Observation.ixx File Reference

Vocabulary for activation observation: the compute pass, the stage, the observer. More...

#include <cstdint>
#include <functional>
#include <initializer_list>
#include <string>
#include <string_view>
import Dnn.ITensor;

Classes
class	Mila::Dnn::Compute::ComputePassMask
	A set of compute passes, small enough to test on the publication path. More...
struct	Mila::Dnn::Compute::ObservableStage
	A stage a component publishes, and the passes it is published on. More...

Namespaces
namespace	Mila
	Mila main API namespace.

Typedefs
using	Mila::Dnn::Compute::ActivationObserver
	Receives one published activation.

Enumerations
enum class	Mila::Dnn::Compute::ComputePass : uint8_t { Forward = 0 , Prefill = 1 , Decode = 2 , Backward = 3 }
	The compute entry point that produced an observed value. More...

Functions
constexpr std::string_view	Mila::Dnn::Compute::computePassToString (ComputePass pass) noexcept

Detailed Description

Vocabulary for activation observation: the compute pass, the stage, the observer.

Design of record is Specifications/Observability.md. These types are shared by Component (which publishes) and IExecutionContext (which transports).

Typedef Documentation

ActivationObserver

using Mila::Dnn::Compute::ActivationObserver

export

Initial value:

std::function<void(
       std::string_view path,
       ComputePass pass,
       std::string_view stage,
       const ITensor& value )>

Receives one published activation.

The tensor is BORROWED and is valid only for the duration of this call. It is also ordered on the publishing component's stream rather than valid on the host: publication never synchronizes, because synchronizing is the clearest way for a probe to change what it observes. An observer needing host-readable values calls IExecutionContext::synchronize() itself, or issues its own copy on that stream.

Enumeration Type Documentation

ComputePass

enum class Mila::Dnn::Compute::ComputePass : uint8_t

exportstrong

The compute entry point that produced an observed value.

Declared complete while only the inference passes are implemented, so that adding Backward later is additive rather than a change to the observer signature.