Post
4505
Kimi K2 tech report is full of gems as always. Here are my notes on it:
> MuonClip: Pretty crazy how after 70k the training stabilizes and the QK-clip is basically inactive. There is also no loss in perf with QK-clip which is not trivial at all (at small scale but with aggressive threshold). Also a cool explanation of why muon makes the logit explode in appendix E (tl;dr is that muon makes the singular value of the update matrix higher)
> Sparsity scaling laws to justify their ratio, they have a very solid training infra that allows the model to be trained at this sparsity level, they could have increased even more but as sparsity increases the training becomes less efficient.
> They diminish the number of attention heads to make it more efficient for long context since attention heads are a big bottleneck for long context. They also remove 2 of the 3 "first dense" layers in the dsv3 arch.
With the sparsity and attention heads (divided by 2) they achieve 83% increased flops compared to deepseek v3 arch at 128k.
> Data: Rephrasing is KEY. They do a lot more synthetic data generation and rephrase their corpus to have different styles, for longer documents they do it by chunk. I'm (half) surprised by the fact that ONLY 1 epoch (assuming same number of training tokens I think?) of data rephrased 10 times has better accuracy than 10 epochs of the same data rephrased once.
> They do rewriting for Math and Knowledge, for Math they apply the ShallowMath recipe and instruct the model to rephrase in a "learning note" style
> They talk about diversity and probably have some internal stuff/eval to test that, as always still a bit unclear for me how to properly measure that.
The infra is also very nice, quick summary:
> PP=16 (1F1B schedule, a bit custom), EP=16, zero1
> No FP8 computation but for storage of specific layers, selective recomputation for inexpensive block, activation offloading to CPU
> MuonClip: Pretty crazy how after 70k the training stabilizes and the QK-clip is basically inactive. There is also no loss in perf with QK-clip which is not trivial at all (at small scale but with aggressive threshold). Also a cool explanation of why muon makes the logit explode in appendix E (tl;dr is that muon makes the singular value of the update matrix higher)
> Sparsity scaling laws to justify their ratio, they have a very solid training infra that allows the model to be trained at this sparsity level, they could have increased even more but as sparsity increases the training becomes less efficient.
> They diminish the number of attention heads to make it more efficient for long context since attention heads are a big bottleneck for long context. They also remove 2 of the 3 "first dense" layers in the dsv3 arch.
With the sparsity and attention heads (divided by 2) they achieve 83% increased flops compared to deepseek v3 arch at 128k.
> Data: Rephrasing is KEY. They do a lot more synthetic data generation and rephrase their corpus to have different styles, for longer documents they do it by chunk. I'm (half) surprised by the fact that ONLY 1 epoch (assuming same number of training tokens I think?) of data rephrased 10 times has better accuracy than 10 epochs of the same data rephrased once.
> They do rewriting for Math and Knowledge, for Math they apply the ShallowMath recipe and instruct the model to rephrase in a "learning note" style
> They talk about diversity and probably have some internal stuff/eval to test that, as always still a bit unclear for me how to properly measure that.
The infra is also very nice, quick summary:
> PP=16 (1F1B schedule, a bit custom), EP=16, zero1
> No FP8 computation but for storage of specific layers, selective recomputation for inexpensive block, activation offloading to CPU
