Expand description
Provider-native NVFP4 matrix products.
TOSA has no FP4 tensor type. Encoding packed weights as an integer TOSA graph would hide the operation from provider admission and explode it into scalar nodes, so Vulkan owns a small artifact format for the exact storage contract used by NVFP4 checkpoints. This is an artifact-format extension, not a virtio-accel wire change.
Structs§
- Nvfp4
Artifact - A validated F32 × NVFP4 projection artifact.
- Nvfp4
MoeArtifact - Two native NVFP4 expert projections with a device-local squared-ReLU intermediate. Each expert occupies one externally bound cache-slot span; the four offsets identify its up/down packed and block-scale planes.
Enums§
- Nvfp4
Activation - Fused projection epilogue. Its integer representation is part of artifact v1.