Nirvedh Meshram

Surprisingly the error came down to this review comment in the original [PR ](https://github.com/iree-org/iree/pull/22523/changes#r2494669095) I am very surprised by this as this is about `overflow` flag in arith.addi When I...

@zjgarvey might be but this issue is consistent and not intermittent, its happening in a very specific case of strided convs @bjacob Thanks, the issue is triaged to optimization after...

Here is what is causing this to fail to bufferize, after `GPUFuseAndHoistParallelLoopsPass` We have the following access ``` %read_write_input = flow.dispatch.tensor.load ... -> tensor %workgroup_scf_forall = scf.forall ... shared_outs(%arg2 =...

@MaheshRavishankar WDYT of the two suggestions from @hanhanW above. Based on our conversations previously we want to support accumulating GEMMS without transforming them to non-accumulating GEMM + elementwise add, That...

@yzhang93 @Max191 wanted to share this issue with you in case you have plans to make this better already, I think we will need to resolve this one to get...

@yzhang93 we would want it to look something like this ``` scf.for %arg7 = %c0 to %c2048 step %c128 { %23 = affine.apply affine_map (d0 + d1 + d2 *...

Sharing the smallest repro with which I am able to captures this issue ``` func.func @run_forward$async_dispatch_27_attention_2x1178x24x64xf16_generic(%12 : tensor, %13 : tensor, %14 : tensor) -> tensor { %cst = arith.constant...

I think the IR is too dated at this point and there is no `iree_linalg_ext.attention ` we would need to revisit this with fresh IR, @monorimet let me know if...

closing as issue (and the op which had the issue) dont seem to be there anymore.

@krzysz00 we are still relying on the flag https://github.com/iree-org/iree/blob/main/compiler/plugins/target/ROCM/ROCMTarget.cpp#L672-L678 We would have to drop the flag and check if the issue is still reproducible