Concepts

Here is a list of all documented concepts with brief descriptions:

NMila	Mila main API namespace
NDnn
NCompute
RGqaOpConcept	Contract for GroupedQueryAttentionOp: positional forward and backward
RIsAcceleratorMemoryResource	Concept for identifying compute accelerator memory resources
RIsCacheCoherent	Concept for cache-coherent memory resources
RIsDeviceMemoryResource	Concept for identifying device-based memory resources
RIsHighBandwidth	Concept for memory resources with high bandwidth characteristics
RIsHostMemoryResource	Concept for identifying host-accessible memory resources
RLinearOpConcept	Contract for LinearOp: typed forward matmul and backward weight/input gradients
ROperationSupported	True iff a concrete OperationTraits specialization exists for the tuple
ROptimizedForCoalescing	Concept for memory resources optimized for coalesced access patterns
RRequiresContextBinding	Concept for memory resources requiring CUDA context binding
RSamplingOpConcept	Contract for SamplingOp: in-place token sampling from a logits tensor
RSupportsConcurrentKernels	Concept for memory resources supporting concurrent kernel execution
RSupportsPeerAccess	Concept for memory resources optimized for device-to-device transfers
RSupportsSIMD	Concept for CPU memory resources with SIMD support
RSupportsTextureMemory	Concept for memory resources supporting texture memory access
RSupportsThreading	Concept for memory resources with threading support
RSupportsUnifiedMemory	Concept for CUDA memory resources with unified memory support
RUnaryOpConcept	Contract for policy-free unary ops (Softmax, RmsNorm, LayerNorm, Residual, ...)
NQuant
NKvCache
RKvCachePolicy	Minimum contract for all KV cache management policies
RQuantKvPolicy	Concept for quantization-based KV cache compression policies
NWeight
RAttentionPrecisionRoles	Fused QKV projection and the attention output projection
RDecoderPrecisionPlan	What an attention-plus-feed-forward decoder block requires
REmbeddingPrecisionRole	The embedding table, required at the network level rather than the block
RFeedForwardPrecisionRoles	Fused gate+up projection and the down projection of a gated feed-forward network
RHasCodebookTable	The format decodes through a per-tensor table, which travels as its own tensor
RHasHighBitPlane	The format spills one bit per element into a second byte-aligned plane, so the primary tensor's kStorageBitsPerElement does not account for the whole code
RLanguageModelHeadRole	The output (language-model head) table
RWeightQuantPolicy
RDeltaNetPrecisionRoles	Roles a Gated DeltaNet block instantiates, on top of the feed-forward pair
RDeviceOnlyTensorDataType	Concept identifying device-only abstract data types
RHostCompatibleTensorDataType	Concept identifying host-compatible abstract data types
RisValidTensor	Primary tensor configuration validation concept
RPrecisionSupportedOnDevice	Concept to validate precision is supported on a device at compile-time
RTokenSink	Satisfied by any callable accepting a span of decoded tokens
RTokenStreamer	Satisfied by any callable accepting a single decoded token
RValidFloatTensorDataType	Concept constraining abstract data types to floating-point formats
RValidIntegerTensorDataType	Concept constraining abstract data types to integer formats