Skip to content

ModelOpt / FP8

ModelOpt and FP8 checkpoint flows are practical quantization families in the codebase. They are useful when a deployment has validated low-precision transformer components or when online quantization is acceptable for the target model.

Role in Sol-Engine

The reported full optimization stacks focus on selective NVFP4 for Cosmos3-Super and LTX-2.3, but ModelOpt / FP8 remains part of the broader local quantization search space.

Practical knobs

  • transformer override path.
  • transformer weights path.
  • ignored layer patterns.
  • layerwise offload compatibility.
  • BF16 fallback modules.

Validation

Quantized checkpoints should be validated in the same pipeline where they will be composed. A checkpoint that is acceptable alone may become unacceptable after cache or sparse attention is added.

References