[Perf] Fix and reapply move apply w8a8 block fp8 linear to class

[ElizaWszola]

This PR resolves the issue (#25623) raised in #25607 by using a single padding-cutlass custom op on Hopper.

We are waiting to merge this after v0.11.0 release out of an abundance of caution.

[ProExpertProg]

@ElizaWszola now that cutlass 4.2 has landed, can you remove the hopper padding in a follow-up?

[mxz297]

This PR seems to break AMD.

@HAIAI what's the status of AMD CI so that we can catch such breakage?

[ProExpertProg]

@mxz297 can you create an issue with repro steps?

[zou3519]

I have a separate report of this breaking AMD. What happened in my case is that Dynamo is putting a namedtuple into the graph, which is causing AOTAutograd cache to bypass, which breaks vLLM's precompilation.

I don't have a repro, will find someone who does. The root issue would require a change to PyTorch to fix. Assuming we can repro this and this is indeed the issue, should we consider reverting this PR?

[bradleyhd]

@ProExpertProg FYI I have been attempting an OSS repro but so far it has been elusive. Our internal repro points to the issue described above with dynamo

[ProExpertProg]

@zou3519 we need this PR, we should try to just avoid the named tuple instead, perhaps by wrapping into a custom op at a lower level. Do you know where the named tuple is used/created?

[zou3519]

This is the namedtuple in question:

vllm/vllm/model_executor/layers/quantization/utils/quant_utils.py

Line 28 in b82f430

class GroupShape(_GroupShape):

[ProExpertProg]

I see - I thought that we converted it to list[int] before passing it to custom ops. Or is that not enough?

[ElizaWszola]

Worst case scenario, it should be relatively easy to replace all GroupShape objects with list[int] without having to revert the whole PR

[roywei]

Hi @ElizaWszola this is affection our AMD production inference, while we work on improving OSS AMD CI, can we try your proposed list[int] solution for now?

[bradleyhd]

Seems #25270 is related? the custom nametuple doesn't play nice with torch.compile...

[ProExpertProg]

@bradleyhd #25270 was only an issue if a custom op signature takes a named tuple as a param. Can you (or @roywei or @zou3519) post more info (repro stack trace), either here or in a new issue? We want to avoid ripping out GroupShape more than is necessary

[Lucaskabela]

I believe I have a fix - not sure if we have an open issue or not but here is the PR (@bradleyhd verified patching this fixes his failure) - #26535