Fix Inspector plot crash on non-finite metrics (pytorch#23400)

### Summary

Follow-up to pytorch#23276. While testing that one I noticed
`compare_results(plot=True)` (and so `inspector_cli --compare_results`)
still crashes when a metric isn't finite. SNR is inf when the run output
matches the reference exactly, and -inf when the reference is all zeros.
`plot_metric` then passes that into `plt.ylim` and matplotlib raises
`ValueError: Axis limits cannot be NaN or Inf`. So the plot fails in the
best case, when outputs match perfectly.

This plots non-finite values as 0 (same as we do for None), while the
bar label still shows the real value (`inf`, `-inf`, `nan`). Also fixed
the `plot_metric` type hint to `List[Optional[float]]` to match what
compare_results passes in.

### Test plan

Added test_compare_results_plot_non_finite_metrics in
devtools/inspector/tests/inspector_utils_test.py. It covers identical
tensors (SNR inf) and an all-zero reference (SNR -inf) with plot=True.
It fails on main with the ValueError above and passes with the fix, and
the rest of inspector_utils_test.py passes (91 tests). ufmt and flake8
are clean with the versions from requirements-lintrunner.txt.


cc @Gasoonjia @nil-is-all

Author: u7k4rs6

devtools/inspector/_inspector_utils.py

@@ -394,16 +394,23 @@ def style_text_size(val, size=12):
         )
 
 
-def plot_metric(result: List[float], metric_name: str):
+def plot_metric(result: List[Optional[float]], metric_name: str):
+    # Nothing to plot (no outputs to compare), so don't write an empty chart.
+    if len(result) == 0:
+        return
+
     import matplotlib.pyplot as plt
     import numpy as np
 
     # Clear the current figure, otherwise this plot will be on top of previous plots
     plt.clf()
     plt.figure(figsize=(8, 6))
 
-    # Non-tensor outputs have no metric value (None); plot them as 0.
-    values = [v if v is not None else 0.0 for v in result]
+    # Non-tensor outputs have no metric value (None), and metrics can be
+    # non-finite (SNR is inf for identical tensors, NaN for all-zero ones).
+    # Plot these as 0; matplotlib cannot use inf/NaN as axis limits. The bar
+    # label still shows the actual value.
+    values = [v if v is not None and math.isfinite(v) else 0.0 for v in result]
     x_axis = np.arange(len(values))
     bars = plt.bar(x_axis, values, width=0.5)
     plt.grid(True, which="major", axis="y")

devtools/inspector/tests/inspector_utils_test.py

@@ -6,6 +6,8 @@
 
 # pyre-unsafe
 
+import math
+import os
 import tempfile
 import unittest
 from typing import Dict, Tuple
@@ -42,6 +44,7 @@
     map_runtime_aot_intermediate_outputs,
     merge_runtime_overlapping_debug_handles,
     NodeFilter,
+    plot_metric,
     propagate_back_debug_handle,
     TimeScale,
 )
@@ -246,6 +249,36 @@ def test_compare_results_with_non_tensor_output(self):
                 self.assertIsNotNone(values[0])
                 self.assertIsNone(values[1])
 
+    def test_compare_results_plot_non_finite_metrics(self):
+        # SNR is inf for identical tensors and -inf when the reference is all
+        # zeros; plotting must not fail on non-finite axis limits.
+        import matplotlib
+
+        matplotlib.use("Agg")
+        a = torch.rand(4, 4)
+        zeros = torch.zeros(4, 4)
+        for ref, out, expected_snr in ((a, a.clone(), math.inf), (zeros, a, -math.inf)):
+            results = compare_results([ref], [out], plot=True)
+            self.assertEqual(results["snr"], [expected_snr])
+
+    def test_plot_metric_empty_result(self):
+        # No values to plot: return without creating a figure or writing a file.
+        import matplotlib
+
+        matplotlib.use("Agg")
+        import matplotlib.pyplot as plt
+
+        plt.close("all")
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            cwd = os.getcwd()
+            os.chdir(tmp_dir)
+            try:
+                plot_metric([], "snr")
+            finally:
+                os.chdir(cwd)
+            self.assertEqual(os.listdir(tmp_dir), [])
+        self.assertEqual(plt.get_fignums(), [])
+
     def test_merge_overlapping_debug_handles_basic(self):
         big_tensor = torch.rand(100, 100)
         intermediate_outputs = {