skypilot-nightly 1.0.0.dev20250716__tar.gz → 1.0.0.dev20250717__tar.gz

{skypilot_nightly-1.0.0.dev20250716/sky/setup_files → skypilot_nightly-1.0.0.dev20250717}/MANIFEST.in

@@ -17,3 +17,4 @@ include sky/utils/kubernetes/*
 include sky/server/html/*
 recursive-include sky/dashboard/out *
 include sky/users/*.conf
+include sky/utils/*.sh

{skypilot_nightly-1.0.0.dev20250716/skypilot_nightly.egg-info → skypilot_nightly-1.0.0.dev20250717}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: skypilot-nightly
-Version: 1.0.0.dev20250716
+Version: 1.0.0.dev20250717
 Summary: SkyPilot: Run AI on Any Infra — Unified, Faster, Cheaper.
 Author: SkyPilot Team
 License: Apache 2.0
@@ -54,6 +54,8 @@ Requires-Dist: sqlalchemy_adapter
 Requires-Dist: prometheus_client>=0.8.0
 Requires-Dist: passlib
 Requires-Dist: pyjwt
+Requires-Dist: gitpython
+Requires-Dist: types-paramiko
 Provides-Extra: aws
 Requires-Dist: awscli>=1.27.10; extra == "aws"
 Requires-Dist: botocore>=1.29.10; extra == "aws"

{skypilot_nightly-1.0.0.dev20250716 → skypilot_nightly-1.0.0.dev20250717}/sky/__init__.py

@@ -5,7 +5,7 @@ from typing import Optional
 import urllib.request
 
 # Replaced with the current commit when building the wheels.
-_SKYPILOT_COMMIT_SHA = '41c25f4073c1d5024d415c63f8e7d584fbb5517a'
+_SKYPILOT_COMMIT_SHA = '5ed1775c386bbd0cc4b9d1c80fc0d2d91b352870'
 
 
 def _get_git_commit():
@@ -35,7 +35,7 @@ def _get_git_commit():
 
 
 __commit__ = _get_git_commit()
-__version__ = '1.0.0.dev20250716'
+__version__ = '1.0.0.dev20250717'
 __root_dir__ = os.path.dirname(os.path.abspath(__file__))
 
 
@@ -88,6 +88,7 @@ from sky.admin_policy import UserRequest
 from sky.catalog import list_accelerators
 from sky.client.sdk import api_cancel
 from sky.client.sdk import api_info
+from sky.client.sdk import api_login
 from sky.client.sdk import api_server_logs
 from sky.client.sdk import api_start
 from sky.client.sdk import api_status
@@ -206,6 +207,7 @@ __all__ = [
     'api_status',
     'api_cancel',
     'api_info',
+    'api_login',
     'api_start',
     'api_stop',
     'api_server_logs',

{skypilot_nightly-1.0.0.dev20250716 → skypilot_nightly-1.0.0.dev20250717}/sky/backends/backend.py

@@ -1,6 +1,6 @@
 """Sky backend interface."""
 import typing
-from typing import Dict, Generic, Optional, Tuple
+from typing import Any, Dict, Generic, Optional, Tuple, Union
 
 from sky.usage import usage_lib
 from sky.utils import cluster_utils
@@ -90,8 +90,10 @@ class Backend(Generic[_ResourceHandleType]):
 
     @timeline.event
     @usage_lib.messages.usage.update_runtime('sync_workdir')
-    def sync_workdir(self, handle: _ResourceHandleType, workdir: Path) -> None:
-        return self._sync_workdir(handle, workdir)
+    def sync_workdir(self, handle: _ResourceHandleType,
+                     workdir: Union[Path, Dict[str, Any]],
+                     envs_and_secrets: Dict[str, str]) -> None:
+        return self._sync_workdir(handle, workdir, envs_and_secrets)
 
     @timeline.event
     @usage_lib.messages.usage.update_runtime('sync_file_mounts')
@@ -165,7 +167,9 @@ class Backend(Generic[_ResourceHandleType]):
     ) -> Tuple[Optional[_ResourceHandleType], bool]:
         raise NotImplementedError
 
-    def _sync_workdir(self, handle: _ResourceHandleType, workdir: Path) -> None:
+    def _sync_workdir(self, handle: _ResourceHandleType,
+                      workdir: Union[Path, Dict[str, Any]],
+                      envs_and_secrets: Dict[str, str]) -> None:
         raise NotImplementedError
 
     def _sync_file_mounts(

{skypilot_nightly-1.0.0.dev20250716 → skypilot_nightly-1.0.0.dev20250717}/sky/backends/cloud_vm_ray_backend.py

@@ -3240,10 +3240,59 @@ class CloudVmRayBackend(backends.Backend['CloudVmRayResourceHandle']):
             common_utils.remove_file_if_exists(lock_path)
 
     def _sync_workdir(self, handle: CloudVmRayResourceHandle,
-                      workdir: Path) -> None:
+                      workdir: Union[Path, Dict[str, Any]],
+                      envs_and_secrets: Dict[str, str]) -> None:
         # Even though provision() takes care of it, there may be cases where
         # this function is called in isolation, without calling provision(),
         # e.g., in CLI.  So we should rerun rsync_up.
+        if isinstance(workdir, dict):
+            self._sync_git_workdir(handle, envs_and_secrets)
+        else:
+            self._sync_path_workdir(handle, workdir)
+
+    def _sync_git_workdir(self, handle: CloudVmRayResourceHandle,
+                          envs_and_secrets: Dict[str, str]) -> None:
+        style = colorama.Style
+        ip_list = handle.external_ips()
+        assert ip_list is not None, 'external_ips is not cached in handle'
+
+        log_path = os.path.join(self.log_dir, 'workdir_sync.log')
+
+        # TODO(zhwu): refactor this with backend_utils.parallel_cmd_with_rsync
+        runners = handle.get_command_runners()
+
+        def _sync_git_workdir_node(
+                runner: command_runner.CommandRunner) -> None:
+            # Type assertion to help mypy understand the type
+            assert hasattr(
+                runner, 'git_clone'
+            ), f'CommandRunner should have git_clone method, ' \
+                f'got {type(runner)}'
+            runner.git_clone(
+                target_dir=SKY_REMOTE_WORKDIR,
+                log_path=log_path,
+                stream_logs=False,
+                max_retry=3,
+                envs_and_secrets=envs_and_secrets,
+            )
+
+        num_nodes = handle.launched_nodes
+        plural = 's' if num_nodes > 1 else ''
+        logger.info(
+            f'  {style.DIM}Syncing workdir (to {num_nodes} node{plural}): '
+            f'{SKY_REMOTE_WORKDIR}{style.RESET_ALL}')
+        os.makedirs(os.path.expanduser(self.log_dir), exist_ok=True)
+        os.system(f'touch {log_path}')
+        num_threads = subprocess_utils.get_parallel_threads(
+            str(handle.launched_resources.cloud))
+        with rich_utils.safe_status(
+                ux_utils.spinner_message('Syncing workdir', log_path)):
+            subprocess_utils.run_in_parallel(_sync_git_workdir_node, runners,
+                                             num_threads)
+        logger.info(ux_utils.finishing_message('Synced workdir.', log_path))
+
+    def _sync_path_workdir(self, handle: CloudVmRayResourceHandle,
+                           workdir: Path) -> None:
         fore = colorama.Fore
         style = colorama.Style
         ip_list = handle.external_ips()

{skypilot_nightly-1.0.0.dev20250716 → skypilot_nightly-1.0.0.dev20250717}/sky/backends/docker_utils.py

@@ -168,7 +168,7 @@ def build_dockerimage(task: task_mod.Task,
                                         build_dir=temp_dir)
 
     dst = os.path.join(temp_dir, SKY_DOCKER_WORKDIR)
-    if task.workdir is not None:
+    if task.workdir is not None and isinstance(task.workdir, str):
         # Copy workdir contents to tempdir
         shutil.copytree(os.path.expanduser(task.workdir), dst)
     else:

{skypilot_nightly-1.0.0.dev20250716 → skypilot_nightly-1.0.0.dev20250717}/sky/backends/local_docker_backend.py

@@ -178,7 +178,8 @@ class LocalDockerBackend(backends.Backend['LocalDockerResourceHandle']):
         return handle, False
 
     def _sync_workdir(self, handle: LocalDockerResourceHandle,
-                      workdir: Path) -> None:
+                      workdir: Union[Path, Dict[str, Any]],
+                      envs_and_secrets: Dict[str, str]) -> None:
         """Workdir is sync'd by adding to the docker image.
 
         This happens in the execute step.

{skypilot_nightly-1.0.0.dev20250716 → skypilot_nightly-1.0.0.dev20250717}/sky/catalog/common.py

@@ -13,6 +13,7 @@ from sky import sky_logging
 from sky.adaptors import common as adaptors_common
 from sky.clouds import cloud as cloud_lib
 from sky.skylet import constants
+from sky.utils import annotations
 from sky.utils import common_utils
 from sky.utils import registry
 from sky.utils import rich_utils
@@ -125,17 +126,21 @@ class LazyDataFrame:
 
     We don't need to load the catalog for every SkyPilot call, and this class
     allows us to load the catalog only when needed.
+
+    Use update_if_stale_func to pass in a function that decides whether to
+    update the catalog on disk, updates it if needed, and returns
+    a bool indicating whether the update was done.
     """
 
-    def __init__(self, filename: str, update_func: Callable[[], None]):
+    def __init__(self, filename: str, update_if_stale_func: Callable[[], bool]):
         self._filename = filename
         self._df: Optional['pd.DataFrame'] = None
-        self._update_func = update_func
+        self._update_if_stale_func = update_if_stale_func
 
+    @annotations.lru_cache(scope='request')
     def _load_df(self) -> 'pd.DataFrame':
-        if self._df is None:
+        if self._update_if_stale_func() or self._df is None:
             try:
-                self._update_func()
                 self._df = pd.read_csv(self._filename)
             except Exception as e:  # pylint: disable=broad-except
                 # As users can manually modify the catalog, read_csv can fail.
@@ -193,55 +198,60 @@ def read_catalog(filename: str,
         return last_update + pull_frequency_hours * 3600 < time.time()
 
     def _update_catalog():
+        # Fast path: Exit early to avoid lock contention.
+        if not _need_update():
+            return False
+
         # Atomic check, to avoid conflicts with other processes.
         with filelock.FileLock(meta_path + '.lock'):
-            if _need_update():
-                url = f'{constants.HOSTED_CATALOG_DIR_URL}/{constants.CATALOG_SCHEMA_VERSION}/{filename}'  # pylint: disable=line-too-long
-                url_fallback = f'{constants.HOSTED_CATALOG_DIR_URL_S3_MIRROR}/{constants.CATALOG_SCHEMA_VERSION}/{filename}'  # pylint: disable=line-too-long
-                headers = {'User-Agent': 'SkyPilot/0.7'}
-                update_frequency_str = ''
-                if pull_frequency_hours is not None:
-                    update_frequency_str = (
-                        f' (every {pull_frequency_hours} hours)')
-                with rich_utils.safe_status(
-                        ux_utils.spinner_message(
-                            f'Updating {cloud} catalog: {filename}') +
-                        f'{update_frequency_str}'):
-                    try:
-                        r = requests.get(url=url, headers=headers)
-                        if r.status_code == 429:
-                            # fallback to s3 mirror, github introduced rate
-                            # limit after 2025-05, see
-                            # https://github.com/skypilot-org/skypilot/issues/5438
-                            # for more details
-                            r = requests.get(url=url_fallback, headers=headers)
-                        r.raise_for_status()
-                    except requests.exceptions.RequestException as e:
-                        error_str = (f'Failed to fetch {cloud} catalog '
-                                     f'{filename}. ')
-                        if os.path.exists(catalog_path):
-                            logger.warning(
-                                f'{error_str}Using cached catalog files.')
-                            # Update catalog file modification time.
-                            os.utime(catalog_path, None)  # Sets to current time
-                        else:
-                            logger.error(
-                                f'{error_str}Please check your internet '
-                                'connection.')
-                            with ux_utils.print_exception_no_traceback():
-                                raise e
+            # Double check after acquiring the lock.
+            if not _need_update():
+                return False
+
+            url = f'{constants.HOSTED_CATALOG_DIR_URL}/{constants.CATALOG_SCHEMA_VERSION}/{filename}'  # pylint: disable=line-too-long
+            url_fallback = f'{constants.HOSTED_CATALOG_DIR_URL_S3_MIRROR}/{constants.CATALOG_SCHEMA_VERSION}/{filename}'  # pylint: disable=line-too-long
+            headers = {'User-Agent': 'SkyPilot/0.7'}
+            update_frequency_str = ''
+            if pull_frequency_hours is not None:
+                update_frequency_str = (
+                    f' (every {pull_frequency_hours} hours)')
+            with rich_utils.safe_status(
+                    ux_utils.spinner_message(
+                        f'Updating {cloud} catalog: {filename}') +
+                    f'{update_frequency_str}'):
+                try:
+                    r = requests.get(url=url, headers=headers)
+                    if r.status_code == 429:
+                        # fallback to s3 mirror, github introduced rate
+                        # limit after 2025-05, see
+                        # https://github.com/skypilot-org/skypilot/issues/5438
+                        # for more details
+                        r = requests.get(url=url_fallback, headers=headers)
+                    r.raise_for_status()
+                except requests.exceptions.RequestException as e:
+                    error_str = (f'Failed to fetch {cloud} catalog '
+                                 f'{filename}. ')
+                    if os.path.exists(catalog_path):
+                        logger.warning(
+                            f'{error_str}Using cached catalog files.')
+                        # Update catalog file modification time.
+                        os.utime(catalog_path, None)  # Sets to current time
                     else:
-                        # Download successful, save the catalog to a local file.
-                        os.makedirs(os.path.dirname(catalog_path),
-                                    exist_ok=True)
-                        with open(catalog_path, 'w', encoding='utf-8') as f:
-                            f.write(r.text)
-                        with open(meta_path + '.md5', 'w',
-                                  encoding='utf-8') as f:
-                            f.write(hashlib.md5(r.text.encode()).hexdigest())
-                logger.debug(f'Updated {cloud} catalog {filename}.')
-
-    return LazyDataFrame(catalog_path, update_func=_update_catalog)
+                        logger.error(f'{error_str}Please check your internet '
+                                     'connection.')
+                        with ux_utils.print_exception_no_traceback():
+                            raise e
+                else:
+                    # Download successful, save the catalog to a local file.
+                    os.makedirs(os.path.dirname(catalog_path), exist_ok=True)
+                    with open(catalog_path, 'w', encoding='utf-8') as f:
+                        f.write(r.text)
+                    with open(meta_path + '.md5', 'w', encoding='utf-8') as f:
+                        f.write(hashlib.md5(r.text.encode()).hexdigest())
+            logger.debug(f'Updated {cloud} catalog {filename}.')
+        return True
+
+    return LazyDataFrame(catalog_path, update_if_stale_func=_update_catalog)
 
 
 def _get_instance_type(

{skypilot_nightly-1.0.0.dev20250716 → skypilot_nightly-1.0.0.dev20250717}/sky/catalog/data_fetchers/fetch_gcp.py

@@ -198,6 +198,7 @@ SERIES_TO_DESCRIPTION = {
     'n1': 'N1 Predefined Instance',
     'n2': 'N2 Instance',
     'n2d': 'N2D AMD Instance',
+    'n4': 'N4 Instance',
     't2a': 'T2A Arm Instance',
     't2d': 'T2D AMD Instance',
 }

{skypilot_nightly-1.0.0.dev20250716 → skypilot_nightly-1.0.0.dev20250717}/sky/catalog/gcp_catalog.py

@@ -37,20 +37,37 @@ _image_df = common.read_catalog('gcp/images.csv',
 _quotas_df = common.read_catalog('gcp/accelerator_quota_mapping.csv',
                                  pull_frequency_hours=_PULL_FREQUENCY_HOURS)
 
-# We will select from the following three CPU instance families:
+# We will select from the following six CPU instance families:
 _DEFAULT_INSTANCE_FAMILY = [
-    # This is the latest general-purpose instance family as of Mar 2023.
-    # CPU: Intel Ice Lake 8373C or Cascade Lake 6268CL.
+    # This is a widely used general-purpose instance family as of July 2025.
+    # CPU: Primarily Intel Ice Lake (3rd Gen Intel Xeon Scalable Processors)
+    #  or Cascade Lake (2nd Gen Intel Xeon Scalable Processors).
     # Memory: 4 GiB RAM per 1 vCPU;
     'n2-standard',
-    # This is the latest memory-optimized instance family as of Mar 2023.
-    # CPU: Intel Ice Lake 8373C or Cascade Lake 6268CL.
+    # This is a memory-optimized instance family as of July 2025.
+    # CPU: Primarily Intel Ice Lake (3rd Gen Intel Xeon Scalable Processors)
+    # or Cascade Lake (2nd Gen Intel Xeon Scalable Processors).
     # Memory: 8 GiB RAM per 1 vCPU;
     'n2-highmem',
-    # This is the latest compute-optimized instance family as of Mar 2023.
-    # CPU: Intel Ice Lake 8373C or Cascade Lake 6268CL.
+    # This is a compute-optimized instance family as of July 2025.
+    # CPU: Primarily Intel Ice Lake (3rd Gen Intel Xeon Scalable Processors)
+    #  or Cascade Lake (2nd Gen Intel Xeon Scalable Processors).
     # Memory: 1 GiB RAM per 1 vCPU;
     'n2-highcpu',
+    # This is the latest general-purpose instance family as of July 2025.
+    # CPU: Intel 5th Gen Xeon Scalable processor (Emerald Rapids).
+    # Memory: 4 GiB RAM per 1 vCPU;
+    'n4-standard',
+    # This is the latest general-purpose instance family
+    # with a higher vCPU to memory ratio as of July 2025.
+    # CPU: Intel 5th Gen Xeon Scalable processor (Emerald Rapids).
+    # Memory: 2 GiB RAM per 1 vCPU;
+    'n4-highcpu',
+    # This is the latest general-purpose instance family
+    # with a lower vCPU to memory ratio as of July 2025.
+    # CPU: Intel 5th Gen Xeon Scalable processor (Emerald Rapids).
+    # Memory: 8 GiB RAM per 1 vCPU;
+    'n4-highmem',
 ]
 # n2 is not allowed for launching GPUs for now.
 _DEFAULT_HOST_VM_FAMILY = (

{skypilot_nightly-1.0.0.dev20250716 → skypilot_nightly-1.0.0.dev20250717}/sky/catalog/kubernetes_catalog.py

@@ -195,6 +195,10 @@ def _list_accelerators(
                 accelerator_name = lf.get_accelerator_from_label_value(
                     node.metadata.labels.get(key))
 
+                # Heterogenous cluster may have some nodes with empty labels.
+                if not accelerator_name:
+                    continue
+
                 # Exclude multi-host TPUs from being processed.
                 # TODO(Doyoung): Remove the logic when adding support for
                 # multi-host TPUs.
@@ -212,7 +216,7 @@ def _list_accelerators(
                     kubernetes_utils.get_node_accelerator_count(
                         node.status.allocatable))
 
-                if accelerator_name and accelerator_count > 0:
+                if accelerator_count > 0:
                     # TPUs are counted in a different way compared to GPUs.
                     # Multi-node GPUs can be split into smaller units and be
                     # provisioned, but TPUs are considered as an atomic unit.