated on the Hugging Face Open LLM Leaderboard, thereβs an easy way to compare them: use the model Comparator tool: open-llm-leaderboard/comparator
Letβs walk through an exampleπ
Letβs compare two solid options:
- Qwen2.5-1.5B-Instruct from Alibaba Cloud Qwen (1.5B params)
- gemma-2-2b-it from Google (2.5B params)
For an assistant, you want a model thatβs great at instruction following. So, how do these two models stack up on the IFEval task?
What about other evaluations?
Both models are close in performance on many other tasks, showing minimal differences. Surprisingly, the 1.5B Qwen model performs just as well as the 2.5B Gemma in many areas, even though it's smaller in size! π
This is a great example of how parameter size isnβt everything. With efficient design and training, a smaller model like Qwen2.5-1.5B can match or even surpass larger models in certain tasks.
Looking for other comparisons? Drop your model suggestions below! π
Need an LLM assistant but unsure which hashtag#smolLM to run locally? With so many models available, how can you decide which one suits your needs best? π€
If the model youβre interested in is evaluated on the Hugging Face Open LLM Leaderboard, thereβs an easy way to compare them: use the model Comparator tool: open-llm-leaderboard/comparator
Letβs walk through an exampleπ
Letβs compare two solid options:
- Qwen2.5-1.5B-Instruct from Alibaba Cloud Qwen (1.5B params)
- gemma-2-2b-it from Google (2.5B params)
For an assistant, you want a model thatβs great at instruction following. So, how do these two models stack up on the IFEval task?
What about other evaluations?
Both models are close in performance on many other tasks, showing minimal differences. Surprisingly, the 1.5B Qwen model performs just as well as the 2.5B Gemma in many areas, even though it's smaller in size! π
This is a great example of how parameter size isnβt everything. With efficient design and training, a smaller model like Qwen2.5-1.5B can match or even surpass larger models in certain tasks.
Looking for other comparisons? Drop your model suggestions below! π
reacted to
ImranzamanML's
post with π₯
18 days ago
Post
1675
Today lets discuss about 32-bit (FP32) and 16-bit (FP16) floating-point!
Floating-point numbers are used to represent real numbers (like decimals) and they consist of three parts:
32-bit Floating Point (FP32)
Total bits: 32 bits
Sign bit: 1 bit
Exponent: 8 bits
Mantissa: 23 bits
For example:
A number like -15.375 would be represented as:
Sign bit: 1 (negative number)
Exponent: Stored after being adjusted by a bias (127 in FP32).
Mantissa: The significant digits after converting the number to binary.
16-bit Floating Point (FP16)
Total bits: 16 bits
Sign bit: 1 bit
Exponent: 5 bits
Mantissa: 10 bits
Example:
A number like -15.375 would be stored similarly:
Sign bit: 1 (negative number)
Exponent: Uses 5 bits, limiting the range compared to FP32.
Mantissa: Only 10 bits for precision.
Precision and Range
FP32: Higher precision and larger range, with about 7 decimal places of accuracy.
FP16: Less precision (around 3-4 decimal places), smaller range but faster computations and less memory use.
Floating-point numbers are used to represent real numbers (like decimals) and they consist of three parts:
Sign bit:
Indicates whether the number is positive (0) or negative (1).
Exponent:
Determines the scale of the number (i.e., how large or small it is by shifting the decimal point).
Mantissa (or fraction):
Represents the actual digits of the number.32-bit Floating Point (FP32)
Total bits: 32 bits
Sign bit: 1 bit
Exponent: 8 bits
Mantissa: 23 bits
For example:
A number like -15.375 would be represented as:
Sign bit: 1 (negative number)
Exponent: Stored after being adjusted by a bias (127 in FP32).
Mantissa: The significant digits after converting the number to binary.
16-bit Floating Point (FP16)
Total bits: 16 bits
Sign bit: 1 bit
Exponent: 5 bits
Mantissa: 10 bits
Example:
A number like -15.375 would be stored similarly:
Sign bit: 1 (negative number)
Exponent: Uses 5 bits, limiting the range compared to FP32.
Mantissa: Only 10 bits for precision.
Precision and Range
FP32: Higher precision and larger range, with about 7 decimal places of accuracy.
FP16: Less precision (around 3-4 decimal places), smaller range but faster computations and less memory use.
-
- 3 replies
reacted to
Tonic's
post with π
19 days ago
Post
1431
hey there folks,
twitter is aweful isnt it ? just getting into the habbit of using hf/posts for shares π¦π¦
Tonic/on-device-granite-3.0-1b-a400m-instruct
new granite on device instruct model demo , hope you like it ππ
twitter is aweful isnt it ? just getting into the habbit of using hf/posts for shares π¦π¦
Tonic/on-device-granite-3.0-1b-a400m-instruct
new granite on device instruct model demo , hope you like it ππ
reacted to
abhishek's
post with π
19 days ago
reacted to
prithivMLmods's
post with π
22 days ago
Post
2346
SambaNova βοΈ
β‘ Inference API with cURL Demo: prithivMLmods/sambanova-inference-api
πSambanova API Documentation : (grab your APIs here) https://cloud.sambanova.ai/apis
Sambanova's Inference API.
SambaNova X Gradio
π Documentation: https://community.sambanova.ai/docs
β‘ Inference API with cURL Demo: prithivMLmods/sambanova-inference-api
πSambanova API Documentation : (grab your APIs here) https://cloud.sambanova.ai/apis
export SAMBANOVA_API_KEY=<your token>Sambanova's Inference API.
pip install sambanova-gradioSambaNova X Gradio
import gradio as gr
import sambanova_gradio
gr.load(
name='Meta-Llama-3.1-405B-Instruct',
src=sambanova_gradio.registry,
).launch()π Documentation: https://community.sambanova.ai/docs
reacted to
Tonic's
post with π
23 days ago
Post
959
if you're encountering 500 errors on spaces that seem to work otherwise , kindly consider screenshotting and sharing the link here : https://discord.com/channels/879548962464493619/1295847667515129877
-
- 7 replies
reacted to
celinah's
post with β€οΈ
23 days ago
Post
1040
π£ πππππππππππ_πππ v0.26.0 is out with some new features and improvements!
β¨ π§πΌπ½ ππΆπ΄π΅πΉπΆπ΄π΅ππ:
- πΒ Multiple access tokens support: Easily manage multiple access tokens with new CLI commands. Perfect for handling multiple tokens with specific permissions in production or when collaborating with external teams.
- πΌοΈ Conversational VLMs inference is now supported withΒ InferenceClient's chat completion!
- π Daily Papers API: Seamlessly search and retrieve detailed paper information from the Hub!
Weβve also introduced multiple bug fixes and quality-of-life improvements - thanks to the awesome contributions from our community! π€
Check out the release notes here: Wauplin/huggingface_hub#9
and you can try it out now π
β¨ π§πΌπ½ ππΆπ΄π΅πΉπΆπ΄π΅ππ:
- πΒ Multiple access tokens support: Easily manage multiple access tokens with new CLI commands. Perfect for handling multiple tokens with specific permissions in production or when collaborating with external teams.
- πΌοΈ Conversational VLMs inference is now supported withΒ InferenceClient's chat completion!
- π Daily Papers API: Seamlessly search and retrieve detailed paper information from the Hub!
Weβve also introduced multiple bug fixes and quality-of-life improvements - thanks to the awesome contributions from our community! π€
Check out the release notes here: Wauplin/huggingface_hub#9
and you can try it out now π
pip install huggingface_hub==0.26.0
reacted to
merve's
post with π₯
23 days ago
Post
1927
It's raining depth estimation models βοΈ
DepthPro is a zero-shot depth estimation model by Apple, it's fast, sharp and accurate π₯
Demo: akhaliq/depth-pro
Model: apple/DepthPro
Paper page: Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (2410.02073)
The model consists of two encoders: an encoder for patches and an image encoder πΌοΈ The outputs of both are merged to decode to depth maps and get the focal length.
The model outperforms the previous state-of-the-art models in average of various benchmarks π
DepthPro is a zero-shot depth estimation model by Apple, it's fast, sharp and accurate π₯
Demo: akhaliq/depth-pro
Model: apple/DepthPro
Paper page: Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (2410.02073)
The model consists of two encoders: an encoder for patches and an image encoder πΌοΈ The outputs of both are merged to decode to depth maps and get the focal length.
The model outperforms the previous state-of-the-art models in average of various benchmarks π
reacted to
dvilasuero's
post with π
23 days ago
Post
950
Big news! You can now build strong ML models without days of human labelling
You simply:
- Define your dataset, including annotation guidelines, labels and fields
- Optionally label some records manually.
- Use an LLM to auto label your data with a human (you? your team?) in the loop!
Get started with this blog post:
https://huggingface.co./blog/sdiazlor/custom-text-classifier-ai-human-feedback
You simply:
- Define your dataset, including annotation guidelines, labels and fields
- Optionally label some records manually.
- Use an LLM to auto label your data with a human (you? your team?) in the loop!
Get started with this blog post:
https://huggingface.co./blog/sdiazlor/custom-text-classifier-ai-human-feedback
reacted to
davidberenstein1957's
post with β
23 days ago
Post
1664
You can now build a custom text classifier without days of human labeling!
π LLMs work reasonably well as text classifiers.
π They are expensive to run at scale and their performance drops in specialized domains.
π Purpose-built classifiers have low latency and can potentially run on CPU.
π They require labeled training data.
Combine the best of both worlds: the automatic labeling capabilities of LLMs and the high-quality annotations from human experts to train and deploy a specialized model.
Blog: https://huggingface.co./blog/sdiazlor/custom-text-classifier-ai-human-feedback
π LLMs work reasonably well as text classifiers.
π They are expensive to run at scale and their performance drops in specialized domains.
π Purpose-built classifiers have low latency and can potentially run on CPU.
π They require labeled training data.
Combine the best of both worlds: the automatic labeling capabilities of LLMs and the high-quality annotations from human experts to train and deploy a specialized model.
Blog: https://huggingface.co./blog/sdiazlor/custom-text-classifier-ai-human-feedback
posted
an
update
24 days ago
Post
1443
nvidia/Llama-3.1-Nemotron-70B-Instruct-HF is now available in HuggingChat if you want to try it!
https://huggingface.co./chat/models/nvidia/Llama-3.1-Nemotron-70B-Instruct-HF
https://huggingface.co./chat/models/nvidia/Llama-3.1-Nemotron-70B-Instruct-HF