Fix reuse of pipelines including IterBlock

[russellb]

When running sdg.generate() more than once, a pipeline including an
IterBlock would fail because the block_kwargs were removed the
first time the block was used. This change removes the special
handling for IterBlock because it does not appear to be necessary.

The new code ends up calling the block's constructor with all of the
same arguments, but now block_kwargs doesn't disappear.

Signed-off-by: Russell Bryant [email protected]

[markmc]

You've also fixed the fact that we were instantiating the blocks twice:

     block = block_type(**block_config)
     if block_type == IterBlock:
         ...
     else:
         block = block_type(**block_config)

I did wonder about this special handling of IterBlock, but it's all confusing enough I couldn't quickly convince myself that it wasn't needed. I sort of looks like a leftover from an earlier implementation where the IterBlock's config just got passed down to the LLMBlock with block_kwargs added? Or something

Anyhow, lgtm 👍

[markmc]

Some special handling of IterBlock turns out to be needed in #86 ... so #86 should replace this PR, at least how it's currently implemented

[russellb]

Some special handling of IterBlock turns out to be needed in #86 ... so #86 should replace this PR, at least how it's currently implemented

marked this as a draft while we work through #86. This will help remind me to come back around and check if it's still relevant later.

[markmc]

IterBlock was removed by #105