Cute Dsl FMHA Params

struct LlmFmhaParams

Everything the dense LLM descriptors need, gathered once per CuteDslFMHARunner::run() call.

Public Members

void const *qPtr = {}

void const *kvPtr = {}

void *oPtr = {}

int32_t const *cuKVSeqLens = {}

int32_t batchSize = {}

int32_t seqLenQ = {}

int32_t numQHeads = {}

int32_t numKVHeads = {}

int32_t headDim = {}

int32_t kvCacheCapacity = {}

int32_t windowSizeLeft = {}

float attentionScale = {}

float scaleQ = {}

float scaleK = {}

float scaleV = {}

float invScaleO = {}

float skipSoftmaxThresholdLog2 = {}

Skip-softmax (BLASST) threshold as log2(lambda); 0.0 = disabled. Only the *_skipsoftmax launchers forward it — the other wrappers have no such argument.

cudaStream_t stream = {}

struct LlmFmhaPagedParams

Everything the paged LLM descriptors need, gathered once per CuteDslFMHARunner::runPaged() call.

Public Members

void const *qPtr = {}

void const *pagedKVPoolPtr = {}

int32_t const *kvCachePageList = {}

void *oPtr = {}

int32_t const *cuQSeqLens = {}

int32_t const *cuKVSeqLens = {}

int32_t const *bidirectionalBlockBegin = {}

int32_t const *bidirectionalBlockEnd = {}

int32_t batchSize = {}

int32_t seqLenQ = {}

int32_t numQHeads = {}

int32_t numKVHeads = {}

int32_t headDim = {}

int32_t numPages = {}

int32_t maxPagesPerSeq = {}

int32_t tokensPerPage = {}

int32_t windowSizeLeft = {}

float attentionScale = {}

float scaleQ = {}

float scaleK = {}

float scaleV = {}

float invScaleO = {}

float skipSoftmaxThresholdLog2 = {}

Skip-softmax (BLASST) threshold as log2(lambda); 0.0 = disabled. Only the *_skipsoftmax_paged launchers forward it — the other wrappers have no such argument.

cudaStream_t stream = {}

struct VitFmhaParams

Everything the ViT descriptors need, gathered once per CuteDslFMHARunner ViT run() call. The ViT AOT variants are plain MHA, hence a single head count.

Public Members

void const *qPtr = {}

void const *kPtr = {}

void const *vPtr = {}

void *oPtr = {}

int32_t const *cuSeqLens = {}

int32_t totalSeqLen = {}

int32_t numHeads = {}

int32_t headDim = {}

int32_t maxSeqLen = {}

int32_t batchSize = {}

float scaleSoftmaxLog2 = {}

float attentionScale = {}

float scaleOutput = {}

cudaStream_t stream = {}

struct FmhaV2LlmParams

Everything the FMHA-v2 dense LLM descriptors need, gathered once per CuteDslFMHAV2Runner::run() call.

Public Members

void const *qPtr = {}

void const *kPtr = {}

void const *vPtr = {}

void *oPtr = {}

int32_t const *cuKVSeqLens = {}

int32_t batchSize = {}

int32_t seqLenQ = {}

int32_t kvSeqLen = {}

int32_t numQHeads = {}

int32_t numKVHeads = {}

int32_t headDim = {}

int32_t windowSizeLeft = {}

float attentionScale = {}

float scaleQ = {}

float scaleK = {}

float scaleV = {}

float invScaleO = {}

cudaStream_t stream = {}

int32_t const *blockBegin = {}

Per-query-row [begin, end) block intervals, packed [B, S_q]. Both null for the plain causal/sliding variants; both set for the bidirectional variants.

int32_t const *blockEnd = {}

struct FmhaV2VitParams

Everything the FMHA-v2 ViT descriptors need, gathered once per CuteDslFMHAV2Runner ViT run() call. Unlike VitFmhaParams these variants are GQA, so Q and KV head counts differ.

Public Members

void const *qPtr = {}

void const *kPtr = {}

void const *vPtr = {}

void *oPtr = {}

int32_t const *cuSeqLens = {}

int32_t totalSeqLen = {}

int32_t numQHeads = {}

int32_t numKVHeads = {}

int32_t headDim = {}

int32_t maxSeqLen = {}

int32_t batchSize = {}

float scaleSoftmaxLog2 = {}

float attentionScale = {}

float scaleOutput = {}

cudaStream_t stream = {}