xmanager-slurm 0.4.8__py3-none-any.whl → 0.4.9__py3-none-any.whl

xm_slurm/executors.py

@@ -57,8 +57,13 @@ class Slurm(xm.Executor):
 
     requeue: bool = True  # Is this job ellible for requeueing?
     requeue_on_exit_code: int = 42  # The exit code that triggers requeueing
+    requeue_on_timeout: bool = True  # Should the job requeue upon timeout minus the grace period
     requeue_max_attempts: int = 5  # How many times to attempt requeueing
 
+    @property
+    def requeue_timeout(self) -> dt.timedelta:
+        return self.time - self.timeout_signal_grace_period
+
     def __post_init__(self) -> None:
         if not isinstance(self.time, dt.timedelta):
             raise TypeError(f"time must be a `datetime.timedelta`, got {type(self.time)}")

xm_slurm/templates/slurm/fragments/monitor.bash.j2

@@ -1,4 +1,4 @@
-{% macro monitor(requeue_max_attempts, requeue_exit_code) -%}
+{% macro monitor(requeue_max_attempts, requeue_exit_code, requeue_on_timeout, requeue_timeout) -%}
 __xm_slurm_wait_for_children() {
   if [[ -n "${SLURM_ARRAY_JOB_ID:-}" ]]; then
     local -r JOB_ID="${SLURM_ARRAY_JOB_ID}_${SLURM_ARRAY_TASK_ID}"
@@ -7,30 +7,69 @@ __xm_slurm_wait_for_children() {
   fi
 
   # If there are no child jobs we should error out
-  if [ -z "$(jobs -p)" ]; then
+  children=( $(jobs -p) )
+{% raw %}
+  if [ ${#children[@]} -eq 0 ]; then
+{% endraw %}
     echo "ERROR: no child jobs exist..." >&2
-    exit -1
+    exit 1
   fi
 
-  # Loop through all job IDs in the background job list and wait for them to finish
-  for job in "$(jobs -p)"; do
-    echo "INFO: Waiting for job ${job} to finish..."
-    set +e
-    wait "${job}"
-    local -r JOB_EXIT_CODE="${?}"
-    set -e
+{% if requeue_on_timeout %}
+  # Start a watchdog process to signal timeout.
+  sleep {{ requeue_timeout }} &
+  timeout_pid=$!
+{% endif %}
 
-    if [ "${JOB_EXIT_CODE}" -eq "{{ requeue_exit_code }}" ] && [ "${SLURM_RESTART_COUNT-0}" -le "{{ requeue_max_attempts }}" ]; then
-      echo "INFO: Received requeue exit code {{ requeue_exit_code }} from job ${job}. Requeing Slurm job ${JOB_ID} after ${SLURM_RESTART_COUNT-0} restarts." >&2
+{% raw %}
+  while [ ${#children[@]} -gt 0 ]; do
+{% endraw %}
+    echo "INFO: Waiting for child processes to finish..."
+{% if requeue_on_timeout %}
+    # Wait on either one of the child processes or the timeout process.
+    wait -n -p child_pid "${children[@]}" "${timeout_pid}"
+{% else %}
+    wait -n -p child_pid "${children[@]}"
+{% endif %}
+    local child_exit_code=$?
+
+{% if requeue_on_timeout %}
+    # If the finished process is the watchdog, trigger the timeout handling.
+    if [ "${child_pid}" = "${timeout_pid}" ]; then
+      echo "INFO: Timeout of {{ requeue_timeout }} seconds reached. Killing remaining processes: ${children[*]}" >&2
+      kill "${children[@]}" 2>/dev/null || true
       scontrol requeue "${JOB_ID}"
       exit {{ requeue_exit_code }}
-    elif [ "${JOB_EXIT_CODE}" -ne 0 ]; then
-      echo "ERROR: Job ${job} exited with code ${JOB_EXIT_CODE}." >&2
-      exit "${JOB_EXIT_CODE}"
-    else
-      echo "INFO: Job ${job} exited successfully." >&2
     fi
+{% endif %}
+
+    echo "INFO: Process ${child_pid} finished with exit code ${child_exit_code}."
+
+    # Handle the exit code of the finished process.
+    if [ "${child_exit_code}" -eq "{{ requeue_exit_code }}" ] && [ "${SLURM_RESTART_COUNT:-0}" -le "{{ requeue_max_attempts }}" ]; then
+      echo "INFO: Received requeue exit code {{ requeue_exit_code }} from process ${child_pid}. Requeuing Slurm job ${JOB_ID} after ${SLURM_RESTART_COUNT-0} restarts." >&2
+      scontrol requeue "${JOB_ID}"
+      exit {{ requeue_exit_code }}
+    elif [ "${child_exit_code}" -ne 0 ]; then
+      echo "ERROR: Process ${child_pid} exited with code ${child_exit_code}." >&2
+      exit "${child_exit_code}"
+    fi
+
+    # Remove the finished PID from the array in a concise way.
+    for i in "${!children[@]}"; do
+      if [ "${children[i]}" = "$child_pid" ]; then
+        unset 'children[i]'
+        break
+      fi
+    done
+
+    # Reindex the array.
+    children=( "${children[@]}" )
   done
+
+{% if requeue_on_timeout %}
+  kill "$timeout_pid" 2>/dev/null || true
+{% endif %}
 }
 
 __xm_slurm_wait_for_children

xm_slurm/templates/slurm/job.bash.j2

@@ -73,7 +73,7 @@ echo "[INFO] Start timestamp: $(date)"
 
 {% block monitor -%}
 {% from 'fragments/monitor.bash.j2' import monitor %}
-{{ monitor(job.executor.requeue_max_attempts, job.executor.requeue_on_exit_code) }}
+{{ monitor(job.executor.requeue_max_attempts, job.executor.requeue_on_exit_code, job.executor.requeue_on_timeout, job.executor.requeue_timeout.seconds) }}
 {%- endblock monitor %}
 
 

{xmanager_slurm-0.4.8.dist-info → xmanager_slurm-0.4.9.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: xmanager-slurm
-Version: 0.4.8
+Version: 0.4.9
 Summary: Slurm backend for XManager.
 Project-URL: GitHub, https://github.com/jessefarebro/xm-slurm
 Author-email: Jesse Farebrother <jfarebro@cs.mcgill.ca>

{xmanager_slurm-0.4.8.dist-info → xmanager_slurm-0.4.9.dist-info}/RECORD

@@ -6,7 +6,7 @@ xm_slurm/constants.py,sha256=zefVtlFdflgSolie5g_rVxWV-Zpydxapchm3y0a2FDc,999
 xm_slurm/dependencies.py,sha256=-5gN_tpfs3dOA7H5_MIHO2ratb7F5Pm_yjkR5rZcgI8,6421
 xm_slurm/executables.py,sha256=fGmrFBl-258bMn6ip5adYeM7xxUHAeIbDN9zD2FDGtY,6373
 xm_slurm/execution.py,sha256=c0aV1h2tKQFyAGM6JLd16MWFgpRLKAbcutZz17xPUSw,31400
-xm_slurm/executors.py,sha256=fMtxGUCi4vEKmb_p4JEpqPUTh7L_f1LcR_TamMLAWNg,4667
+xm_slurm/executors.py,sha256=bUgKcgtvf-nPGjcuHRzUAqD1r3_vwea_h-Y9MAB-Kqo,4887
 xm_slurm/experiment.py,sha256=94r0mhtUPUzw4eaUEz0kpsufC25wEGqlDhV4Fcr1ukY,39883
 xm_slurm/filesystem.py,sha256=4rKtq3t-KDgxJbSGt6JVyRJT_3lCN_vIKTcwKHpTo3I,4389
 xm_slurm/job_blocks.py,sha256=_F8CKCs5BQFj40a2-mjG71HfacvWoBXBDPDKEaKTbXc,616
@@ -39,13 +39,13 @@ xm_slurm/templates/docker/uv.Dockerfile,sha256=L2UJMX2c8waMdrRhiqPytQe3pTBu6u5Pp
 xm_slurm/templates/slurm/entrypoint.bash.j2,sha256=MRdSVwgGrgQdpEhqfkP35IidgsblrtVXB1YWzvE9hkk,666
 xm_slurm/templates/slurm/job-array.bash.j2,sha256=smxmSSzBEUHm6MJF-nYPVVjK6CLKrb1fRxF_tfrzAX8,552
 xm_slurm/templates/slurm/job-group.bash.j2,sha256=Cp8YhNOxYqaOkl4MFjQlcaLMGZwdDh97m8OGT5RWbAo,1101
-xm_slurm/templates/slurm/job.bash.j2,sha256=d35VYHdAKkgVK8s4XnUDJwQR0gLnDWRJu-Ldz-qALmQ,1914
-xm_slurm/templates/slurm/fragments/monitor.bash.j2,sha256=HYqYhXsTv8TCed5UaGCZVGIYsqxSKHcnPyNNTHWNvxc,1279
+xm_slurm/templates/slurm/job.bash.j2,sha256=pNKir1tkmRTGDiGxlQ3DkUaW9Zos_gdkkXJC_xX5Cxo,1985
+xm_slurm/templates/slurm/fragments/monitor.bash.j2,sha256=BJ1brSjhESOe9VX_OYaPyy9-qE3uiFlzxp8ZkFcTw8Y,2504
 xm_slurm/templates/slurm/fragments/proxy.bash.j2,sha256=VJLglZo-Nvx9R-qe3rHTxr07CylTQ6Z9NwBzvIpAZrA,814
 xm_slurm/templates/slurm/runtimes/apptainer.bash.j2,sha256=lE2EWVCK2O-n08RL4_MJYIikVTvODjcYKuv7Eh73Q2w,1932
 xm_slurm/templates/slurm/runtimes/podman.bash.j2,sha256=3j7K5eyXt_WhXK0EoMlxnhlmFVJ2JyxRKbsMRaDqzSs,1148
-xmanager_slurm-0.4.8.dist-info/METADATA,sha256=8QQ9xbptuObTCHB8WaYr0rSKOgxK5ojCaeD7mN9Qvl0,1042
-xmanager_slurm-0.4.8.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-xmanager_slurm-0.4.8.dist-info/entry_points.txt,sha256=_HLGmLgxuQLOPmF2gOFYDVq2HqtMVD_SzigHvUh8TCY,49
-xmanager_slurm-0.4.8.dist-info/licenses/LICENSE.md,sha256=IxstXr3MPHwTJ5jMrByHrQsR1ZAGQ2U_uz_4qzI_15Y,11756
-xmanager_slurm-0.4.8.dist-info/RECORD,,
+xmanager_slurm-0.4.9.dist-info/METADATA,sha256=WWPRzVrTsK5t8kD732EIejSlNgQ8KP01Ln7eM1Mj4e4,1042
+xmanager_slurm-0.4.9.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+xmanager_slurm-0.4.9.dist-info/entry_points.txt,sha256=_HLGmLgxuQLOPmF2gOFYDVq2HqtMVD_SzigHvUh8TCY,49
+xmanager_slurm-0.4.9.dist-info/licenses/LICENSE.md,sha256=IxstXr3MPHwTJ5jMrByHrQsR1ZAGQ2U_uz_4qzI_15Y,11756
+xmanager_slurm-0.4.9.dist-info/RECORD,,