Praveen G

_Initial analysis_ As part of the DeduplicateExecutables pass, the following dispatch gets changed as follows ``` %1520 = flow.dispatch @prefill_bs8$async_dispatch_805::@prefill_bs8$async_dispatch_805_matmul_like_Dx131072x5120_f16xf16xf32[%1519, %13](%1519, %1518, %__hoisted_tensor_131072x5120xf16_578, %13) : (index, tensor

I have a smaller reproducer for the above issue. Thanks @pashu123 for helping in debugging this. ``` module { func.func @prefill_bs2(%arg0: !torch.vtensor, %arg1: !torch.vtensor) -> !torch.vtensor attributes {torch.assume_strict_symbolic_shapes} { %965...

@MaheshRavishankar @qedawkins I was not able to get a repro for IREE at this moment, but was able to identify that the issue is due to **sharedMemBytes** being set to...

MLIR files [attention_and_matmul_spec_gfx942.txt](https://github.com/user-attachments/files/20617477/attention_and_matmul_spec_gfx942.txt) [flux_vae_bs1_1024x1024_bf16.txt](https://github.com/user-attachments/files/20617478/flux_vae_bs1_1024x1024_bf16.txt) [flux_t5xxl_bs1_512_bf16.txt](https://github.com/user-attachments/files/20617480/flux_t5xxl_bs1_512_bf16.txt) [flux_schnell_sampler_bs1_512_1024x1024_bf16.txt](https://github.com/user-attachments/files/20617479/flux_schnell_sampler_bs1_512_1024x1024_bf16.txt) [flux_dev_sampler_bs1_512_1024x1024_bf16.txt](https://github.com/user-attachments/files/20617476/flux_dev_sampler_bs1_512_1024x1024_bf16.txt) [flux_clip_bs1_77_bf16.txt](https://github.com/user-attachments/files/20617481/flux_clip_bs1_77_bf16.txt)

@MaheshRavishankar Had a call with @Muzammiluddin-Syed-ECE to reproduce the error using shortfin. I was verifying compiling the above mlir files and running using the actual inputs with and without the...

Removing the field **block_shared_memory_size:uint32 (deprecated);** from ~/iree/runtime/src/iree/schemas/hip_executable_def.fbs with revert being undone, It fails in shortfin with the following error ``` ValueError: iree/runtime/src/iree/hal/drivers/hip/native_executable.c:109: INVALID_ARGUMENT; flatbuffer verification failed: vtable offset out of...

> If it still happens after recompile, that means there is something stale (?). I would try printf debugging for the culprit here > > [iree/runtime/src/iree/hal/drivers/hip/native_executable.c](https://github.com/iree-org/iree/blob/1b0acbf490b08c2a85ef24fb694f35cc5782b063/runtime/src/iree/hal/drivers/hip/native_executable.c#L109) > > Line 109...