feat: DeepSeek MoE v1

[llllvvuu]

DeepSeek is still releasing models on their older MoE architecture.

mlx_lm.convert --hf-path deepseek-ai/DeepSeek-Prover-V1.5-RL --mlx-path DeepSeek-Prover-V1.5-RL-8bit --q-bits 8 -q
mlx_lm.generate --model DeepSeek-Prover-V1.5-RL-8bit --ignore-chat-template --max-tokens 512 --prompt 'import Mathlib
import Aesop

set_option maxHeartbeats 0

open BigOperators Real Nat Topology Rat

/-- The second and fourth terms of a geometric sequence are $2$ and $6$. Which of the following is a possible first term?
Show that it is $\frac{2\sqrt{3}}{3}$.-/
theorem amc12b_2003_p6 (a r : ℝ) (u : ℕ → ℝ) (h₀ : ∀ k, u k = a * r ^ k) (h₁ : u 1 = 2)
  (h₂ : u 3 = 6) : u 0 = 2 / Real.sqrt 3 ∨ u 0 = -(2 / Real.sqrt 3) := by'

With MoE:

mlx_lm.convert --hf-path deepseek-ai/deepseek-moe-16b-chat --mlx-path deepseek-moe-16b-4bit-chat -q
mlx_lm.generate --model deepseek-moe-16b-4bit-chat --prompt "What is mixture-of-experts?" --max-tokens 512 

[awni]

This looks great! Thanks for the addition!

Also it would be super useful to add the pre-converted models to the MLX Community if you are interested. You can pass for example --upload-repo mlx-community/deepseek-moe-16b-4bit-chat to the convert script and it will auto-upload it.