skypilot-nightly 1.0.0.dev20251013__py3-none-any.whl → 1.0.0.dev20251015__py3-none-any.whl

sky/skylet/constants.py

@@ -226,7 +226,9 @@ RAY_INSTALLATION_COMMANDS = (
     f'{SKY_UV_PIP_CMD} list | grep "ray " | '
     f'grep {SKY_REMOTE_RAY_VERSION} 2>&1 > /dev/null '
     f'|| {RAY_STATUS} || '
-    f'{SKY_UV_PIP_CMD} install -U ray[default]=={SKY_REMOTE_RAY_VERSION}; '  # pylint: disable=line-too-long
+    # The pydantic-core==2.41.3 for arm seems corrupted
+    # so we need to avoid that specific version.
+    f'{SKY_UV_PIP_CMD} install -U "ray[default]=={SKY_REMOTE_RAY_VERSION}" "pydantic-core==2.41.1"; '  # pylint: disable=line-too-long
     # In some envs, e.g. pip does not have permission to write under /opt/conda
     # ray package will be installed under ~/.local/bin. If the user's PATH does
     # not include ~/.local/bin (the pip install will have the output: `WARNING:
@@ -402,10 +404,25 @@ OVERRIDEABLE_CONFIG_KEYS_IN_TASK: List[Tuple[str, ...]] = [
 ]
 # When overriding the SkyPilot configs on the API server with the client one,
 # we skip the following keys because they are meant to be client-side configs.
-SKIPPED_CLIENT_OVERRIDE_KEYS: List[Tuple[str, ...]] = [('api_server',),
-                                                       ('allowed_clouds',),
-                                                       ('workspaces',), ('db',),
-                                                       ('daemons',)]
+# Also, we skip the consolidation mode config as those should be only set on
+# the API server side.
+SKIPPED_CLIENT_OVERRIDE_KEYS: List[Tuple[str, ...]] = [
+    ('api_server',),
+    ('allowed_clouds',),
+    ('workspaces',),
+    ('db',),
+    ('daemons',),
+    # TODO(kevin,tian): Override the whole controller config once our test
+    # infrastructure supports setting dynamic server side configs.
+    # Tests that are affected:
+    # - test_managed_jobs_ha_kill_starting
+    # - test_managed_jobs_ha_kill_running
+    # - all tests that use LOW_CONTROLLER_RESOURCE_ENV or
+    #   LOW_CONTROLLER_RESOURCE_OVERRIDE_CONFIG (won't cause test failure,
+    #   but the configs won't be applied)
+    ('jobs', 'controller', 'consolidation_mode'),
+    ('serve', 'controller', 'consolidation_mode'),
+]
 
 # Constants for Azure blob storage
 WAIT_FOR_STORAGE_ACCOUNT_CREATION = 60

sky/skylet/log_lib.py

@@ -271,7 +271,6 @@ def run_with_log(
                 stdout, stderr = context_utils.pipe_and_wait_process(
                     ctx,
                     proc,
-                    cancel_callback=subprocess_utils.kill_children_processes,
                     stdout_stream_handler=stdout_stream_handler,
                     stderr_stream_handler=stderr_stream_handler)
             elif process_stream:

sky/skylet/log_lib.pyi

@@ -42,7 +42,7 @@ class _ProcessingArgs:
         ...
 
 
-def _get_context() -> Optional[context.Context]:
+def _get_context() -> Optional[context.SkyPilotContext]:
     ...
 
 

sky/utils/asyncio_utils.py

@@ -0,0 +1,18 @@
+"""Asyncio utilities."""
+
+import asyncio
+import functools
+
+
+def shield(func):
+    """Shield the decorated async function from cancellation.
+
+    Note that filelock.AsyncFileLock is not cancellation safe, thus the
+    function calls filelock.AsyncFileLock must be shielded.
+    """
+
+    @functools.wraps(func)
+    async def async_wrapper(*args, **kwargs):
+        return await asyncio.shield(func(*args, **kwargs))
+
+    return async_wrapper

sky/utils/common.py

@@ -42,6 +42,8 @@ def refresh_server_id() -> None:
     JOB_CONTROLLER_NAME = f'{JOB_CONTROLLER_PREFIX}{SERVER_ID}'
 
 
+# TODO(kevin): Remove this side effect and have callers call
+# refresh_server_id() explicitly as needed.
 refresh_server_id()
 
 

sky/utils/context.py

@@ -5,13 +5,12 @@ from collections.abc import Mapping
 import contextvars
 import copy
 import functools
-import inspect
 import os
 import pathlib
 import subprocess
 import sys
-from typing import (Callable, Dict, Iterator, MutableMapping, Optional, TextIO,
-                    TYPE_CHECKING, TypeVar)
+from typing import (Any, Callable, Coroutine, Dict, Iterator, MutableMapping,
+                    Optional, TextIO, TYPE_CHECKING, TypeVar)
 
 from typing_extensions import ParamSpec
 
@@ -19,7 +18,7 @@ if TYPE_CHECKING:
     from sky.skypilot_config import ConfigContext
 
 
-class Context(object):
+class SkyPilotContext(object):
     """SkyPilot typed context vars for threads and coroutines.
 
     This is a wrapper around `contextvars.ContextVar` that provides a typed
@@ -114,7 +113,14 @@ class Context(object):
             self._log_file_handle.close()
             self._log_file_handle = None
 
-    def copy(self) -> 'Context':
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        del exc_type, exc_val, exc_tb
+        self.cleanup()
+
+    def copy(self) -> 'SkyPilotContext':
         """Create a copy of the context.
 
         Changes to the current context after this call will not affect the copy.
@@ -123,18 +129,18 @@ class Context(object):
         The new context will get an independent copy of the config context.
         Cancellation of the current context will not be propagated to the copy.
         """
-        new_context = Context()
+        new_context = SkyPilotContext()
         new_context.redirect_log(self._log_file)
         new_context.env_overrides = self.env_overrides.copy()
         new_context.config_context = copy.deepcopy(self.config_context)
         return new_context
 
 
-_CONTEXT = contextvars.ContextVar[Optional[Context]]('sky_context',
-                                                     default=None)
+_CONTEXT = contextvars.ContextVar[Optional[SkyPilotContext]]('sky_context',
+                                                             default=None)
 
 
-def get() -> Optional[Context]:
+def get() -> Optional[SkyPilotContext]:
     """Get the current SkyPilot context.
 
     If the context is not initialized, get() will return None. This helps
@@ -200,7 +206,7 @@ class ContextualEnviron(MutableMapping[str, str]):
 
     def __iter__(self) -> Iterator[str]:
 
-        def iter_from_context(ctx: Context) -> Iterator[str]:
+        def iter_from_context(ctx: SkyPilotContext) -> Iterator[str]:
             deleted_keys = set()
             for key, value in ctx.env_overrides.items():
                 if value is None:
@@ -311,56 +317,56 @@ def contextual(func: Callable[P, T]) -> Callable[P, T]:
     context that inherits the values from the existing context.
     """
 
+    def run_in_context(*args: P.args, **kwargs: P.kwargs) -> T:
+        # Within the new contextvars Context, set up the SkyPilotContext.
+        original_ctx = get()
+        with initialize(original_ctx):
+            return func(*args, **kwargs)
+
     @functools.wraps(func)
     def wrapper(*args: P.args, **kwargs: P.kwargs) -> T:
+        # Create a copy of the current contextvars Context so that setting the
+        # SkyPilotContext does not affect the caller's context in async
+        # environments.
+        context = contextvars.copy_context()
+        return context.run(run_in_context, *args, **kwargs)
+
+    return wrapper
+
+
+def contextual_async(
+    func: Callable[P, Coroutine[Any, Any, T]]
+) -> Callable[P, Coroutine[Any, Any, T]]:
+    """Decorator to initialize a context before executing the function.
+
+    If a context is already initialized, this decorator will create a new
+    context that inherits the values from the existing context.
+    """
+
+    async def run_in_context(*args: P.args, **kwargs: P.kwargs) -> T:
+        # Within the new contextvars Context, set up the SkyPilotContext.
         original_ctx = get()
-        initialize(original_ctx)
-        ctx = get()
-        cleanup_after_await = False
-
-        def cleanup():
-            try:
-                if ctx is not None:
-                    ctx.cleanup()
-            finally:
-                # Note: _CONTEXT.reset() is not reliable - may fail with
-                # ValueError: <Token ... at ...> was created in a different
-                # Context
-                # We must make sure this happens because otherwise we may try to
-                # write to the wrong log.
-                _CONTEXT.set(original_ctx)
-
-        # There are two cases:
-        # 1. The function is synchronous (that is, return type is not awaitable)
-        #    In this case, we use a finally block to cleanup the context.
-        # 2. The function is asynchronous (that is, return type is awaitable)
-        #    In this case, we need to construct an async def wrapper and await
-        #    the value, then call the cleanup function in the finally block.
-
-        async def await_with_cleanup(awaitable):
-            try:
-                return await awaitable
-            finally:
-                cleanup()
-
-        try:
-            ret = func(*args, **kwargs)
-            if inspect.isawaitable(ret):
-                cleanup_after_await = True
-                return await_with_cleanup(ret)
-            else:
-                return ret
-        finally:
-            if not cleanup_after_await:
-                cleanup()
+        with initialize(original_ctx):
+            return await func(*args, **kwargs)
+
+    @functools.wraps(func)
+    async def wrapper(*args: P.args, **kwargs: P.kwargs) -> T:
+        # Create a copy of the current contextvars Context so that setting the
+        # SkyPilotContext does not affect the caller's context in async
+        # environments.
+        context = contextvars.copy_context()
+        return await context.run(run_in_context, *args, **kwargs)
 
     return wrapper
 
 
-def initialize(base_context: Optional[Context] = None) -> None:
+def initialize(
+        base_context: Optional[SkyPilotContext] = None) -> SkyPilotContext:
     """Initialize the current SkyPilot context."""
-    new_context = base_context.copy() if base_context is not None else Context()
+    new_context = base_context.copy(
+    ) if base_context is not None else SkyPilotContext()
     _CONTEXT.set(new_context)
+    return new_context
 
 
 class _ContextualStream:

sky/utils/context_utils.py

@@ -60,7 +60,7 @@ def passthrough_stream_handler(in_stream: IO[Any], out_stream: IO[Any]) -> str:
 
 
 def pipe_and_wait_process(
-        ctx: context.Context,
+        ctx: context.SkyPilotContext,
         proc: subprocess.Popen,
         poll_interval: float = 0.5,
         cancel_callback: Optional[Callable[[], None]] = None,
@@ -113,7 +113,7 @@ def pipe_and_wait_process(
         return stdout, stderr
 
 
-def wait_process(ctx: context.Context,
+def wait_process(ctx: context.SkyPilotContext,
                  proc: subprocess.Popen,
                  poll_interval: float = 0.5,
                  cancel_callback: Optional[Callable[[], None]] = None):

sky/utils/controller_utils.py

@@ -72,7 +72,8 @@ class _ControllerSpec:
     """Spec for skypilot controllers."""
     controller_type: str
     name: str
-    cluster_name: str
+    _cluster_name_func: Callable[[], str]
+    _cluster_name_from_server: Optional[str]  # For client-side only
     in_progress_hint: Callable[[bool], str]
     decline_cancel_hint: str
     _decline_down_when_failed_to_fetch_status_hint: str
@@ -93,6 +94,24 @@ class _ControllerSpec:
         return self._check_cluster_name_hint.format(
             cluster_name=self.cluster_name)
 
+    @property
+    def cluster_name(self) -> str:
+        """The cluster name of the controller.
+
+        On the server-side, the cluster name is the actual cluster name,
+        which is read from common.(JOB|SKY_SERVE)_CONTROLLER_NAME.
+
+        On the client-side, the cluster name may not be accurate,
+        as we may not know the exact name, because we are missing
+        the server-side common.SERVER_ID. We have to wait until
+        we get the actual cluster name from the server.
+        """
+        return (self._cluster_name_from_server if self._cluster_name_from_server
+                is not None else self._cluster_name_func())
+
+    def set_cluster_name_from_server(self, cluster_name: str) -> None:
+        self._cluster_name_from_server = cluster_name
+
 
 # TODO: refactor controller class to not be an enum.
 class Controllers(enum.Enum):
@@ -102,7 +121,8 @@ class Controllers(enum.Enum):
     JOBS_CONTROLLER = _ControllerSpec(
         controller_type='jobs',
         name='managed jobs controller',
-        cluster_name=common.JOB_CONTROLLER_NAME,
+        _cluster_name_func=lambda: common.JOB_CONTROLLER_NAME,
+        _cluster_name_from_server=None,
         in_progress_hint=lambda _:
         ('* {job_info}To see all managed jobs: '
          f'{colorama.Style.BRIGHT}sky jobs queue{colorama.Style.RESET_ALL}'),
@@ -133,7 +153,8 @@ class Controllers(enum.Enum):
     SKY_SERVE_CONTROLLER = _ControllerSpec(
         controller_type='serve',
         name='serve controller',
-        cluster_name=common.SKY_SERVE_CONTROLLER_NAME,
+        _cluster_name_func=lambda: common.SKY_SERVE_CONTROLLER_NAME,
+        _cluster_name_from_server=None,
         in_progress_hint=(
             lambda pool:
             (f'* To see detailed pool status: {colorama.Style.BRIGHT}'
@@ -166,7 +187,9 @@ class Controllers(enum.Enum):
         default_autostop_config=serve_constants.CONTROLLER_AUTOSTOP)
 
     @classmethod
-    def from_name(cls, name: Optional[str]) -> Optional['Controllers']:
+    def from_name(cls,
+                  name: Optional[str],
+                  expect_exact_match: bool = True) -> Optional['Controllers']:
         """Check if the cluster name is a controller name.
 
         Returns:
@@ -187,7 +210,11 @@ class Controllers(enum.Enum):
         elif name.startswith(common.JOB_CONTROLLER_PREFIX):
             controller = cls.JOBS_CONTROLLER
             prefix = common.JOB_CONTROLLER_PREFIX
-        if controller is not None and name != controller.value.cluster_name:
+
+        if controller is not None and expect_exact_match:
+            assert name == controller.value.cluster_name, (
+                name, controller.value.cluster_name)
+        elif controller is not None and name != controller.value.cluster_name:
             # The client-side cluster_name is not accurate. Assume that `name`
             # is the actual cluster name, so need to set the controller's
             # cluster name to the input name.
@@ -201,7 +228,7 @@ class Controllers(enum.Enum):
                                                                          prefix)
 
             # Update the cluster name.
-            controller.value.cluster_name = name
+            controller.value.set_cluster_name_from_server(name)
         return controller
 
     @classmethod
@@ -228,7 +255,7 @@ def get_controller_for_pool(pool: bool) -> Controllers:
 def high_availability_specified(cluster_name: Optional[str]) -> bool:
     """Check if the controller high availability is specified in user config.
     """
-    controller = Controllers.from_name(cluster_name)
+    controller = Controllers.from_name(cluster_name, expect_exact_match=False)
     if controller is None:
         return False
 
@@ -411,7 +438,7 @@ def check_cluster_name_not_controller(
     Returns:
       None, if the cluster name is not a controller name.
     """
-    controller = Controllers.from_name(cluster_name)
+    controller = Controllers.from_name(cluster_name, expect_exact_match=False)
     if controller is not None:
         msg = controller.value.check_cluster_name_hint
         if operation_str is not None:

sky/utils/locks.py

@@ -243,6 +243,7 @@ class PostgresLock(DistributedLock):
         if not self._acquired or not self._connection:
             return
 
+        connection_lost = False
         try:
             cursor = self._connection.cursor()
             cursor.execute('SELECT pg_advisory_unlock(%s)', (self._lock_key,))
@@ -252,8 +253,11 @@ class PostgresLock(DistributedLock):
             # Lost connection to the database, likely the lock is force unlocked
             # by other routines.
             logger.debug(f'Failed to release postgres lock {self.lock_id}: {e}')
+            connection_lost = True
         finally:
-            self._close_connection()
+            # Invalidate if connection was lost to prevent SQLAlchemy from
+            # trying to reset a dead connection
+            self._close_connection(invalidate=connection_lost)
 
     def force_unlock(self) -> None:
         """Force unlock the postgres advisory lock."""
@@ -292,13 +296,24 @@ class PostgresLock(DistributedLock):
         finally:
             self._close_connection()
 
-    def _close_connection(self) -> None:
-        """Close the postgres connection."""
+    def _close_connection(self, invalidate: bool = False) -> None:
+        """Close the postgres connection.
+
+        Args:
+            invalidate: If True, invalidate connection instead of closing it.
+                Use this when the connection might be broken (e.g., after
+                pg_terminate_backend) to prevent SQLAlchemy from trying to
+                reset it (which would result in an error being logged).
+        """
         if self._connection:
             try:
-                self._connection.close()
+                if invalidate:
+                    self._connection.invalidate()
+                else:
+                    self._connection.close()
             except Exception as e:  # pylint: disable=broad-except
-                logger.debug(f'Failed to close postgres connection: {e}')
+                logger.debug(
+                    f'Failed to invalidate or close postgres connection: {e}')
             self._connection = None
 
     def is_locked(self) -> bool:

sky/utils/subprocess_utils.py

@@ -10,7 +10,8 @@ import sys
 import threading
 import time
 import typing
-from typing import Any, Callable, Dict, List, Optional, Protocol, Tuple, Union
+from typing import (Any, Callable, Dict, List, Optional, Protocol, Set, Tuple,
+                    Union)
 
 import colorama
 
@@ -107,7 +108,7 @@ def get_parallel_threads(cloud_str: Optional[str] = None) -> int:
 
 
 def run_in_parallel(func: Callable,
-                    args: List[Any],
+                    args: Union[List[Any], Set[Any]],
                     num_threads: Optional[int] = None) -> List[Any]:
     """Run a function in parallel on a list of arguments.
 
@@ -128,7 +129,7 @@ def run_in_parallel(func: Callable,
     if len(args) == 0:
         return []
     if len(args) == 1:
-        return [func(args[0])]
+        return [func(list(args)[0])]
 
     processes = (num_threads
                  if num_threads is not None else get_parallel_threads())

{skypilot_nightly-1.0.0.dev20251013.dist-info → skypilot_nightly-1.0.0.dev20251015.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: skypilot-nightly
-Version: 1.0.0.dev20251013
+Version: 1.0.0.dev20251015
 Summary: SkyPilot: Run AI on Any Infra — Unified, Faster, Cheaper.
 Author: SkyPilot Team
 License: Apache 2.0
@@ -156,51 +156,51 @@ Requires-Dist: aiosqlite; extra == "server"
 Requires-Dist: greenlet; extra == "server"
 Provides-Extra: shadeform
 Provides-Extra: all
-Requires-Dist: python-dateutil; extra == "all"
-Requires-Dist: ibm-platform-services>=0.48.0; extra == "all"
+Requires-Dist: pydo>=0.3.0; extra == "all"
 Requires-Dist: tomli; python_version < "3.11" and extra == "all"
-Requires-Dist: anyio; extra == "all"
-Requires-Dist: azure-storage-blob>=12.23.1; extra == "all"
-Requires-Dist: aiosqlite; extra == "all"
-Requires-Dist: casbin; extra == "all"
+Requires-Dist: ibm-platform-services>=0.48.0; extra == "all"
+Requires-Dist: aiohttp; extra == "all"
+Requires-Dist: vastai-sdk>=0.1.12; extra == "all"
+Requires-Dist: google-cloud-storage; extra == "all"
 Requires-Dist: pyjwt; extra == "all"
-Requires-Dist: azure-cli>=2.65.0; extra == "all"
-Requires-Dist: pyopenssl<24.3.0,>=23.2.0; extra == "all"
-Requires-Dist: passlib; extra == "all"
+Requires-Dist: google-api-python-client>=2.69.0; extra == "all"
+Requires-Dist: websockets; extra == "all"
+Requires-Dist: protobuf<7.0.0,>=5.26.1; extra == "all"
+Requires-Dist: kubernetes!=32.0.0,>=20.0.0; extra == "all"
+Requires-Dist: msgraph-sdk; extra == "all"
+Requires-Dist: runpod>=1.6.1; extra == "all"
 Requires-Dist: azure-core>=1.31.0; extra == "all"
-Requires-Dist: azure-mgmt-network>=27.0.0; extra == "all"
-Requires-Dist: vastai-sdk>=0.1.12; extra == "all"
+Requires-Dist: aiosqlite; extra == "all"
+Requires-Dist: ibm-vpc; extra == "all"
 Requires-Dist: ibm-cloud-sdk-core; extra == "all"
-Requires-Dist: cudo-compute>=0.1.10; extra == "all"
-Requires-Dist: pyvmomi==8.0.1.0.2; extra == "all"
+Requires-Dist: azure-storage-blob>=12.23.1; extra == "all"
+Requires-Dist: sqlalchemy_adapter; extra == "all"
+Requires-Dist: oci; extra == "all"
 Requires-Dist: grpcio>=1.63.0; extra == "all"
-Requires-Dist: boto3>=1.26.1; extra == "all"
-Requires-Dist: nebius>=0.2.47; extra == "all"
-Requires-Dist: greenlet; extra == "all"
-Requires-Dist: azure-common; extra == "all"
-Requires-Dist: botocore>=1.29.10; extra == "all"
 Requires-Dist: azure-mgmt-compute>=33.0.0; extra == "all"
-Requires-Dist: ibm-vpc; extra == "all"
-Requires-Dist: oci; extra == "all"
-Requires-Dist: kubernetes!=32.0.0,>=20.0.0; extra == "all"
-Requires-Dist: azure-identity>=1.19.0; extra == "all"
-Requires-Dist: ibm-cos-sdk; extra == "all"
-Requires-Dist: google-api-python-client>=2.69.0; extra == "all"
 Requires-Dist: colorama<0.4.5; extra == "all"
+Requires-Dist: python-dateutil; extra == "all"
+Requires-Dist: cudo-compute>=0.1.10; extra == "all"
+Requires-Dist: botocore>=1.29.10; extra == "all"
+Requires-Dist: azure-core>=1.24.0; extra == "all"
 Requires-Dist: awscli>=1.27.10; extra == "all"
-Requires-Dist: docker; extra == "all"
-Requires-Dist: aiohttp; extra == "all"
+Requires-Dist: passlib; extra == "all"
 Requires-Dist: msrestazure; extra == "all"
-Requires-Dist: protobuf<7.0.0,>=5.26.1; extra == "all"
-Requires-Dist: ecsapi>=0.2.0; extra == "all"
-Requires-Dist: google-cloud-storage; extra == "all"
-Requires-Dist: azure-core>=1.24.0; extra == "all"
-Requires-Dist: sqlalchemy_adapter; extra == "all"
-Requires-Dist: runpod>=1.6.1; extra == "all"
-Requires-Dist: msgraph-sdk; extra == "all"
 Requires-Dist: ray[default]>=2.6.1; extra == "all"
-Requires-Dist: pydo>=0.3.0; extra == "all"
-Requires-Dist: websockets; extra == "all"
+Requires-Dist: greenlet; extra == "all"
+Requires-Dist: docker; extra == "all"
+Requires-Dist: ecsapi>=0.2.0; extra == "all"
+Requires-Dist: azure-mgmt-network>=27.0.0; extra == "all"
+Requires-Dist: ibm-cos-sdk; extra == "all"
+Requires-Dist: pyopenssl<24.3.0,>=23.2.0; extra == "all"
+Requires-Dist: casbin; extra == "all"
+Requires-Dist: boto3>=1.26.1; extra == "all"
+Requires-Dist: pyvmomi==8.0.1.0.2; extra == "all"
+Requires-Dist: nebius>=0.2.47; extra == "all"
+Requires-Dist: azure-identity>=1.19.0; extra == "all"
+Requires-Dist: azure-cli>=2.65.0; extra == "all"
+Requires-Dist: azure-common; extra == "all"
+Requires-Dist: anyio; extra == "all"
 Dynamic: author
 Dynamic: classifier
 Dynamic: description
@@ -250,7 +250,8 @@ Dynamic: summary
 ----
 
 :fire: *News* :fire:
-- [Oct 2025] Run large-scale **LLM training with TorchTitan** on any AI infra: [**example**](./llm/torchtitan/)
+- [Oct 2025] Train and serve [Andrej Karpathy's](https://x.com/karpathy/status/1977755427569111362) **nanochat** - the best ChatGPT that $100 can buy: [**example**](./llm/nanochat)
+- [Oct 2025] Run large-scale **LLM training with TorchTitan** on any AI infra: [**example**](./examples/training/torchtitan)
 - [Sep 2025] Scaling AI infrastructure at Abridge - **10x faster development** with SkyPilot: [**blog**](https://blog.skypilot.co/abridge/)
 - [Sep 2025] Network and Storage Benchmarks for LLM training on the cloud: [**blog**](https://maknee.github.io/blog/2025/Network-And-Storage-Training-Skypilot/)
 - [Aug 2025] Serve and finetune **OpenAI GPT-OSS models** (gpt-oss-120b, gpt-oss-20b) with one command on any infra: [**serve**](./llm/gpt-oss/) + [**LoRA and full finetuning**](./llm/gpt-oss-finetuning/)