Dflash Target KV Cache Update Plugin#

class DFlashTargetKVCacheUpdatePlugin : public nvinfer1::IPluginV3, public nvinfer1::IPluginV3OneCore, public nvinfer1::IPluginV3OneBuildV2, public nvinfer1::IPluginV3OneRuntime#

TensorRT plugin for DFlash target KV cache update (V3 — IPluginV3).

Inputs: 0: k_delta [B, L, numKVHeads, headDim] FP16 1: v_delta [B, L, numKVHeads, headDim] FP16 2: past_key_value Paged KV pool [2, numPages, kTOKENS_PER_PAGE, numKVHeads, headDim] FP16 — the same KVCacheManager allocation as the AttentionPlugin binding (K/V split OUTERMOST), bound unreshaped. maxBatch/capPadded are recovered at enqueue time from numPages (dim 1) and the build-time pages_per_slot attribute (maxBatch = numPages / pagesPerSlot, cap = pagesPerSlot * kTOKENS_PER_PAGE) — see setPagesPerSlot(). 3: rope_cos_sin [ropeBatch, cosSinSeqLen, rotaryDim] FP32 — cosSinSeqLen <= cap 4: delta_start_positions [B] INT32 5: delta_lengths [B] INT32

Outputs: 0: present_key_value same shape/dtype as past_key_value (aliased)

BREAKING ABI NOTE (paged-KV substrate): past_key_value/present_key_value moved to the paged pool layout above and a new serialized pages_per_slot field was added (see setPagesPerSlot()). The plugin version string was deliberately NOT bumped &#8212; this project always regenerates ONNX and rebuilds engines together with the runtime, so an ABI break here is accepted rather than versioned. Any ONNX/engine older than this change must be re-exported and rebuilt; it will not load correctly against this plugin.

Public Functions

DFlashTargetKVCacheUpdatePlugin(std::string const &name)#
DFlashTargetKVCacheUpdatePlugin(
std::string const &name,
nvinfer1::PluginFieldCollection const *fc
)#
DFlashTargetKVCacheUpdatePlugin() = delete#
DFlashTargetKVCacheUpdatePlugin(
DFlashTargetKVCacheUpdatePlugin const&
) = delete#
~DFlashTargetKVCacheUpdatePlugin() override = default#
nvinfer1::IPluginCapability *getCapabilityInterface(
nvinfer1::PluginCapabilityType type
) noexcept override#
nvinfer1::IPluginV3 *clone() noexcept override#
char const *getPluginName() const noexcept override#
char const *getPluginVersion() const noexcept override#
char const *getPluginNamespace() const noexcept override#
int32_t getNbOutputs() const noexcept override#
int32_t getOutputDataTypes(
nvinfer1::DataType *outputTypes,
int32_t nbOutputs,
nvinfer1::DataType const *inputTypes,
int32_t nbInputs
) const noexcept override#
int32_t getOutputShapes(
nvinfer1::DimsExprs const *inputs,
int32_t nbInputs,
nvinfer1::DimsExprs const *shapeInputs,
int32_t nbShapeInputs,
nvinfer1::DimsExprs *outputs,
int32_t nbOutputs,
nvinfer1::IExprBuilder &exprBuilder
) noexcept override#
bool supportsFormatCombination(
int32_t pos,
nvinfer1::DynamicPluginTensorDesc const *inOut,
int32_t nbInputs,
int32_t nbOutputs
) noexcept override#
int32_t configurePlugin(
nvinfer1::DynamicPluginTensorDesc const *in,
int32_t nbInputs,
nvinfer1::DynamicPluginTensorDesc const *out,
int32_t nbOutputs
) noexcept override#
size_t getWorkspaceSize(
nvinfer1::DynamicPluginTensorDesc const *inputs,
int32_t nbInputs,
nvinfer1::DynamicPluginTensorDesc const *outputs,
int32_t nbOutputs
) const noexcept override#
int32_t getAliasedInput(int32_t outputIndex) noexcept override#
int32_t enqueue(
nvinfer1::PluginTensorDesc const *inputDesc,
nvinfer1::PluginTensorDesc const *outputDesc,
void const *const *inputs,
void *const *outputs,
void *workspace,
cudaStream_t stream
) noexcept override#
int32_t onShapeChange(
nvinfer1::PluginTensorDesc const *in,
int32_t nbInputs,
nvinfer1::PluginTensorDesc const *out,
int32_t nbOutputs
) noexcept override#
nvinfer1::IPluginV3 *attachToContext(
nvinfer1::IPluginResourceContext *context
) noexcept override#
nvinfer1::PluginFieldCollection const *getFieldsToSerialize(
) noexcept override#
void setPluginNamespace(char const *pluginNamespace) noexcept#
void setPagesPerSlot(int32_t pagesPerSlot) noexcept#

Set the pool’s pages-per-slot (capPadded / kTOKENS_PER_PAGE), a build-time-only fact not derivable from the pool-shaped past_key_value binding alone (dim 1 there is numPages = maxBatch * pagesPerSlot). Called by the builder (llmBuilder.cpp) right after ONNX parsing, from mBuilderConfig.maxKVCacheCapacity; propagated through clone() and engine serialization (getFieldsToSerialize) so the deserialized runtime plugin has it too.

class DFlashTargetKVCacheUpdatePluginCreator : public nvinfer1::IPluginCreatorV3One#

Public Functions

DFlashTargetKVCacheUpdatePluginCreator()#
~DFlashTargetKVCacheUpdatePluginCreator() override = default#
char const *getPluginName() const noexcept override#
char const *getPluginVersion() const noexcept override#
nvinfer1::PluginFieldCollection const *getFieldNames(
) noexcept override#
char const *getPluginNamespace() const noexcept override#
void setPluginNamespace(char const *pluginNamespace) noexcept#
nvinfer1::IPluginV3 *createPlugin(
char const *name,
nvinfer1::PluginFieldCollection const *fc,
nvinfer1::TensorRTPhase phase
) noexcept override#