Research Scientist @ Meta (Superintelligence Labs) | LLM Training Efficiency and Optimizer Design | Large Batch Scaling | Distributed AI Systems
San Jose, CA
Joined April 2010
- Tweets122
- Following552
- Followers224
- Likes188
Pinned Tweet
1/11 How should the gradient noise scale (GNS) be defined for non-Euclidean optimizers, such as SignSGD and Muon? Key insights from our recent work on extending GNS to stochastic sign and spectral descent; see arxiv.org/abs/2602.03001.
10/11 The benefits also extend to vision tasks: on the ImageWoof SimpleViT workload, the adaptive strategy achieves a significant reduction in optimizer steps while preserving or improving final validation accuracy.
1
2
133
11/11 This research was a collaborative effort across Mila and Meta. Credit to all authors: @_Hiroki11x , @s_gupta1945, Youssef Briki, Ioannis Mitliagkas, Irina Rish, @paramsraman and @hjmshi. The paper is accepted to ICML 2026.
4
193
Parameswaran Raman retweeted
1/ today we're releasing muse spark, the first model from MSL. nine months ago we rebuilt our ai stack from scratch. new infrastructure, new architecture, new data pipelines. muse spark is the result of that work, and now it powers meta ai. 🧵
We propose GPA (Generalized Primal Averaging), a new optimizer for LLM Training, making interesting connections to DiLoCo and Schedule-Free!
Paper: arxiv.org/abs/2512.17131 and Code: github.com/facebookresearch/…. Checkout below thread for more details.
1/10 Are DiLoCo and Schedule-Free actually related? A brief history and unusually late advertisement for our work: Smoothing DiLoCo with Primal Averaging for Faster Training of LLMs (see arxiv.org/abs/2512.17131).
3
167
Parameswaran Raman retweeted
1/14 Is Muon “better” than Shampoo?
We argue that their relationship parallels Adam's relationship with Signum. Analogous to @lukas_balles and Hennig’s (2018) decomposition of Adam into element-wise scaled Signum, we can decompose Shampoo as left- and right-adapted Muon.
Parameswaran Raman retweeted
Why do gradients increase near the end of training?
Read the paper to find out!
We also propose a simple fix to AdamW that keeps gradient norms better behaved throughout training.
arxiv.org/abs/2506.02285
Parameswaran Raman retweeted
wow. The new model from @LumaLabsAI extending images into videos is really something else. I understood intuitively that this would become possible very soon, but it's still something else to see it and think through future iterations of.
A few more examples around, e.g. the girl in front of the house on fire
teapawt.net/CharaspowerAI/status/1…
I've been wanting to get to the bottom of this story for so long. #DreamMachine #LumaAI 😂
Parameswaran Raman retweeted
I absolutely love this education demo of @OpenAI. Let's make it available in all languages, and personalised to each country.
For example, we could do Spanish for kids and teenagers in Bolivia, giving them access to a technical education that otherwise would not be available to them.
Voice opens up the opportunity to do this even with old not-smart phones, e.g. a farmer in Ghana could start a conversation to get assistance on how to run a farm more efficiently and make more money.
There's a great opportunity here to help people in the entire world, and the AI community should embrace it.
Parameswaran Raman retweeted
In arxiv.org/abs/2404.10575, we propose an MCMC sampler for contrastive learning to look for negative samples - works especially well with small batch size and we showed stationary point convergence.
If you are at #AISTATS2024, checkout our work "Krylov cubic regularized Newton: A subspace second-order method with dimension-free convergence rate" where we present a novel subspace method that converges fast by selecting a subspace with a handful of dimensions (size m <<< d ).
Parameswaran Raman retweeted
Congrats to @AIatMeta on Llama 3 release!! 🎉
ai.meta.com/blog/meta-llama-…
Notes:
Releasing 8B and 70B (both base and finetuned) models, strong-performing in their model class (but we'll see when the rankings come in @ @lmsysorg :))
400B is still training, but already encroaching GPT-4 territory (e.g. 84.8 MMLU vs. 86.5 4Turbo).
Tokenizer: number of tokens was 4X'd from 32K (Llama 2) -> 128K (Llama 3). With more tokens you can compress sequences more in length, cites 15% fewer tokens, and see better downstream performance.
Architecture: no major changes from the Llama 2. In Llama 2 only the bigger models used Grouped Query Attention (GQA), but now all models do, including the smallest 8B model. This is a parameter sharing scheme for the keys/values in the Attention, which reduces the size of the KV cache during inference. This is a good, welcome, complexity reducing fix and optimization.
Sequence length: the maximum number of tokens in the context window was bumped up to 8192 from 4096 (Llama 2) and 2048 (Llama 1). This bump is welcome, but quite small w.r.t. modern standards (e.g. GPT-4 is 128K) and I think many people were hoping for more on this axis. May come as a finetune later (?).
Training data. Llama 2 was trained on 2 trillion tokens, Llama 3 was bumped to 15T training dataset, including a lot of attention that went to quality, 4X more code tokens, and 5% non-en tokens over 30 languages. (5% is fairly low w.r.t. non-en:en mix, so certainly this is a mostly English model, but it's quite nice that it is > 0).
Scaling laws. Very notably, 15T is a very very large dataset to train with for a model as "small" as 8B parameters, and this is not normally done and is new and very welcome. The Chinchilla "compute optimal" point for an 8B model would be train it for ~200B tokens. (if you were only interested to get the most "bang-for-the-buck" w.r.t. model performance at that size). So this is training ~75X beyond that point, which is unusual but personally, I think extremely welcome. Because we all get a very capable model that is very small, easy to work with and inference. Meta mentions that even at this point, the model doesn't seem to be "converging" in a standard sense. In other words, the LLMs we work with all the time are significantly undertrained by a factor of maybe 100-1000X or more, nowhere near their point of convergence. Actually, I really hope people carry forward the trend and start training and releasing even more long-trained, even smaller models.
Systems. Llama 3 is cited as trained with 16K GPUs at observed throughput of 400 TFLOPS. It's not mentioned but I'm assuming these are H100s at fp16, which clock in at 1,979 TFLOPS in NVIDIA marketing materials. But we all know their tiny asterisk (*with sparsity) is doing a lot of work, and really you want to divide this number by 2 to get the real TFLOPS of ~990. Why is sparsity counting as FLOPS? Anyway, focus Andrej. So 400/990 ~= 40% utilization, not too bad at all across that many GPUs! A lot of really solid engineering is required to get here at that scale.
TLDR: Super welcome, Llama 3 is a very capable looking model release from Meta. Sticking to fundamentals, spending a lot of quality time on solid systems and data work, exploring the limits of long-training models. Also very excited for the 400B model, which could be the first GPT-4 grade open source release. I think many people will ask for more context length.
Personal ask: I think I'm not alone to say that I'd also love much smaller models than 8B, for educational work, and for (unit) testing, and maybe for embedded applications etc. Ideally at ~100M and ~1B scale.
Talk to it at meta.ai
Integration with github.com/pytorch/torchtune