Add new slurm sbatch profile

config/slurm.yaml

@@ -0,0 +1,53 @@
+#
+# SBatch arguments for different run profile
+#
+
+multi-node-full:
+  # DGX run: 2 nodes x 8 A100 80Go SXM4
+  - --partition=staff-idt
+  - -w cn-d[003-004]
+  - --ntasks=1
+  - --gpus-per-task=a100l:8
+  - --exclusive
+  - --nodes=2
+  - --cpus-per-task=128
+  - --time=1:30:00
+  - --ntasks-per-node=1
+  - --mem=0
+
+single-node-full:
+  # DGX run: 1 node x 8 A100 80Go SXM4
+  - --partition=staff-idt
+  - -w cn-d[003-004]
+  - --ntasks=1
+  - --gpus-per-task=a100l:8
+  - --exclusive
+  - --nodes=1
+  - --cpus-per-task=128
+  - --time=1:30:00
+  - --ntasks-per-node=1
+  - --mem=0
+
+multi-node-small:
+  # Any GPU, 2 nodes x 2 GPU
+  - --partition=staff-idt
+  - --ntasks=1
+  - --gpus-per-task=2
+  - --exclusive
+  - --nodes=2
+  - --cpus-per-task=16
+  - --time=1:30:00
+  - --ntasks-per-node=1
+  - --mem=64G
+
+single-node-small:
+  # Any GPU, 1 node x 2 GPU
+  - --partition=staff-idt
+  - --ntasks=1
+  - --gpus-per-task=2
+  - --exclusive
+  - --nodes=1
+  - --cpus-per-task=16
+  - --time=1:30:00
+  - --ntasks-per-node=1
+  - --mem=64G

milabench/_version.py

@@ -1,5 +1,5 @@
 """This file is generated, do not modify"""
 
-__tag__ = "v0.1.0-12-g39e7cce9"
-__commit__ = "39e7cce9aec8a9e1ae7713137f287353ce718875"
-__date__ = "2024-06-17 13:41:35 -0400"
+__tag__ = "v0.1.0-20-g7246295a"
+__commit__ = "7246295a356186b55fa4b2b75480e3700c279b15"
+__date__ = "2024-06-20 09:18:17 -0400"

milabench/cli/pr.py

@@ -4,7 +4,7 @@
 from coleo import Option, tooled
 
 from ..common import _short_make_report
-from ..schedule import post_comment_on_pr
+from .schedule import post_comment_on_pr
 
 
 # fmt: off

milabench/cli/schedule.py

@@ -5,6 +5,7 @@
 
 import importlib_resources
 import requests
+import yaml
 from coleo import Option, tooled
 
 
@@ -14,6 +15,7 @@ class Arguments:
     sync: bool = False
     dry : bool = False
     args: list = field(default_factory=list)
+    profile: str = None
 # fmt: on
 
 
@@ -25,11 +27,29 @@ def arguments():
     # Print the command and return without running it
     dry: Option & bool = False
 
-    # pip arguments
+    # sbatch run profile
+    profile: Option & str = None
+
+    # script arguments
     # [remainder]
     args: Option = []
 
-    return Arguments(sync, dry, args)
+    return Arguments(sync, dry, args, profile)
+
+
+def get_sbatch_profiles(profile, default):
+    ROOT = os.path.dirname(__file__)
+    default_scaling_config = os.path.join(ROOT, "..", "..", "config", "slurm.yaml")
+
+    with open(default_scaling_config, "r") as fp:
+        sbatch_profiles = yaml.safe_load(fp)
+
+    args = sbatch_profiles.get(profile)
+
+    if args is None:
+        args = sbatch_profiles.get(default)
+
+    return args 
 
 
 @tooled
@@ -39,9 +59,9 @@ def cli_schedule(args=None):
     if args is None:
         args = arguments()
 
-    launch_milabench(args.args, sbatch_args=None, dry=args.dry, sync=args.sync)
-
+    sbatch_args = get_sbatch_profiles(args.profile, "single-node-small")
 
+    launch_milabench(args.args, sbatch_args=sbatch_args, dry=args.dry, sync=args.sync)
 
 
 def popen(cmd, callback=None):
@@ -120,6 +140,7 @@ class SetupOptions:
     config: str = "milabench/config/standard.yaml"
     env: str = "./env"
     python: str = "3.9"
+    fun: str = "run"
 
     def deduce_remote(self, current_branch):
         prefix = "refs/heads/"
@@ -164,35 +185,25 @@ def arguments(self):
             self.env,
             "-p",
             self.python,
+            "-f",
+            self.fun
         ]
 
 
 def launch_milabench(args, sbatch_args=None, dry: bool = False, sync: bool = False):
     sbatch_script = (
-        importlib_resources.files(__name__) / "scripts" / "milabench_run.bash"
+        os.path.abspath(importlib_resources.files(__name__) / ".." / "scripts" / "milabench_run.bash")
     )
     sbatch_script = str(sbatch_script)
 
-    # salloc --gres=gpu:rtx8000:1 --mem=64G --cpus-per-gpu=4
-
-    if sbatch_args is None:
-        sbatch_args = [
-            "--ntasks=1",
-            "--gpus-per-task=rtx8000:2",
-            "--cpus-per-task=8",
-            "--time=01:30:00",
-            "--ntasks-per-node=1",
-            "--mem=64G",
-        ]
-
     script_args = SetupOptions()
     script_args.deduce_from_repository()
     script_args = script_args.arguments()
 
     cmd = sbatch_args + [sbatch_script] + script_args + args
+    print("sbatch " + " ".join(cmd))
 
     if dry:
-        print("sbatch " + " ".join(cmd))
         code = 0
     else:
         code, _ = sbatch(cmd, sync=sync, tags=None)

milabench/cli/slurm.py

@@ -2,9 +2,8 @@
 import os
 
 from coleo import tooled
-from voir.instruments.gpu import get_gpu_info
 
-from ..slurm import expand_node_list
+from ..system import get_gpu_capacity
 
 
 @tooled
@@ -26,18 +25,73 @@ def make_node(i, ip):
 
         return node
 
-    capacity = float("+inf")
-
-    for _, v in get_gpu_info("cuda")["gpus"].items():
-        capacity = min(v["memory"]["total"], capacity)
-
     # nvidia-smi --query-gpu=memory.total --format=csv
     system = {
         "arch": "cuda",
-        "gpu": {"capacity": f"{int(capacity)} MiB"},
         "nodes": [make_node(i, ip) for i, ip in enumerate(node_list)],
     }
 
+    capacity = get_gpu_capacity()
+    if capacity > 0:
+        system["gpu"] = {
+            "capacity": f"{capacity} MiB"
+        }
+
     import yaml
 
     print(yaml.dump({"system": system}))
+
+
+def expand_range(s):
+    numbers = []
+    count = 0
+
+    for i in s.split(","):
+        if "-" not in i:
+            count = len(i)
+            numbers.append(i)
+        else:
+            start, end = i.split("-")
+            count = len(start)
+
+            for n in range(int(start), int(end) + 1):
+                numbers.append(f"{n:0{count}d}")
+
+    return numbers
+
+
+def expand_node_list(node_list):
+    nodes = []
+    s = 0
+
+    while s < len(node_list):
+        if node_list[s] == ",":
+            s += 1
+
+        next = node_list.find(",", s)
+        range_start = node_list.find("[", s)
+        range_end = node_list.find("]", s)
+
+        # Found a range
+        if range_start != -1 and (next == -1 or range_start < next):
+            node_name = node_list[s:range_start]
+
+            range = node_list[range_start + 1 : range_end]
+
+            for i in expand_range(range):
+                nodes.append(f"{node_name}{i}")
+
+            # eat the ]
+            s = range_end + 1
+
+        else:
+            if next == -1:
+                next = len(node_list)
+
+            node_name = node_list[s:next]
+            nodes.append(node_name)
+
+            # eat the ,
+            s = next + 1
+
+    return nodes

milabench/config.py

@@ -26,7 +26,7 @@ def get_run_count():
 
 def get_base_folder():
     config = config_global.get()
-    return XPath(config["dirs"]["base"])
+    return XPath(config["_defaults"]["dirs"]["base"])
 
 def relative_to(pth, cwd):
     pth = XPath(pth).expanduser()