Paulo Valente

@polvalente

Software Engineer Maintainer of Nx, the Numerical Elixir ecosystem Sponsors go toward further Nx development: github.com/sponsors/polva…

Rio de Janeiro, Brazil

github.com/sponsors/polva…

Replies Reposts Media

Paulo Valente
@polvalente
Jul 7
Handrolled implementation is now at 320 tok/s for the 0.6B-4bit variant, while plain Axon rewrites can get the bumblebee model to 100 tok/s on the non-quantized variant! New Hex version shipping tomorrow.
Paulo Valente
@polvalente
Jul 6
In the past few days I shipped a proper EMLX compiler that leverages MLX tracing compiler, with support to all Nx constructs. This was inspired by the other MLX-based backend. PR below has Qwen3 perf comparison. We can now reach 170+ tok/s! github.com/elixir-nx/emlx…
Paulo Valente
@polvalente
Jul 6
In the past few days I shipped a proper EMLX compiler that leverages MLX tracing compiler, with support to all Nx constructs. This was inspired by the other MLX-based backend. PR below has Qwen3 perf comparison. We can now reach 170+ tok/s!
feat: lower while and runtime calls by polvalente · Pull Request #123 · elixir-nx/emlx
From github.com
Paulo Valente
@polvalente
May 10
EMLX just got a nice boost: github.com/elixir-nx/emlx… This PR introduces the EMLXAxon library which provides a few Axon rewriters based on EMLX Fast callbacks. Local benchmarks show 4x perf on Bumblebee code and 6x if using a port of Qwen3-0.6B-4bit 0.3 soon!
Paulo Valente
@polvalente
Apr 24
github.com/elixir-nx/emlx… Quantization support just got merged! It still needs a bit of validation, but EMLX 0.3 is close!
Paulo Valente
@polvalente
Apr 24
This week EMLX got quite a bit of work done. github.com/elixir-nx/emlx This includes: Updated to MLX 0.31.2, added support for CommandQueues for thread isolation, native LinAlg operations, zero-copy to_binary, and more! 0.3 coming soon, right after quantization ships!
Paulo Valente
@polvalente
Apr 24
This week EMLX got quite a bit of work done. github.com/elixir-nx/emlx This includes: Updated to MLX 0.31.2, added support for CommandQueues for thread isolation, native LinAlg operations, zero-copy to_binary, and more! 0.3 coming soon, right after quantization ships!