mamba
mamba copied to clipboard
big scan dim cause cuda error
when i use input with shape [b, c ,l] c>80000 in this function
_out, x, *rest = selective_scan_cuda.fwd(u, delta, A, B, C, D, z, delta_bias, delta_softplus)
cause
and i checked the memory of input is about 1 mb ,so it can't be a oom problem, and all the input located in cuda, so not a location problem.
but with a smaller dim,(about 50000), it works.
so, is there a limitation in input dim of mamba , and how can i use a big dim tensor in mamba?
Limit is probably < 64000.
Limit is probably < 64000.
ok, thanks a lot.
Could be wrong, but I think that this is just a bug. I submitted a PR, which fixes at least one illegal memory access error in Mamba2. I noticed that this happened randomly at high lengths due to using more of HBM, and leaking into system memory.