update hotfix branch

examples/gpt-demo/01-GPT-Training.ipynb

@@ -184,14 +184,14 @@
    "lastKernelId": null
   },
   "kernelspec": {
-   "display_name": "scorep-python3",
-   "language": "python3",
-   "name": "scorep-python3"
+   "display_name": "JUmPER",
+   "language": "python",
+   "name": "jumper"
   },
   "language_info": {
    "file_extension": ".py",
    "mimetype": "text/plain",
-   "name": "Any text"
+   "name": "python"
   }
  },
  "nbformat": 4,

pyproject.toml

@@ -7,7 +7,7 @@ build-backend = "setuptools.build_meta:__legacy__"
 
 [project]
 name='jumper-kernel'
-version='1.0.0'
+version='1.1.0'
 authors=[
   {name='Elias Werner',email='[email protected]'},
 ]
@@ -21,13 +21,14 @@ classifiers=[
 
 dependencies = [
   "ipykernel",
+  "ipywidgets",
+  "ipympl",
   "jupyter-client",
   "astunparse",
   "dill",
   "itables",
   "matplotlib",
   "pandas",
-  "ipywidgets",
   "pynvml" # we need that only for Nvidia GPU systems
 ]
 

src/jumper/perfdatahandler.py

@@ -3,7 +3,6 @@
 import pickle
 import codecs
 import time
-from datetime import datetime
 import os
 import subprocess
 import sys
@@ -28,8 +27,9 @@ class PerformanceDataHandler:
     def __init__(self):
         self.code_history = []
         self.performance_data_history = []
+        self.time_indices = []
         self.nodelist = None
-        # for local it's none, otherwise points to slurm/ssh/... monitor
+        # for local, it's none, otherwise points to slurm/ssh/... monitor
         self.monitor_module = None
         # the object from the monitor module
         self.monitor = None
@@ -78,8 +78,12 @@ def get_perfdata_history(self):
     def get_code_history(self):
         return self.code_history
 
-    def append_code(self, time, code):
-        self.code_history.append([time, code])
+    def get_time_indices(self):
+        return self.time_indices
+
+    def append_code(self, time_, code, time_indices=None):
+        self.code_history.append([time_, code])
+        self.time_indices.append(time_indices)
 
     def get_perfdata_aggregated(self):
         perfdata_aggregated = []
@@ -157,7 +161,12 @@ def get_perfdata_aggregated(self):
                 # add cell index and the number of measurements
                 # we will use that in the visualization to generate
                 # a color transition in the graphs and add the cell index
-                time_indices[node].append((idx, len(perfdata[node][2])))
+                if self.time_indices[idx]:
+                    # for cells tracked in multi cell mode, we can use the sub
+                    # indices created
+                    time_indices[node].extend(self.time_indices[idx][node])
+                else:
+                    time_indices[node].append((idx, len(perfdata[node][2])))
 
         return perfdata_aggregated, time_indices
 
@@ -326,7 +335,7 @@ def start_perfmonitor(self, pid):
 
         self.starttime = time.perf_counter()
 
-    def end_perfmonitor(self, code):
+    def end_perfmonitor(self):
         duration = time.perf_counter() - self.starttime
 
         if self.monitor_module:
@@ -352,6 +361,4 @@ def end_perfmonitor(self, code):
         performance_data_nodes = self.parse_perfdata_from_stdout(
             stdout_data_node
         )
-        if performance_data_nodes:
-            self.append_code(datetime.now(), code)
         return performance_data_nodes, duration

src/jumper/userpersistence.py

@@ -101,6 +101,7 @@ def jupyter_dump(self):
         Generate code for kernel ghost cell to dump notebook persistence for
         subprocess.
         """
+
         jupyter_dump_ = (
             "import sys\n"
             "import os\n"
@@ -111,7 +112,7 @@ def jupyter_dump(self):
             f"'{self.paths['jupyter']['sys_path']}',{self.marshaller})\n"
             f"dump_variables({str(self.jupyter_variables)},globals(),"
             f"'{self.paths['jupyter']['var']}',"
-            f"{self.marshaller})"
+            f"{self.marshaller})\n"
         )
 
         return jupyter_dump_

src/jumper/visualization.py

@@ -32,6 +32,7 @@ def plot_graph(ax, metric, perfdata, time_indices=None, color=None):
             int(os.environ.get("JUMPER_REPORT_FREQUENCY", 2)),
         )
     ]
+
     if metric == perfmetrics["cpu_agg"]:
         ax.plot(
             x_scale, perfdata[0][0][-3], label="Mean", color=(0.20, 0.47, 1.00)
@@ -147,24 +148,26 @@ def plot_graph(ax, metric, perfdata, time_indices=None, color=None):
     ax.legend()
     ax.grid(True)
 
-    # in multi node case, we have to iterate over the indices (time_indices)
-    # and not only 0 here
-    current_index = 0
-    target_index = -1
-    transition_offset = (x_scale[1] - x_scale[0]) / 2
-    start_offset = 0
+    # colorization of the plot in case of multiple cells
     if time_indices:
+        # in multi node case, we have to iterate over the indices (
+        # time_indices) and not only 0 here
+        current_index = 0
+        target_index = 0
+        transition_offset = (x_scale[1] - x_scale[0]) / 2
+        start_offset = 0
         last_idx = time_indices[0][-1][0]
 
-        for cell_idx, n_ms in time_indices[0]:
-
+        # for multi cell mode, we might have sub indices
+        for cell_idx, values in enumerate(time_indices[0]):
+            sub_idx, n_ms = values
             target_index = target_index + n_ms
             # don't use offset for last cell
-            if cell_idx == last_idx:
+            if sub_idx == last_idx:
                 transition_offset = 0
             ax.axvspan(
                 x_scale[current_index] + start_offset,
-                x_scale[target_index] + transition_offset,
+                x_scale[min(target_index, len(x_scale)-1)] + transition_offset,
                 facecolor=color[cell_idx],
                 alpha=0.3,
             )
@@ -174,7 +177,7 @@ def plot_graph(ax, metric, perfdata, time_indices=None, color=None):
                 + start_offset
                 + (
                     (
-                        x_scale[target_index]
+                        x_scale[min(target_index, len(x_scale)-1)]
                         + transition_offset
                         - x_scale[current_index]
                         + start_offset
@@ -188,7 +191,7 @@ def plot_graph(ax, metric, perfdata, time_indices=None, color=None):
             ax.text(
                 text_x_pos,
                 text_y_pos,
-                "#" + str(cell_idx),
+                "#" + str(sub_idx),
                 style="italic",
                 bbox={"facecolor": "lightgrey", "alpha": 0.5, "pad": 2},
             )

tests/kernel/multicell.yaml

@@ -3,7 +3,7 @@
   - - "Multicell mode enabled. The following cells will be marked for instrumented execution."
 -
   - "c = np.sum(c_mtx)"
-  - - "Cell marked for multicell mode. It will be executed at position 1"
+  - - "Cell marked for multicell mode. It will be executed at position 0"
 -
   - "%%abort_multicellmode"
   - - "Multicell mode aborted."
@@ -15,23 +15,23 @@
     with scorep.instrumenter.enable():
         c = np.sum(c_mtx)
     c_vec = np.arange(b, c)
-  - - "Cell marked for multicell mode. It will be executed at position 1"
+  - - "Cell marked for multicell mode. It will be executed at position 0"
 -
   - |-
     print('c =', c)
     print('Sum(c_vec) =', c_vec.sum())
-  - - "Cell marked for multicell mode. It will be executed at position 2"
+  - - "Cell marked for multicell mode. It will be executed at position 1"
 -
   - "%%finalize_multicellmode"
   - - "\0"
-    - "Executing cell 1\n"
+    - "Executing cell 0\n"
     - "with scorep.instrumenter.enable():\n"
     - "    c = np.sum(c_mtx)\n"
     - "c_vec = np.arange(b, c)\n"
     - "----------------------------------\n"
     - "\n"
     - "\n"
-    - "Executing cell 2\n"
+    - "Executing cell 1\n"
     - "print('c =', c)\n"
     - "print('Sum(c_vec) =', c_vec.sum())\n"
     - "----------------------------------\n"

tests/test_kernel.py

@@ -34,6 +34,7 @@ def check_stream_output(self, code, expected_output, stream="stdout"):
             # some messages can be of type 'execute_result'
             # type instead of stdout
             # self.assertEqual(msg["content"]["name"], stream)
+
             if msg["header"]["msg_type"] == "stream":
                 self.assertEqual(msg["content"]["name"], stream)
                 self.assertEqual(msg["content"]["text"], expected_msg)
@@ -42,6 +43,7 @@ def check_stream_output(self, code, expected_output, stream="stdout"):
                     msg["content"]["data"]["text/plain"], expected_msg
                 )
 
+
     def check_from_file(self, filename):
 
         with open(filename, "r") as file: