Sync quant dialect changes 20x #104
Conversation
Full revamp of the 'quant' dialect. This is an implementation for the RFC at https://discourse.llvm.org/t/rfc-improvements-in-the-quant-dialect/79942
|
@nikita-kud can you please provide feedback on this new change? |
There was a problem hiding this comment.
@nikita-kud can you please provide feedback on this new change?
In fact, I've already done that in Teams..
Zoran:
What we propose is for us to firstly bring the bigger quant dialect refactoring https://github.com/llvm/llvm-project/pull/100667/files into our LLVM fork https://jira.devtools.intel.com/browse/EISW-159244
and after that it we'll apply the sub-channel quantization changes as it s done on the public llvm repo.
The problem for me is that this refactoring is not part of LLVM19. So you have to merge it twice, for LLVM18 and then move this patch to LLVM19. Unless we skip version 19 and move to 20 and we discuss this option above:
Gelb:
Well I agree, more commits = more difficult to track down to a specific one that breaks something
Zoran:
This will simplify the llvm version update wrt to the quant dialect;
So, on the one hand, I agree, on the other hand, I believe we should first upgrade LLVM to version 19.
Summary
Upon the
Sub-channel quantized typePull Request being blocked, we decided that synchronizing thequant dialectchanges from llvm-project upstream can reduce the effort for the upcoming LLVM update to our fork. The changes introduced are cherry-picked from llvm/llvm-project#100667 and as an addition custom npu patches have been applied.With this early integration, we aim to have that Pull Request unblocked and continue the integration of the new data type to our fork.
JIRA ticket
EISW-159244
Related PR in NPU Compiler and/or OpenVINO repository with sub-module update
Other related tickets