Ssd Varlen Metadata#
- void trt_edgellm::mamba::fillUniformValidLens(
- int32_t *d_valid_lens,
- int32_t batch,
- int32_t seq_len,
- cudaStream_t stream
Fill
d_valid_lens[0..batch)with uniformseq_lenvalue. Used by the runner when caller’scontext_lengthsis null (uniform batch); the kernel’s padded_mode end-of-seq clamp still requires a valid_lens tensor.
- void trt_edgellm::mamba::buildSSDPackedVarlenMetadata(
- int32_t *d_seq_idx,
- int32_t *d_chunk_indices,
- int32_t *d_chunk_offsets,
- int32_t *d_seq_chunk_cumsum,
- int32_t const *d_cu_seqlens,
- int32_t num_seqs,
- int32_t total_tokens,
- int32_t chunk_size,
- int32_t total_slots,
- cudaStream_t stream
Packed counterpart of buildSSDVarlenMetadata: sequence b owns tokens
[d_cu_seqlens[b], d_cu_seqlens[b + 1])of one flat axis oftotal_tokenstokens.total_slotsmust be at leastssdPackedLogicalChunkUpperBound(num_seqs, total_tokens, chunk_size); offsets are clamped to the token axis and writes stay withintotal_slots. Single-block kernel:num_seqs <= 1024.
- void trt_edgellm::mamba::fillValidLensFromCuSeqlens(
- int32_t *d_valid_lens,
- int32_t const *d_cu_seqlens,
- int32_t num_seqs,
- int32_t total_tokens,
- cudaStream_t stream
Fill
d_valid_lens[b]with the length of packed sequence b, clamped to the token axis like the chunk metadata.
- void trt_edgellm::mamba::buildSSDVarlenMetadata(
- int32_t *d_seq_idx,
- int32_t *d_chunk_indices,
- int32_t *d_chunk_offsets,
- int32_t *d_seq_chunk_cumsum,
- int32_t const *d_context_lengths,
- int32_t batch,
- int32_t seq_len,
- int32_t chunk_size,
- cudaStream_t stream
Build varlen metadata fully on-device (CUDA-graph-compatible – no host sync). The logical chunk upper bound is
batch * (ceil(seq_len / chunk_size) + maybe_unaligned_chunk); trailing slack in chunk_indices is filled with sentinel -1 so the kernel’schunk_indices[physical_chunk+1]lookup is safe up to the upper bound.