[Bug] Inconsistent Results between Direct Optimization and Sequential Optimization in TVM

Jupiterghy commented 5 months ago

When applying optimization passes in TVM, there is a discrepancy in the results between directly applying opt_a(opt_b(mod)) and using a sequential optimization approach, where seq_ab = tvm.ir.transform.Sequential([opt_a, opt_b]) and seq_ba = tvm.ir.transform.Sequential([opt_b, opt_a]) are used.

Additionally, this issue seems to occur specifically when one of the optimizations is FakeQuantizationToInteger.

Actual behavior

Both structure and inference results of two mods which are applied optimizations by two ways are inconsistent.

Traceback information:

AssertionError: 
Not equal to tolerance rtol=0.01, atol=0.01

Mismatched elements: 1 / 18 (5.56%)
Max absolute difference: 1.3370061e+08
Max relative difference: 3.0688565e+10
 x: array([[[1.859827e-001]],

       [[3.138568e-311]],...
 y: array([[[ 1.859827e-001]],

       [[ 1.022716e-321]],...

Environment

Operating System: Ubuntu 18.04.5
TVM version: 0.15.dev0
ONNX: 1.15.0

Steps to reproduce

Download the ONNX model

Execute the script:


import onnx
import tvm
from tvm import relay
import numpy as np

def compare_outputs(output1, output2, rtol=1e-2, atol=1e-2): if len(output1) != len(output2): raise ValueError("Number of outputs in the two lists is different.") for i in range(len(output1)): output1_np = np.asarray(output1[i]) output2_np = np.asarray(output2[i]) np.testing.assert_allclose(output1_np, output2_np, rtol=rtol, atol=atol)

def compile_onnx(mod, params, inputs): mod = relay.transform.InferType()(mod) exec_mod = 'graph' target = 'llvm' ctx = tvm.cpu(0)

with tvm.transform.PassContext(opt_level=0):
    executor = relay.build_module.create_executor(
        exec_mod, mod, ctx, target, params
    ).evaluate()
output = executor(**inputs)
if isinstance(output, (tvm.runtime.container.ADT, list)):
    output = [r.numpy() for r in output]
elif output is not None:
    output = [output.numpy()]
return output

if name == "main": onnx_file = "model.onnx" onnx_model = onnx.load(onnx_file)

shape_dict = {'v13_0': [1], 'v11_0': [1, 38, 27, 1], 'v7_0': [18, 1, 1]}
inputs = {}
inputs['v13_0'] = np.random.random([1])
inputs['v11_0'] = np.random.random([1, 38, 27, 1])
inputs['v7_0'] = np.random.random([18, 1, 1])

mod, params = relay.frontend.from_onnx(onnx_model, shape_dict, freeze_params=True)

opt_a = tvm.relay.transform.AlterOpLayout()
opt_b = tvm.relay.transform.FakeQuantizationToInteger()
mod = tvm.relay.transform.InferType()(mod)
module_ab = opt_b(opt_a(mod))
module_ba = opt_a(opt_b(mod))
assert tvm.ir.structural_equal(module_ab, module_ba) #same
outputs_ab = compile_onnx(module_ab, params, inputs)
outputs_ba = compile_onnx(module_ba, params, inputs)
compare_outputs(outputs_ab, outputs_ba) #same

seq_ab = tvm.ir.transform.Sequential([opt_a, opt_b])
seq_ba = tvm.ir.transform.Sequential([opt_b, opt_a])
with tvm.transform.PassContext(opt_level=0):
    module_ab_2 = seq_ab(mod)
    module_ba_2 = seq_ba(mod)
assert tvm.ir.structural_equal(module_ab, module_ab_2) # assertion error, inconsistent
assert tvm.ir.structural_equal(module_ba, module_ba_2) # assertion error, inconsistent
assert tvm.ir.structural_equal(module_ab_2, module_ba_2) # assertion error, inconsistent
outputs_ab_2 = compile_onnx(module_ab_2, params, inputs)
outputs_ba_2 = compile_onnx(module_ba_2, params, inputs)
compare_outputs(outputs_ab, outputs_ab_2) # assertion error, inconsistent
compare_outputs(outputs_ba, outputs_ba_2) # assertion error, inconsistent
compare_outputs(outputs_ab_2, outputs_ba_2) # assertion error, inconsistent



### Triage

* needs-triage

vinx13 commented 5 months ago

If you call a pass directly (instead of using Sequential, it will bypass the check for opt_level, required_pass, etc.

Jupiterghy commented 4 months ago

If you call a pass directly (instead of using Sequential, it will bypass the check for opt_level, required_pass, etc.

Thank you for your response. However, I'm curious to understand why executing optimizations in the Sequential manner still results in inconsistency with different orderings.

apache / tvm