🐛 Describe the bug

Describe the bug

torch.compile(backend='inductor') silently produces incorrect results when adaptive_avg_pool2d is fused with downstream flatten + sum operations. The bug triggers when the channel dimension is not a power of 2 (e.g., C=33, 65, 129).

This is a silent correctness bug — no error is raised, but the fused kernel computes wrong output values. The adaptive_avg_pool2d kernel alone produces correct results; the error is introduced during fusion with the subsequent reduction.

Minimal repro

import torch

def f(x):
    y = torch.nn.functional.adaptive_avg_pool2d(x, 7)
    return y.flatten(1).sum(dim=-1)

torch.manual_seed(42)
x = torch.randn(2, 33, 8, 8, dtype=torch.float64, device='cuda')

# Eager — correct
print(f"Eager: {f(x).tolist()}")
# [-52.27041819362077, 30.75847231895647]

# Inductor — wrong
torch._dynamo.reset()
compiled = torch.compile(f, backend='inductor')(x)
print(f"Inductor: {compiled.tolist()}")
# [-52.27041819362077, 34.325837206874446]  ← second element is wrong

print(f"Max diff: {(f(x) - compiled).abs().max().item()}")
# 3.567365  ← catastrophic for fp64

Key observations

1. Pool alone is correct; fusion introduces the error:

# These are all correct (diff ≈ 0):
torch.compile(lambda x: F.adaptive_avg_pool2d(x, 7))(x)           # pool only
torch.compile(lambda x: F.adaptive_avg_pool2d(x, 7).sum())(x)     # pool + sum
torch.compile(lambda x: F.adaptive_avg_pool2d(x, 7).flatten(1))(x) # pool + flatten

# This is WRONG (diff ≈ 3.6):
torch.compile(lambda x: F.adaptive_avg_pool2d(x, 7).flatten(1).sum(-1))(x)

2. Channel dimension determines whether the bug triggers:

3. Spatial dimensions also matter:

With C=33, output_size=7:

• H=7 (no resize needed): diff=0 ✅
• H=8 (8→7, non-integer stride): diff=7.37 ❌
• H=14 (14→7, integer stride): diff=0 ✅ on pt211, diff≠0 on nightly
• H=15 (15→7, non-integer stride): diff=8.86 ❌

4. CUDA only — CPU produces correct results.

5. Deterministic across random seeds — not a race condition.

Versions

• PyTorch 2.11.0+cu126: bug present
• PyTorch 2.12.0.dev20260410+cu126 (nightly): bug present
• CUDA 12.6, Tesla T4
• OS: Ubuntu 20.04, Python 3.11

cc @ptrblck @msaroufim @eqy @jerryzh168 @tinglvv @nWEIdia @mikaylagawarecki @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @aakhundov @coconutruben @jataylo