Aritra (@ariG23498) / X

Aritra

9,002 posts

Aritra

@ariG23498

MLE at @huggingface

OOMs

Joined August 2016

Pinned
Aritra
@ariG23498
19h
It is profiling time! ⏰ In Part 2 we cover: > trace a Linear layer > talk about mul + add vs linear > gemm epilogues (my favourite part) > trace MLP > torch compile vs liger kernels > use hugging face kernels Bookmark this for a fun weekend read! 😎
00:00
5.8K
Aritra
@ariG23498
Sep 21, 2024
I am joining @huggingface as an Open-Source ML Release Engineer! Excited to collaborate with amazing teams and external groups to push open-source ML forward. Grateful to @RisingSayak and @carrigmat for their constant guidance.
63K
Aritra
@ariG23498
Sep 25, 2024
My favourite bit about the Llama 3.2 release is the small models. Both 1B and 3B despite being quite small are very capable. While there are benchmarks that prove my point, I took them on a spin for something totally different, assisted decoding. [1/N] 🧵⤵️
123K
Aritra
@ariG23498
Sep 26, 2024
Look maa! I am in the release announcement 🤗
45K
Aritra
@ariG23498
Oct 4, 2024
In the Large Language Model world, there are mostly two variants of the same model. * base model: completes a prompt (llama-3.2-1B) * instruction tuned model: completes an instruction (llama-3.2-1B-Instruct) [1/N] 🧵⬇️
107K
Aritra
@ariG23498
Oct 1, 2024
A simple hack to calculating how much VRAM you would need to run a model. X Billion params: * X/2 gb for int4 * X gb for fp8 * 2*X gb for bf16/fp16 * 4*x gb for fp32 Think in terms of bytes a data type would need. (example attached)
25K
Aritra
@ariG23498
May 16, 2022
This is the first time I have recieved a signed copy. Humbled! CC: @huggingface
Aritra
@ariG23498
Jun 18, 2025
As promised here is the video classification fine tuning notebook with @AIatMeta's VJEPA 2. Links and some description 🧶⤵️
26K
Aritra
@ariG23498
Jun 4, 2025
Documented all my learning while we implemented KV Cache from scratch in nanoVLM. We achieve 38% speedup in generation speed. This blog post covers: > attention mechanism and its redundancies > top notch ascii art explaining concepts > a lot of code pin pointing to the changes
16K
Aritra
@ariG23498
Sep 27, 2024
Next week there will be a lot of opportunities to work in open source! Watch this space 🤗
16K
Aritra
@ariG23498
Jul 8, 2025
"Why is the training so slow?" We figure out that starving the model from data, or providing it with padding tokens leads to training delays. We publish a write up which talks about data efficiency, and how we apply them to nanoVLM. Spoiler: We use knapsack algorithm. 🧵⤵️
14K
Aritra
@ariG23498
Dec 23, 2024
Controlling text generation has always been a problem. What if I wanted my LLM to always end with a particular phrase, or wanted a longer generation? In this blog post (with @a_erdem4) we venture on covering logits processing, a simple way to achieve this control. 🧵👇🏻 [1/N]
44K
Aritra
@ariG23498
Dec 2, 2024
The @Alibaba_Qwen team is doing so much for the community by keeping research open and constructive. They listen to the community and put efforts in building competitive models. I was intrigued by their latest `Qwen/QwQ-32B-Preview` model and wanted to play with it. [1/N]
19K
Aritra
@ariG23498
Jun 2, 2022
I do not know who needs to hear it but 👇 Overfit your model on a single batch of data first before anything. Thank you @karpathy karpathy.github.io/2019/04/25/rec…