Open akrzgc opened 1 year ago
Looks great! Minor comments:
(1) some stuff wrt naming and hinting on linear algebra stuff (likely not too useful) and
(2) potential performance issue with at[i].set(.)
vs jax.lax.dynamic_update_slice(.)
(could you try to run e.g. benzene on IPU with both, if the cycle count is +-5% my concerns are misplaced).
We need a before/after graph or table to show the effect of the change, all other things being fixed
We need a before/after graph or table to show the effect of the change, all other things being fixed
Do you mean before/after of ipu_eigh
input?
Should I attach it to this review or would you like it to be plotted each time nanoDFT.py
is run?
The change:
ipu_eigh
initializationipu_eigh_its
parameter so that we can now select the number ofipu_eigh
iterations from CLI--visualise_eigvects true