Skip to content

Latest commit

 

History

History
74 lines (50 loc) · 2.79 KB

model_customization.md

File metadata and controls

74 lines (50 loc) · 2.79 KB
Raw

Model Customization

By default, torchchat (and PyTorch) defaults to unquantized eager execution.

This page goes over the different options torchchat provides for customizing the model execution for inference.

  • Device
  • Compilation
  • Model Precision
  • Quantization

Device

python3 (chat | generate | browser | server | export | eval) --device [cpu | cuda | mps] ...

To leverage a specific accelerator, the target device can be set.

By default, torchchat defaults to the fastest executor available in the system, chosen in this order: cuda, mps, and cpu.

Compilation: JIT-compiled execution

python3 (chat | generate | browser | server | eval) [--compile][--compile_prefill] ...

To improve performance, you can compile the model with --compile; trading off the time to first token processed with time per token.

To improve performance further, at the cost of increased compile time, you may also compile the prefill with --compile_prefill.

To learn more about compilation, check out: https://pytorch.org/get-started/pytorch-2.0/

For CPU, you can use --max-autotune to further improve the performance with --compile and compile-prefill.

See max-autotune on CPU tutorial.

Model Precision

python3 (chat | generate | browser | server | export | eval) --dtype [fast | fast16 | bf16 | fp16 | fp32] ...

To reduce the memory bandwidth requirement and to take advantage of higher density compute available, the model can use lower precision floating point representations. For example, many GPUs and some of the CPUs have good support for bfloat16 and float16.

Unlike gpt-fast which uses bfloat16 as default, torchchat uses the dtype "fast16". This picks the best performing 16-bit floating point type available (for execution with Executorch, macOS/ARM and Linux/x86 platforms). For example on macOS, support depends on the OS version, with versions starting with 14.0 supporting bfloat16 as support, and float16 for earlier OS version based on system support for these data types.

The "fast" data type is also provided as a virtual data type that defaults to the best floating point data type available on the selected device. Currently, this behaves the same as "fast16", but with "fp32" when exporting to ExecuTorch.

Quantization

python3 (chat | generate | browser | server | export | eval) [--quantize] <quant.json> ...

To further minimize memory requirements, accelerate inference speeds, and decrease power consumption the model can also be quantized. Torchchat leverages torchao for quantization.

See the quantization guide for examples and more details.