Skip to main content

Module etalon_tensor_bridge

Module etalon_tensor_bridge 

Source
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§

AttentionParams
Legacy attention configuration. Use FusedAttentionParams for GPU uniform bindings.
CudaMegaPassExecutor
Scalar transformer reference executor with a resident KV cache.
DirectMlComputeLane
DirectML / DX12 compute executor lane supporting both tiled host execution and Etalon GPU compute pipeline dispatch.
DmlGemmOp
DirectML 2D matrix multiplication operator ($C = A_{m \times k} \times B_{k \times n}$).
DmlGemmParams
ForwardArena
Unified forward execution arena on an Etalon device.
FusedAttentionPipeline
GPU compute pipeline for fused FlashAttention dispatch.
LogitMaskHeader
80-byte std140-aligned logit mask header containing the 512-bit sieve bitmask and autoregressive sampling hyper-parameters.
PagedKvBlockTable
TensorView
TopkParams
Uniform parameters matching the WGSL Params layout in topk_reduction.wgsl.
TopkReductionPipeline
Complete GPU compute pipeline for block top-k / argmax reduction.

Enums§

ForwardMode
Operational mode of the forward execution arena.
KvCacheLayout
QuantScheme

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.

Type Aliases§

CapturedCudaMegaPass
DmlLane