chunk scan
A GPU kernel mechanism in Mamba2 that parallelizes the serial prefix scan, using a two-stage strategy of intra-chunk parallelism and inter-chunk serialism to enable efficient parallel computation over long sequences. It is highly sensitive to accumulator initialization and boundary synchronization.
Timeline (last 90 days)
Mamba2's chunk scan mechanism adopts a two-stage strategy of intra-chunk parallelism and inter-chunk serialism, reducing the degree of parallelism from the sequence-length level to the chunk-count level.
NaN is contagious: the result of any arithmetic operation with NaN is NaN; in the cross-chunk reduction of chunk scan, a NaN produced by one chunk will propagate to all subsequent chunks.
All Facts (2)
Mamba2's chunk scan mechanism adopts a two-stage strategy of intra-chunk parallelism and inter-chunk serialism, reducing the degree of parallelism from the sequence-length level to the chunk-count level.
50%UnverifiedNaN is contagious: the result of any arithmetic operation with NaN is NaN; in the cross-chunk reduction of chunk scan, a NaN produced by one chunk will propagate to all subsequent chunks.
50%