Expand description
Bridge between etalon-tensor accelerated operators and QualiaDB inference.
Exposes unified forward execution arenas, FlashAttention compute pipelines, in-shader top-k reduction, and DirectML / CUDA tensor execution lanes.
Structs§
- Attention
Params - Legacy attention configuration. Use
FusedAttentionParamsfor GPU uniform bindings. - Cuda
Mega Pass Executor - Scalar transformer reference executor with a resident KV cache.
- Direct
MlCompute Lane - DirectML / DX12 compute executor lane supporting both tiled host execution and Etalon GPU compute pipeline dispatch.
- DmlGemm
Op - DirectML 2D matrix multiplication operator ($C = A_{m \times k} \times B_{k \times n}$).
- DmlGemm
Params - Forward
Arena - Unified forward execution arena on an Etalon device.
- Fused
Attention Pipeline - GPU compute pipeline for fused FlashAttention dispatch.
- Logit
Mask Header - 80-byte std140-aligned logit mask header containing the 512-bit sieve bitmask and autoregressive sampling hyper-parameters.
- Paged
KvBlock Table - Tensor
View - Topk
Params - Uniform parameters matching the WGSL
Paramslayout intopk_reduction.wgsl. - Topk
Reduction Pipeline - Complete GPU compute pipeline for block top-k / argmax reduction.
Enums§
- Forward
Mode - Operational mode of the forward execution arena.
- KvCache
Layout - Quant
Scheme
Constants§
- FUSED_
ATTENTION_ WGSL - Raw WGSL shader source for the fused FlashAttention kernel with guarded all-masked row handling.
- TOPK_
REDUCTION_ WGSL - Raw WGSL shader source for the GPU block top-k reduction kernel.
Functions§
- build_
forward_ arena_ on - Construct a unified forward execution arena on the given GPU device.
- build_
fused_ attention_ on - Construct a compiled fused FlashAttention compute pipeline on the given device.
- build_
topk_ pipeline_ on - Construct a compiled GPU block top-k reduction compute pipeline on the given device.
- dispatch_
fused_ attention - Dispatch fused attention on an active compute pass.
- dispatch_
topk_ reduction - Dispatch top-k block reduction on an active compute pass.
- online_
softmax_ rescale - Numerical online-softmax rescale calculation with guarded all-masked row handling.
- read_
token_ blocking - Blocking readback of a single decoded token from the forward arena staging buffer.
- record_
readback_ copy - Record a copy of token candidates into the forward arena readback stage.
- try_
shared_ forward_ arena - Attempt to construct a unified forward execution arena using the process-wide shared GPU device.
- try_
shared_ fused_ attention - Attempt to construct a fused FlashAttention compute pipeline using the process-wide shared GPU device.
- try_
shared_ topk_ pipeline - Attempt to construct a top-k reduction compute pipeline using the process-wide shared GPU device.
- upload_
attention_ params - Upload attention uniform parameters to a uniform buffer using the device queue.
- upload_
logit_ mask_ header - Upload logit mask header to a uniform/storage buffer using the device queue.
- upload_
topk_ params - Upload top-k uniform parameters to a uniform buffer using the device queue.