tangyixuan01 opened a new issue, #20105: URL: https://github.com/apache/tvm/issues/20105
Thanks for participating in the TVM community! We use https://discuss.tvm.apache.org/ for any general usage questions and discussions. The issue tracker is used for actionable items such as feature proposals discussion, roadmaps, and bug tracking. You are always welcomed to post on the forum first :smile_cat: Issues that are inactive for a period of time may get closed. We adopt this policy so that we won't lose track of actionable issues that may fall at the bottom of the pile. Feel free to reopen a new one if you feel there is an additional problem that needs attention when an old one gets closed. ### Expected behavior relax.op.nn.adaptive_avg_pool2d should return the average of each adaptive pooling window. For a `(1, 1, 3, 4)` NCHW input and `output_size=(2, 2)`, the standard overlapping windows are: height: [0:2], [1:3] width : [0:2], [2:4] For the input below, the expected output is: input = 10 11 12 13 14 15 16 17 18 19 20 21 expected = 12.5 14.5 16.5 18.5 ### Actual behavior The module builds and runs without an exception, but the LLVM CPU execution returns: 6.5 7.5 8.5 9.5 The maximum absolute difference is `9.0`. The same incorrect result was observed with the project's `default`, `zero`, and `none` Relax pipelines. This is a numerical correctness issue rather than a compilation crash. It can silently propagate incorrect values to later model computations. The original fuzzing seed contained additional `reshape`, `leakyrelu`, `pow`, and `strided_slice` nodes. The first output, directly produced by `adaptive_avg_pool2d`, already mismatched in all 8 elements. The `pow` node was ruled out: it uses a positive integer exponent and its result is dead code, not part of the returned values. ### Environment - OS: Linux x86_64 under WSL2 (`6.6.114.1-microsoft-standard-WSL2`) - Python: 3.10.20 - NumPy: 1.26.4 - TVM: 0.25.dev0 - TVM source under test: local `tvm-relax-src` fork, commit `4d9d129c9` - Target: `llvm`, executed by `relax.VirtualMachine` on CPU - Build used for the reproduction: `/home/cxk/tvm-relax-src/build-gcov` ### Steps to reproduce Save the following as `repro.py` in a configured TVM environment: import numpy as np import tvm from tvm import relax from tvm.script import relax as R data = np.arange(10, 22, dtype="float32").reshape(1, 1, 3, 4) @tvm.script.ir_module class Mod: @R.function def main(x: R.Tensor((1, 1, 3, 4), "float32")) -> R.Tensor((1, 1, 2, 2), "float32"): return R.nn.adaptive_avg_pool2d(x, output_size=(2, 2), layout="NCHW") ex = relax.build(Mod, target="llvm") vm = relax.VirtualMachine(ex, tvm.cpu()) actual = vm["main"](tvm.runtime.tensor(data, tvm.cpu())).numpy()[0, 0] expected = np.array([[12.5, 14.5], [16.5, 18.5]], dtype="float32") print("actual:", actual) print("expected:", expected) print("max_abs_diff:", np.max(np.abs(actual - expected))) np.testing.assert_allclose(actual, expected, rtol=1e-5, atol=1e-5) ### Triage * needs-triage Related issue: (https://github.com/apache/tvm/issues/19520) reports a CUDA compilation crash for `adaptive_avg_pool2d` when the output size does not evenly divide the input. The present report is different in backend and symptom: LLVM/CPU execution succeeds but returns incorrect values. It should be treated as related, not as a duplicate. * needs-triage -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
