Merge branch 'main' of github.com:SpareCores/sc-inspector

Author: daroczig

inspector/inspector.py

@@ -38,9 +38,12 @@
 EXCLUDE_INSTANCES: list[tuple[str, str]] = [
     ("aws", "f2.48xlarge"),
     ("aws", "p4de.24xlarge"),
+    ("aws", "p5.4xlarge"),
     ("aws", "p5.48xlarge"),
     ("aws", "p5e.48xlarge"),
     ("aws", "p5en.48xlarge"),
+    ("aws", "p6-b200.48xlarge"),
+    ("aws", "p6-b300.48xlarge"),
     ("aws", "trn2.48xlarge"),
     ("aws", "u7i-12tb.224xlarge"),
     ("aws", "u7i-8tb.112xlarge"),

inspector/lib.py

@@ -47,6 +47,7 @@
     re.compile(r"creating .* error"),  # Azure errors
     re.compile(r"creating failed"),  # Azure errors
     re.compile(r"error waiting for.*to create"),   # GCP error
+    re.compile(r"Unable to create server"),  # Upcloud error
 }
 # provision machines with storage (GiB)
 VOLUME_SIZE = 128

inspector/repo.py

@@ -36,7 +36,7 @@ def add_token_auth(url: str, token: str) -> str:
     parsed = urlparse(url)
     # works no matter if the original url had a user/pass or not
     domain = parsed.netloc.split("@")[-1]
-    domain = f"{token}@{domain}"
+    domain = f"user:{token}@{domain}"
     unparsed = (parsed[0], domain, parsed[2], parsed[3], parsed[4], parsed[5])
     return urlunparse(unparsed)
 

inspector/tasks.py

@@ -195,7 +195,7 @@
 llm = DockerTask(
     parallel=False,
     # might be slow when testing large models
-    timeout=timedelta(hours=1),
+    timeout=timedelta(hours=1.5),
     minimum_memory=1,
     priority=11,
     image="ghcr.io/sparecores/benchmark-llm:main",