[Bug] [Relax][CUDA] Wrong result for a K=1 matmul chain after AdjustMatmulOrder

Start by reading the AdjustMatmulOrder transform and the DLight reverse_compute_inline path mentioned in the report, then run the supplied reproducer on a CUDA-enabled TVM build. Compare the scheduled TIR for the original and transformed modules, especially where the clipped exp chain appears in reduction initialization and update. Done means both compiled variants match the NumPy result for the K=1 matmul chain.

Expected behavior

The original module and the module after AdjustMatmulOrder should return the same values. Both should match the NumPy calculation in the script below.

Actual behavior

I ran into this with a chain of two matmuls followed by two clipped exp operations. The first matmul has a reduction dimension of 1. Both versions compile and run, but the transformed version returns different values:

OFF first = 0.6067528 max abs error = 0.0
OFF allclose = True
ON first = 0.7993563 max abs error = 0.19286269
ON allclose = False

The OFF result matches NumPy. The ON result does not: all 30 output values fail rtol=0.02, atol=0.002. I can also reproduce this with a different random input.

In the scheduled ON TIR, the exp chain is in both the reduction initialization and update after DLight's reverse_compute_inline. It should run once on the completed matmul result.

Environment

• Apache TVM 0.26.dev0+source.8f328e8 (source build)
• Linux 6.8.0-52-generic, x86-64; Python 3.11.16
• NVIDIA RTX A6000, driver 550.120, CUDA toolkit 12.4, target sm_86
• Relax default CUDA compilation pipeline

Steps to reproduce

Save as repro_k1_matmul_exp.py and run python repro_k1_matmul_exp.py on a CUDA-enabled TVM build:

import numpy as np
import tvm
from tvm import relax as R
from tvm.script import ir as I, relax as S, tirx as T

@I.ir_module
class Before:
    @S.function
    def main(
        a: S.Tensor((1, 10), "float16"),
        b: S.Tensor((20, 1), "float16"),
        c: S.Tensor((20, 3), "float16"),
    ) -> S.Tensor((10, 3), "float32"):
        with S.dataflow():
            outer = S.matmul(b, a, out_dtype=None)
            transposed = S.permute_dims(outer, axes=None)
            matmul = S.matmul(transposed, c, out_dtype="float32")
            first = S.exp(
                S.subtract(
                    S.clip(matmul, T.float64(-0.25), T.float64(0.25)),
                    S.multiply(
                        S.clip(matmul, T.float64(-0.25), T.float64(2.0)),
                        S.const(0.5, "float32"),
                    ),
                )
            )
            second = S.exp(
                S.subtract(
                    S.clip(matmul, T.float64(-0.25), T.float64(0.25)),
                    S.multiply(
                        S.clip(first, T.float64(-0.25), T.float64(2.0)),
                        S.const(0.5, "float32"),
                    ),
                )
            )
            S.output(second)
        return second

dev = tvm.cuda(0)
target = tvm.target.Target({
    "kind": "cuda",
    "arch": "sm_" + dev.compute_version.replace(".", ""),
    "host": {"kind": "llvm", "mcpu": "generic"},
})
with tvm.transform.PassContext(opt_level=2):
    After = R.transform.AdjustMatmulOrder()(Before)

rng = np.random.default_rng(710000)
a, b, c = [
    ((rng.integers(0, 2, size=shape, dtype="int32") * 2 - 1) / 16).astype("float16")
    for shape in ((1, 10), (20, 1), (20, 3))
]
outer = (b @ a).astype("float16")
matmul = outer.T.astype("float32") @ c.astype("float32")
expected = matmul.copy()
for _ in range(2):
    expected = np.exp(
        np.clip(matmul, -0.25, 0.25)
        - 0.5 * np.clip(expected, -0.25, 2.0)
    )

# Freeze both branches before compiling either one.
frozen = {"OFF": tvm.ir.save_json(Before), "ON": tvm.ir.save_json(After)}
for name, serialized in frozen.items():
    with target, tvm.transform.PassContext(
        opt_level=2, disabled_pass=["AdjustMatmulOrder"]
    ):
        ex = tvm.compile(
            tvm.ir.load_json(serialized),
            target=target,
            relax_pipeline="default",
            tir_pipeline="default",
        )
    vm = R.VirtualMachine(ex, dev)
    output = vm["main"](*[
        tvm.runtime.tensor(x.copy(), dev) for x in (a, b, c)
    ]).numpy()
    print(name, "first =", output.flat[0],
          "max abs error =", np.max(np.abs(output - expected)))
    print(name, "allclose =", np.allclose(
        output, expected, rtol=0.02, atol=0.002
    ))

Triage

• needs-triage
• type: bug
• backend:cuda