Timer#

class TimerSession#

RAII timer session for automatic cleanup.

Automatically stops timing when session goes out of scope.

Public Functions

inline TimerSession(std::function<void()> onEnd) noexcept#

Construct active session with callback.

Parameters:

onEnd – Callback to execute on destruction

inline TimerSession(std::nullptr_t) noexcept#

Construct inactive session.

inline ~TimerSession() noexcept#

Destructor - executes callback if active.

TimerSession(TimerSession const&) = delete#

Deleted copy constructor.

TimerSession &operator=(TimerSession const&) = delete#

Deleted copy assignment.

inline TimerSession(TimerSession &&other) noexcept#

Move constructor.

inline TimerSession &operator=(TimerSession &&other) noexcept#

Move assignment operator.

class Timer#

Preallocated event pool. Longer-lived profiling can grow without waiting for earlier stages.

Public Functions

void reserve(size_t maxSamples, size_t maxSequences = 1)#
inline size_t capacity() const noexcept#

Number of timing slots whose CUDA events have been created.

inline void setGpuTimingEnabled(bool enabled) noexcept#

When disabled, profiling keeps host-side acceptance records but creates no CUDA events or stage samples.

void recordAcceptance(
int32_t sequenceIndex,
int32_t proposed,
int32_t accepted,
int32_t output,
int32_t bonus,
int32_t baseLength,
bool provenanceExact = false,
bool proposedExact = false
)#
inline std::vector<AcceptanceSample> const &getAcceptanceSamples(
) const noexcept#
void reset() noexcept#
void beginRequest()#
void setRound(int32_t round) noexcept#
void setPastLengths(std::vector<int32_t> const &lengths)#

Host-known past lengths for the next forward; empty means unavailable without device readback.

TimerSession startStage(
std::string const &stageId,
cudaStream_t stream
)#
TimerSession startForward(
std::string const &modelRole,
int64_t batch,
int64_t tokens,
int64_t selected,
cudaStream_t stream
)#
std::optional<StageTimingData> getTimingData(
std::string const &stageId
) const#
std::unordered_map<std::string, StageTimingData> const &getAllTimingData(
) const#
std::vector<TimingSample> const &getSamples() const#
class TimerProxy#

Keeps legacy profiling call sites attached to the current runtime’s event storage.

Public Functions

inline void select(std::shared_ptr<Timer> timer)#
inline void reserve(size_t size, size_t maxSequences = 1)#
inline size_t capacity() const noexcept#
inline void recordAcceptance(
int32_t sequenceIndex,
int32_t proposed,
int32_t accepted,
int32_t output,
int32_t bonus,
int32_t baseLength,
bool provenanceExact = false,
bool proposedExact = false
)#
inline std::vector<AcceptanceSample> const &getAcceptanceSamples(
) const noexcept#
inline void reset() noexcept#
inline void beginRequest()#
inline void setRound(int32_t round) noexcept#
inline void setPastLengths(std::vector<int32_t> const &lengths)#
inline TimerSession startStage(
std::string const &stage,
cudaStream_t stream
)#
inline TimerSession startForward(
std::string const &modelRole,
int64_t batch,
int64_t tokens,
int64_t selected,
cudaStream_t stream
)#
inline std::optional<StageTimingData> getTimingData(
std::string const &stage
) const#
inline std::unordered_map<std::string, StageTimingData> const &getAllTimingData(
) const#
inline std::vector<TimingSample> const &getSamples() const#
struct TimerPair#

CUDA event pair for timing.

Lazy-initialized timer pair using CUDA events.

Public Functions

inline void initialize()#

Initialize CUDA events.

Throws:

std::runtime_error – if CUDA event creation fails

inline ~TimerPair() noexcept#

Destructor - destroys CUDA events.

TimerPair() = default#

Default constructor.

TimerPair(TimerPair const&) = delete#

Deleted copy constructor.

TimerPair &operator=(TimerPair const&) = delete#

Deleted copy assignment.

inline TimerPair(TimerPair &&other) noexcept#

Move constructor.

inline TimerPair &operator=(TimerPair &&other) noexcept#

Move assignment operator.

Public Members

cudaEvent_t gpuStart = {nullptr}#

Start event.

cudaEvent_t gpuEnd = {nullptr}#

End event.

bool hasStarted = {false}#

Whether timing has started.

bool isInitialized = {false}#

Whether events are initialized.

struct StageTimingData#

Stage timing data.

Stores raw timing measurements and calculates derived values on-demand.

Public Functions

inline void addTiming(float timeMs)#

Add timing measurement.

Parameters:

timeMs – Time in milliseconds

inline void reset() noexcept#

Reset all timing data.

inline float getTotalGpuTimeMs() const noexcept#

Calculate total GPU time.

Returns:

Total time in milliseconds

inline float getAverageTimeMs() const noexcept#

Calculate average time per run.

Returns:

Average time in milliseconds

inline int64_t getTotalRuns() const noexcept#

Get total number of runs.

Returns:

Run count

Public Members

std::vector<float> gpuTimesMs#

GPU time measurements in milliseconds.

struct TimingSample#

One completed invocation; model forwards are children of their semantic stage.

Public Members

std::string stageId#
std::string parentStage#
uint64_t requestId = {0}#
int32_t round = {-1}#
int64_t batchSize = {0}#
int64_t queryTokens = {0}#
int64_t selectedTokens = {0}#
std::vector<int32_t> pastLengths#
bool forward = {false}#
int32_t forwardIndex = {-1}#

Forward ordinal within this request and round; -1 for stages.

float gpuTimeMs = {0}#
struct AcceptanceSample#

Host-visible acceptance facts collected at the production token readback boundary.

Public Members

uint64_t requestId#
int32_t round#
int32_t sequenceIndex#
int32_t proposedDrafts#
int32_t acceptedDrafts#
int32_t outputTokens#
int32_t bonusTokens#
int32_t baseMaterializedLength#
bool provenanceExact#
bool proposedExact#

False when proposedDrafts is a verify-tree capacity, not valid proposal nodes.

bool trt_edgellm::getProfilingEnabled() noexcept#

Thread-local profiling control flag Controls aggregate counters and GPU stage timing; per-request collection is opt-in. This is useful to exclude warmup runs from benchmark statistics

void trt_edgellm::setProfilingEnabled(bool enabled) noexcept#