[GPU][DT] Failed to compile toy_llama.mlir when const-eval is off
- Dominant language
- C++
- Stars
- 3.9k
- Forks
- 1k
- Avg merge
- 4d 16h
- Merged PRs (30d)
- 47
Description
To repro: `iree-compile --iree-hal-target-device=hip --iree-hip-target=gfx942 --iree-hip-encoding-layout-resolver=data-tiling --iree-llvmgpu-test-combine-layout-transformation=true ~/toy_llama.mlir -o /tmp/z.vmfb --iree-opt-data-tiling=false --iree-dispatch-creation-data-tiling --iree-opt-level=O3`
You can find the toy_llama.mlir from https://github.com/iree-org/iree-test-suites/tree/main/sharktank_models/llama3.1/assets
Note that it compiles if the option is on (which is default).
You can reproduce the issue from Codegen: https://gist.github.com/hanhanW/84281926535080272c1d66bcbe098bfd
`iree-compile --compile-from=executable-sources --compile-to=executable-targets repro.mlir`. The failures are mostly about encoding materialization. Some creates invalid expand_shape ops, and some crash.
Contributor guide
Assessment
This issue has not been assessed yet.