[laifenxiawucha]

Summary

Fix unbacked-symbol memo reuse when replaying fake-tensor binding sites.

Root Cause

Fake replay could reuse a bool-mask fake tensor's cached nonzero memo across two aten.index.Tensor nodes that carried distinct old unbacked_bindings. The second replayed node then tried to rename one fresh symbol to two different old symbols, triggering AssertionError: u6 -> u0 during Inductor fake propagation.

Fix

• add targeted fake-tensor unbacked-memo invalidation helpers
• clear those memos before replaying nodes with unbacked_bindings in ShapeProp, FakeTensorProp, PropagateUnbackedSymInts, FakeTensorUpdater, and Inductor fallback example propagation
• add regression tests for FX shape propagation and Inductor bool-mask indexing replay

Verification

• pre-fix reproducer in /home/jyz/cuda-fix-alignment/.venv/bin/python failed with torch._inductor.exc.InductorError: AssertionError: u6 -> u0
• post-fix verification passed by loading the patched Python modules into the same XPU wheel environment and rerunning:

import torch
from torch._dynamo import config as dynamo_config

class M(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.w = torch.nn.Parameter(torch.randn(3, 4, 4, device="xpu"))

    def forward(self, mask):
        y1 = self.w[:, ~mask]
        y2 = (self.w + self.w.transpose(1, 2))[:, ~mask]
        return y1, y2

with dynamo_config.patch(capture_dynamic_output_shape_ops=True):
    m = M().to("xpu")
    mask = torch.eye(4, dtype=torch.bool, device="xpu")
    eager = m(mask)
    compiled = torch.compile(m, backend="inductor", fullgraph=True)
    out = compiled(mask)
    torch.testing.assert_close(out[0].cpu(), eager[0].cpu())
    torch.testing.assert_close(out[1].cpu(), eager[1].cpu())

Issue: https://github.com/intel-sandbox/torch-xpu-ops-exp/issues/1893