Fix freeze pane, add more metrics (#250)

* Fix freeze pane, add more metrics

* changes

* spell

* fix

* add stats

Author: xadupre

CHANGELOGS.rst

@@ -4,6 +4,7 @@ Change Logs
 0.7.14
 ++++++
 
+* :pr:`250`: add variables to track sequence nodes
 * :pr:`249`: patches _maybe_broadcast to support a corner case
 
 0.7.13

_unittests/ut_helpers/test_log_helper.py

@@ -268,7 +268,7 @@ def test_cube_logs_performance_cube_time(self):
         cube = CubeLogsPerformance(dfs, keep_last_date=True)
         cube.load()
         ct = cube.clone()
-        self.assertEqual((52, 111), ct.shape)
+        self.assertEqual((52, 116), ct.shape)
 
     def test_duplicate(self):
         df = pandas.DataFrame(

onnx_diagnostic/helpers/log_helper.py

@@ -1167,7 +1167,7 @@ def to_excel(
                     df.to_excel(
                         writer,
                         sheet_name=name,
-                        freeze_panes=(df.columns.nlevels + df.index.nlevels, df.index.nlevels),
+                        freeze_panes=(df.columns.nlevels + 1, df.index.nlevels),
                     )
                     f_highlights[name] = tview.f_highlight
                     if tview.plots:
@@ -1222,15 +1222,15 @@ def to_excel(
                     writer,
                     sheet_name=name,
                     freeze_panes=(
-                        sbs_raw.columns.nlevels + sbs_raw.index.nlevels,
+                        sbs_raw.columns.nlevels + 1,
                         sbs_raw.index.nlevels,
                     ),
                 )
                 sbs_agg.to_excel(
                     writer,
                     sheet_name=f"{name}-AGG",
                     freeze_panes=(
-                        sbs_agg.columns.nlevels + sbs_agg.index.nlevels,
+                        sbs_agg.columns.nlevels + 1,
                         sbs_agg.index.nlevels,
                     ),
                 )
@@ -1456,9 +1456,7 @@ def _mkc(m, s):
 
 
 class CubeLogsPerformance(CubeLogs):
-    """
-    Processes logs coming from experiments.
-    """
+    """Processes logs coming from experiments."""
 
     def __init__(
         self,
@@ -1511,20 +1509,25 @@ def __init__(
             "n_model_faster2x",
             "n_model_faster3x",
             "n_model_faster4x",
+            "n_model_faster5x",
             "n_node_attention",
             "n_node_attention23",
-            "n_node_rotary_embedding",
-            "n_node_rotary_embedding23",
-            "n_node_layer_normalization",
-            "n_node_layer_normalization23",
+            "n_node_causal_mask",
+            "n_node_constant",
             "n_node_control_flow",
-            "n_node_scatter",
+            "n_node_expand",
             "n_node_function",
+            "n_node_gqa",
             "n_node_initializer",
             "n_node_initializer_small",
-            "n_node_constant",
+            "n_node_layer_normalization",
+            "n_node_layer_normalization23",
+            "n_node_reshape",
+            "n_node_rotary_embedding",
+            "n_node_rotary_embedding23",
+            "n_node_scatter",
+            "n_node_sequence",
             "n_node_shape",
-            "n_node_expand",
             "onnx_n_nodes_no_cst",
             "peak_gpu_torch",
             "peak_gpu_nvidia",
@@ -1690,6 +1693,11 @@ def first_err(df: pandas.DataFrame) -> pandas.Series:
                     "time_latency",
                     gdf(df, "time_latency_eager") > gdf(df, "time_latency", np.inf) * 3.98,
                 ),
+                n_model_faster5x=lambda df: gpreserve(
+                    df,
+                    "time_latency",
+                    gdf(df, "time_latency_eager") > gdf(df, "time_latency", np.inf) * 4.98,
+                ),
                 n_node_attention23=lambda df: gpreserve(
                     df, "time_latency_eager", gdf(df, "op_onnx__Attention")
                 ),
@@ -1720,6 +1728,11 @@ def first_err(df: pandas.DataFrame) -> pandas.Series:
                     + gdf(df, "op_onnx_com.microsoft_DecoderMaskedMultiHeadAttention", 0)
                     + gdf(df, "op_onnx_com.microsoft_SparseAttention", 0),
                 ),
+                n_node_gqa=lambda df: gpreserve(
+                    df,
+                    "time_latency_eager",
+                    gdf(df, "op_onnx_com.microsoft_GroupQueryAttention", 0),
+                ),
                 n_node_layer_normalization=lambda df: gpreserve(
                     df,
                     "time_latency_eager",
@@ -1764,9 +1777,22 @@ def first_err(df: pandas.DataFrame) -> pandas.Series:
                 n_node_shape=lambda df: gpreserve(
                     df, "time_latency_eager", gdf(df, "op_onnx__Shape")
                 ),
+                n_node_reshape=lambda df: gpreserve(
+                    df, "time_latency_eager", gdf(df, "op_onnx__Reshape")
+                ),
                 n_node_expand=lambda df: gpreserve(
                     df, "time_latency_eager", gdf(df, "op_onnx__Expand")
                 ),
+                n_node_causal_mask=lambda df: gpreserve(
+                    df,
+                    "time_latency_eager",
+                    gdf(df, "op_onnx__CausalMask", 0),
+                ),
+                n_node_sequence=lambda df: gpreserve(
+                    df,
+                    "time_latency_eager",
+                    gdf(df, "op_onnx__SequenceAt", 0) + gdf(df, "op_onnx__SplitToSequence", 0),
+                ),
             )
             assert (
                 formula in lambdas

onnx_diagnostic/torch_models/validate.py

@@ -1744,6 +1744,7 @@ def _simplify(p):
             "constant_folding",
             "remove_identity",
             "remove_duplicated_initializer",
+            "remove_duplicated_shape",
             "dynamic_dimension_naming",
             "inline",
             "check",