Problem

test_deepseek_model_pd_ratio_with_output_validation can fail during compilation with
BackendCompilerFailed. The immediate symptom is that generated FX code evaluates
view_default_1 using pow_2 before pow_2 is defined.

The failure is unrelated to whether PeepHolePass is enabled. This Issue is scoped to
the general FX graph-ordering invariant required by the compiler rewrite pipeline.

Reproduction

1. Run the regression case:

   python -m pytest -m nightly tests/regression/cli/test_throughput_optimizer.py::TestThroughputOptimizer::test_deepseek_model_pd_ratio_with_output_validation -vv -s
   

2. On an affected revision, compilation may raise:

   torch._dynamo.exc.BackendCompilerFailed
   ...
   UnboundLocalError: cannot access local variable 'pow_2' where it is not associated with a value
   

Expected

Compiler graph rewrite passes preserve FX topological order: symbolic shape-expression
nodes must be emitted before any view or reshape node that consumes them. The
DeepSeek PD-ratio output-validation regression should compile successfully.

Actual

AOT Autograd can produce a graph whose symbolic shape calculation appears after a
view consuming it. Subsequent rewrite passes can retain or rebuild this invalid order.
Generated code then references the symbolic expression before assignment, causing
BackendCompilerFailed.

Root Cause

• CompilerBackend.compile() entered rewrite passes without normalizing the AOT FX graph.
• LiftCombineQuantPass can rebuild view nodes, then runs dead-code elimination; the
  graph needs to be topologically restored and linted before that boundary.

Proposed Fix

1. Call stable_topo_sort(gm) and gm.graph.lint() before compiler graph rewrites.
2. After LiftCombineQuantPass rebuilds views, call stable_topo_sort(gm) and lint
   before dead-code elimination.
3. Add a focused unit regression with view(..., [-1, operator.pow(x, 2)]), asserting
   pow precedes view and graph lint passes before and after the pass.

Acceptance Criteria

• The dynamic-shape unit regression passes and validates FX graph linting.
• test_deepseek_model_pd_ratio_with_output_validation compiles and passes.
• Existing quantization and graph-rewrite regressions remain green.
• No pass relies on a consumer-before-producer FX node order.

Related Pull Request

• https://gitcode.com/Ascend/msmodeling/merge_requests/802

Evidence

• PR 802 contains the focused compiler and regression-test changes.
• The PR diff after synchronizing with master is limited to 3 files and +59/-2.
• Previous local verification recorded for the PR: targeted topology tests (2 passed),
  related rewrite tests (24 passed), and the DeepSeek PD-ratio regression passed.