Post
65
In celebration of the new storage graph feature on the Hub, here's mine 😊 :
Post inspired by @ZennyKenny
Post inspired by @ZennyKenny
Parag Ekbote PRO
AINovice2005
AI & ML interests
ML Engineer passionate about taking models from research to production. 1 year supporting tech startups. Active OSS contributor.
Recent Activity
posted an update about 15 hours ago
In celebration of the new storage graph feature on the Hub, here's mine 😊 :
Post inspired by @ZennyKenny updated a model 3 days ago
AINovice2005/ModernBERT-base-lora-cicflow-1m-r8 updated a model 3 days ago
AINovice2005/ModernBERT-base-lora-cicflow-1m-r16Organizations
posted an update about 15 hours ago
posted an update 4 days ago
Post
111
I recently created my first storage bucket to store experiment data of my performance analysis of 15 tokenizers across 20 languages.
The setup is simple enough for a new product and can be scalable depending on the use-case 🤗 .
Bucket: https://huggingface.co/buckets/AINovice2005/tokenizer-benchmark
github gist: https://gist.github.com/ParagEkbote/b3877f667f84cbb9a27bdaca94ba662a
Article: https://medium.com/@paragekbote23/one-sentence-fifteen-tokenizers-a-tokenizer-benchmarking-pipeline-with-hf-storage-buckets-2e59790276fd
The setup is simple enough for a new product and can be scalable depending on the use-case 🤗 .
Bucket: https://huggingface.co/buckets/AINovice2005/tokenizer-benchmark
github gist: https://gist.github.com/ParagEkbote/b3877f667f84cbb9a27bdaca94ba662a
Article: https://medium.com/@paragekbote23/one-sentence-fifteen-tokenizers-a-tokenizer-benchmarking-pipeline-with-hf-storage-buckets-2e59790276fd
posted an update 5 days ago
Post
126
I'm pleased to announce that native trackio support has landed in optuna v.4.8.0.
Now you can store HPO trial data to the 🤗 Hub as a dataset: AINovice2005/toy-optuna-hpo-dataset
or Space: AINovice2005/optuna-dashboard
Article which talks about the integration: https://medium.com/@paragekbote23/designing-and-contributing-an-optuna-integration-that-performs-well-in-practice-21a07d77ec2c
Now you can store HPO trial data to the 🤗 Hub as a dataset: AINovice2005/toy-optuna-hpo-dataset
or Space: AINovice2005/optuna-dashboard
Article which talks about the integration: https://medium.com/@paragekbote23/designing-and-contributing-an-optuna-integration-that-performs-well-in-practice-21a07d77ec2c
posted an update 8 days ago
Post
84
Pro tip2: You can treat HF datasets as versioned repos by pinning a specific revision (tag, branch or commit) when downloading files. 🧠
This ensures your data processing pipelines always use the exact dataset state before passing the data to the model. It enables reproducible pipelines and allows for reliable outputs of your ML system.
This ensures your data processing pipelines always use the exact dataset state before passing the data to the model. It enables reproducible pipelines and allows for reliable outputs of your ML system.
from huggingface_hub import hf_hub_download
data_path = hf_hub_download(
repo_id="lysandre/arxiv-nlp",
filename="train.parquet",
repo_type="dataset",
revision="877b84a8f93f2d619faa2a6e514a32beef88ab0a"
)reactedto sdiazlor's post with 🚀 8 days ago
Post
2589
More OSS than ever with the latest pruna 0.3.2 release. It extends existing algorithm families, such as compilers, kernels, and pruners, and adds new ones, including decoders, distillers, enhancers, and recoverers. But it's not only a collection of algorithms; instead, you can easily combine them to get the biggest efficiency win.
Read the full blog here: https://huggingface.co/blog/PrunaAI/pruna-0-3-2-open-source-optimization-algorithms
Read the full blog here: https://huggingface.co/blog/PrunaAI/pruna-0-3-2-open-source-optimization-algorithms
posted an update 11 days ago
Post
372
Pro tip: If you are finetuning any model with tensorboard logs enabled, be sure to upload them to HF Hub as event artifacts, they can be viewed instantly. 🚀
I previously remembered this done in the notus model release: argilla/notus-7b-v1
Examples:
AINovice2005/ModernBERT-base-lora-cicflow-1m-r8
AINovice2005/ModernBERT-base-lora-cicflow-1m-r4
AINovice2005/ModernBERT-base-lora-cicflow-1m-r16
cc: @davidberenstein1957
I previously remembered this done in the notus model release: argilla/notus-7b-v1
Examples:
AINovice2005/ModernBERT-base-lora-cicflow-1m-r8
AINovice2005/ModernBERT-base-lora-cicflow-1m-r4
AINovice2005/ModernBERT-base-lora-cicflow-1m-r16
cc: @davidberenstein1957
repliedto their post 12 days ago
Thanks for reaching out.
This is a specialized kernel for diffusion models. Typically, CUDA kernels are used for speeding up inference, distributed training of large models, etc. So, scope of the kernel can differ for each specific process.
posted an update 13 days ago
Post
471
Just published my first cuda kernel, inspired by Sage Attention. Feel free to try it out ☺️
AINovice2005/attention-int8
AINovice2005/attention-int8
- 2 replies
repliedto danielhanchen's post 20 days ago
Congratulations on reaching this great milestone @danielhanchen 🙌
reactedto danielhanchen's post with 🔥 about 1 month ago
Post
5196
We collaborated with Hugging Face to enable you to train MoE models 12× faster with 35% less VRAM via our new Triton kernels (no accuracy loss). 🤗
Train gpt-oss locally on 12.8GB VRAM with our free notebooks: https://unsloth.ai/docs/new/faster-moe
Train gpt-oss locally on 12.8GB VRAM with our free notebooks: https://unsloth.ai/docs/new/faster-moe
- 1 reply
reactedto singhsidhukuldeep's post with 👍 over 1 year ago
Post
1740
Good folks at @nvidia have just released NVLM 1.0, a family of frontier-class multimodal large language models that achieve state-of-the-art results across vision-language tasks.
Here is how they did it:
1. Model Architecture Design:
- Developed three model architectures:
a) NVLM-D: Decoder-only architecture
b) NVLM-X: Cross-attention-based architecture
c) NVLM-H: Novel hybrid architecture
2. Vision Encoder:
- Used InternViT-6B-448px-V1-5 as the vision encoder
- Implemented dynamic high-resolution (DHR) input handling
3. Language Model:
- Used Qwen2-72B-Instruct as the base LLM
4. Training Data Curation:
- Carefully curated high-quality pretraining and supervised fine-tuning datasets
- Included diverse task-oriented datasets for various capabilities
5. Pretraining:
- Froze LLM and vision encoder
- Trained only modality-alignment modules (e.g., MLP projector, cross-attention layers)
- Used a large batch size of 2048
6. Supervised Fine-Tuning (SFT):
- Unfroze LLM while keeping the vision encoder frozen
- Trained on multimodal SFT datasets and high-quality text-only SFT data
- Implemented 1-D tile tagging for dynamic high-resolution inputs
7. Evaluation:
- Evaluated on multiple vision-language benchmarks
- Compared performance to leading proprietary and open-source models
8. Optimization:
- Iterated on model designs and training approaches
- Used smaller 34B models for faster experimentation before scaling to 72B
9. Now comes the best part...Open-Sourcing:
- Released model weights and full technical details to the research community
The paper provides fascinating insights into architecture design, training data curation, and achieving production-grade multimodality. A must-read for anyone working on multimodal AI!
Here is how they did it:
1. Model Architecture Design:
- Developed three model architectures:
a) NVLM-D: Decoder-only architecture
b) NVLM-X: Cross-attention-based architecture
c) NVLM-H: Novel hybrid architecture
2. Vision Encoder:
- Used InternViT-6B-448px-V1-5 as the vision encoder
- Implemented dynamic high-resolution (DHR) input handling
3. Language Model:
- Used Qwen2-72B-Instruct as the base LLM
4. Training Data Curation:
- Carefully curated high-quality pretraining and supervised fine-tuning datasets
- Included diverse task-oriented datasets for various capabilities
5. Pretraining:
- Froze LLM and vision encoder
- Trained only modality-alignment modules (e.g., MLP projector, cross-attention layers)
- Used a large batch size of 2048
6. Supervised Fine-Tuning (SFT):
- Unfroze LLM while keeping the vision encoder frozen
- Trained on multimodal SFT datasets and high-quality text-only SFT data
- Implemented 1-D tile tagging for dynamic high-resolution inputs
7. Evaluation:
- Evaluated on multiple vision-language benchmarks
- Compared performance to leading proprietary and open-source models
8. Optimization:
- Iterated on model designs and training approaches
- Used smaller 34B models for faster experimentation before scaling to 72B
9. Now comes the best part...Open-Sourcing:
- Released model weights and full technical details to the research community
The paper provides fascinating insights into architecture design, training data curation, and achieving production-grade multimodality. A must-read for anyone working on multimodal AI!
Nice Playlist.
Do you think that golden hour (piano version) can fit in?