skypilot-nightly 1.0.0.dev20250824__py3-none-any.whl → 1.0.0.dev20250826__py3-none-any.whl

sky/serve/load_balancing_policies.py

@@ -3,7 +3,7 @@ import collections
 import random
 import threading
 import typing
-from typing import Dict, List, Optional
+from typing import Any, Dict, List, Optional
 
 from sky import sky_logging
 
@@ -146,3 +146,117 @@ class LeastLoadPolicy(LoadBalancingPolicy, name='least_load', default=True):
         del request  # Unused.
         with self.lock:
             self.load_map[replica_url] -= 1
+
+
+class InstanceAwareLeastLoadPolicy(LeastLoadPolicy,
+                                   name='instance_aware_least_load'):
+    """Instance-aware least load load balancing policy.
+
+    This policy considers the accelerator type and its QPS capabilities
+    when distributing load. It normalizes the load by dividing the current
+    load by the target QPS for that accelerator type.
+    """
+
+    def __init__(self) -> None:
+        super().__init__()
+        self.replica_info: Dict[str, Dict[str, Any]] = {}  # replica_url -> info
+        self.target_qps_per_accelerator: Dict[str, float] = {
+        }  # accelerator_type -> target_qps
+
+    def set_ready_replicas(self, ready_replicas: List[str]) -> None:
+        if set(self.ready_replicas) == set(ready_replicas):
+            return
+        with self.lock:
+            self.ready_replicas = ready_replicas
+            # Clean up load map for removed replicas
+            for r in list(self.load_map.keys()):
+                if r not in ready_replicas:
+                    del self.load_map[r]
+            # Initialize load for new replicas
+            for replica in ready_replicas:
+                if replica not in self.load_map:
+                    self.load_map[replica] = 0
+
+    def set_replica_info(self, replica_info: Dict[str, Dict[str, Any]]) -> None:
+        """Set replica information including accelerator types.
+
+        Args:
+            replica_info: Dict mapping replica URL to replica information
+                         e.g., {'http://url1': {'gpu_type': 'A100'}}
+        """
+        with self.lock:
+            self.replica_info = replica_info
+            logger.debug('Set replica info: %s', self.replica_info)
+
+    def set_target_qps_per_accelerator(
+            self, target_qps_per_accelerator: Dict[str, float]) -> None:
+        """Set target QPS for each accelerator type."""
+        with self.lock:
+            self.target_qps_per_accelerator = target_qps_per_accelerator
+
+    def _get_normalized_load(self, replica_url: str) -> float:
+        """Get normalized load for a replica based on its accelerator type."""
+        current_load = self.load_map.get(replica_url, 0)
+
+        # Get accelerator type for this replica
+        replica_data = self.replica_info.get(replica_url, {})
+        accelerator_type = replica_data.get('gpu_type', 'unknown')
+
+        # Get target QPS for this accelerator type with flexible matching
+        target_qps = self._get_target_qps_for_accelerator(accelerator_type)
+        if target_qps <= 0:
+            logger.warning(
+                'Non-positive target QPS (%s) for accelerator type %s; '
+                'using default value 1.0 to avoid division by zero.',
+                target_qps, accelerator_type)
+            target_qps = 1.0
+
+        # Load is normalized by target QPS
+        normalized_load = current_load / target_qps
+
+        logger.debug(
+            'InstanceAwareLeastLoadPolicy: Replica %s - GPU type: %s, '
+            'current load: %s, target QPS: %s, normalized load: %s',
+            replica_url, accelerator_type, current_load, target_qps,
+            normalized_load)
+
+        return normalized_load
+
+    def _get_target_qps_for_accelerator(self, accelerator_type: str) -> float:
+        """Get target QPS for accelerator type with flexible matching."""
+        # Direct match first
+        if accelerator_type in self.target_qps_per_accelerator:
+            return self.target_qps_per_accelerator[accelerator_type]
+
+        # Try matching by base name (e.g., 'A100' matches 'A100:1')
+        for config_key in self.target_qps_per_accelerator.keys():
+            # Remove count suffix (e.g., 'A100:1' -> 'A100')
+            base_name = config_key.split(':')[0]
+            if accelerator_type == base_name:
+                return self.target_qps_per_accelerator[config_key]
+
+        # Fallback to minimum QPS
+        logger.warning(
+            f'No matching QPS found for accelerator type: {accelerator_type}. '
+            f'Available types: {list(self.target_qps_per_accelerator.keys())}. '
+            f'Using default value 1.0 as fallback.')
+        return 1.0
+
+    def _select_replica(self, request: 'fastapi.Request') -> Optional[str]:
+        del request  # Unused.
+        if not self.ready_replicas:
+            return None
+        with self.lock:
+            # Calculate normalized loads for all replicas
+            replica_loads = []
+            for replica in self.ready_replicas:
+                normalized_load = self._get_normalized_load(replica)
+                replica_loads.append((replica, normalized_load))
+
+            # Select replica with minimum normalized load
+            selected_replica = min(replica_loads, key=lambda x: x[1])[0]
+            logger.debug('Available replicas and loads: %s', replica_loads)
+            logger.debug('Selected replica: %s', selected_replica)
+            return selected_replica
+
+    # pre_execute_hook and post_execute_hook are inherited from LeastLoadPolicy

sky/serve/service.py

@@ -359,7 +359,8 @@ def _start(service_name: str, tmp_task_yaml: str, job_id: int, entrypoint: str):
                         load_balancer_log_file).run,
                     args=(controller_addr, load_balancer_port,
                           service_spec.load_balancing_policy,
-                          service_spec.tls_credential))
+                          service_spec.tls_credential,
+                          service_spec.target_qps_per_replica))
                 load_balancer_process.start()
 
             if not is_recovery:

sky/serve/service_spec.py

@@ -3,7 +3,7 @@ import json
 import os
 import textwrap
 import typing
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, List, Optional, Union
 
 from sky import serve
 from sky.adaptors import common as adaptors_common
@@ -33,7 +33,7 @@ class SkyServiceSpec:
         max_replicas: Optional[int] = None,
         num_overprovision: Optional[int] = None,
         ports: Optional[str] = None,
-        target_qps_per_replica: Optional[float] = None,
+        target_qps_per_replica: Optional[Union[float, Dict[str, float]]] = None,
         post_data: Optional[Dict[str, Any]] = None,
         tls_credential: Optional[serve_utils.TLSCredential] = None,
         readiness_headers: Optional[Dict[str, str]] = None,
@@ -109,7 +109,8 @@ class SkyServiceSpec:
         self._max_replicas: Optional[int] = max_replicas
         self._num_overprovision: Optional[int] = num_overprovision
         self._ports: Optional[str] = ports
-        self._target_qps_per_replica: Optional[float] = target_qps_per_replica
+        self._target_qps_per_replica: Optional[Union[float, Dict[
+            str, float]]] = target_qps_per_replica
         self._post_data: Optional[Dict[str, Any]] = post_data
         self._tls_credential: Optional[serve_utils.TLSCredential] = (
             tls_credential)
@@ -241,6 +242,26 @@ class SkyServiceSpec:
         service_config['load_balancing_policy'] = config.get(
             'load_balancing_policy', None)
 
+        # Validate instance-aware settings
+        target_qps_per_replica = service_config['target_qps_per_replica']
+        load_balancing_policy = service_config['load_balancing_policy']
+
+        if isinstance(target_qps_per_replica, dict):
+            if load_balancing_policy != 'instance_aware_least_load':
+                with ux_utils.print_exception_no_traceback():
+                    raise ValueError(
+                        'When using dict type target_qps_per_replica, '
+                        'load_balancing_policy must be '
+                        '"instance_aware_least_load".')
+
+        if load_balancing_policy == 'instance_aware_least_load':
+            if not isinstance(target_qps_per_replica, dict):
+                with ux_utils.print_exception_no_traceback():
+                    raise ValueError(
+                        'When using "instance_aware_least_load" policy, '
+                        'target_qps_per_replica must be a '
+                        'dict mapping GPU types to QPS values.')
+
         tls_section = config.get('tls', None)
         if tls_section is not None:
             service_config['tls_credential'] = serve_utils.TLSCredential(
@@ -435,7 +456,8 @@ class SkyServiceSpec:
         return self._ports
 
     @property
-    def target_qps_per_replica(self) -> Optional[float]:
+    def target_qps_per_replica(
+            self) -> Optional[Union[float, Dict[str, float]]]:
         return self._target_qps_per_replica
 
     @property

sky/server/rest.py

@@ -48,6 +48,13 @@ _session.headers[constants.API_VERSION_HEADER] = str(constants.API_VERSION)
 _session.headers[constants.VERSION_HEADER] = (
     versions.get_local_readable_version())
 
+# Enumerate error types that might be transient and can be addressed by
+# retrying.
+_transient_errors = [
+    requests.exceptions.RequestException,
+    ConnectionError,
+]
+
 
 class RetryContext:
 
@@ -87,14 +94,10 @@ def retry_transient_errors(max_retries: int = 3,
         if isinstance(e, requests.exceptions.HTTPError):
             # Only server error is considered as transient.
             return e.response.status_code >= 500
-        if isinstance(e, exceptions.ClientError):
-            return False
-        # It is hard to enumerate all other errors that are transient, e.g.
-        # broken pipe, connection refused, etc. Instead, it is safer to assume
-        # all other errors might be transient since we only retry for 3 times
-        # by default. For permanent errors that we do not know now, we can
-        # exclude them here in the future.
-        return True
+        for error in _transient_errors:
+            if isinstance(e, error):
+                return True
+        return False
 
     def decorator(func: F) -> F:
 

sky/utils/schemas.py

@@ -668,8 +668,24 @@ def get_service_schema():
                         'minimum': 0,
                     },
                     'target_qps_per_replica': {
-                        'type': 'number',
-                        'minimum': 0,
+                        'anyOf': [
+                            {
+                                'type': 'number',
+                                'minimum': 0,
+                            },
+                            {
+                                'type': 'object',
+                                'patternProperties': {
+                                    # Pattern for accelerator types like
+                                    # "H100:1", "A100:1", "H100", "A100"
+                                    '^[A-Z0-9]+(?::[0-9]+)?$': {
+                                        'type': 'number',
+                                        'minimum': 0,
+                                    }
+                                },
+                                'additionalProperties': False,
+                            }
+                        ]
                     },
                     'dynamic_ondemand_fallback': {
                         'type': 'boolean',

{skypilot_nightly-1.0.0.dev20250824.dist-info → skypilot_nightly-1.0.0.dev20250826.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: skypilot-nightly
-Version: 1.0.0.dev20250824
+Version: 1.0.0.dev20250826
 Summary: SkyPilot: Run AI on Any Infra — Unified, Faster, Cheaper.
 Author: SkyPilot Team
 License: Apache 2.0

{skypilot_nightly-1.0.0.dev20250824.dist-info → skypilot_nightly-1.0.0.dev20250826.dist-info}/RECORD

@@ -1,4 +1,4 @@
-sky/__init__.py,sha256=ErYYKEgFO1itJZeoNiYaFkEnJ9HkapcA0vIZ-tQx1g8,6615
+sky/__init__.py,sha256=wU94TjC0ZwBhdANnFHf83Ng7EochDDpzei9uwvl2U3g,6615
 sky/admin_policy.py,sha256=BeSowGnWfDj58ALiNf3cc2N4gMQWzjO6aXnX7vaqYhk,9782
 sky/authentication.py,sha256=V7zGSV7bqcAKC_EGOOS0KhJ01ZFLnme0WnjLFO7zavs,25603
 sky/check.py,sha256=Z7D6txaOAEL7fyEQ8q-Zxk1aWaHpEcl412Rj2mThbQ0,31025
@@ -36,7 +36,7 @@ sky/adaptors/vsphere.py,sha256=zJP9SeObEoLrpgHW2VHvZE48EhgVf8GfAEIwBeaDMfM,2129
 sky/backends/__init__.py,sha256=tpa9gAygQopsiBUUuy3wVmr4E05FoPTFHIWqEo4i-u0,627
 sky/backends/backend.py,sha256=DcZZYgBvhghjLnD1C2TvgAisPpcqT0oYNJ3p_P-0UT4,8098
 sky/backends/backend_utils.py,sha256=TTwnjhW7hVNnrVQe8dyVxHssSxL-_4mZiT-qaFQgur8,157967
-sky/backends/cloud_vm_ray_backend.py,sha256=X-FJiOX82-Z8Ut2zX7T6u3T9rN4gxD7huwqrjriIVI8,275131
+sky/backends/cloud_vm_ray_backend.py,sha256=7gZ_kjHYeoz_473nDxLHF4lbgBVg_3BGFr-B_ScyhMo,276537
 sky/backends/docker_utils.py,sha256=_EhM6NStZDAwcegppQqExaB5iuSn1qL4xFFUqXAz2Uk,8392
 sky/backends/local_docker_backend.py,sha256=i_Ff6MUNiIJKknTJw0YtQcvDlba2F7f44lx_eZKfP40,17409
 sky/backends/wheel_utils.py,sha256=2ik8vImmlYLCygRqqTI1ukfZNpqjpdjvj2kHWznFjs8,11104
@@ -71,7 +71,7 @@ sky/catalog/data_fetchers/fetch_fluidstack.py,sha256=hsqpQi_YUI-qil3zLCEGatrR7Bk
 sky/catalog/data_fetchers/fetch_gcp.py,sha256=iFOpPq52c2ENg-oV9OXqFqPMG4Dpv5VGtRhcxkMir90,31950
 sky/catalog/data_fetchers/fetch_hyperbolic.py,sha256=VLBJXWvrKJH5QbMUiREPEMfO3x37_XsqMAmOo7bCQtc,4537
 sky/catalog/data_fetchers/fetch_ibm.py,sha256=WPzR1y5ZaTdv-R3HLIdSUnOfWh4N9cqzKoKiKJQkjFk,7414
-sky/catalog/data_fetchers/fetch_lambda_cloud.py,sha256=MUzogyLruLQmIt-To6TsfnGPgv_nnlp49XYbeshsd7I,5003
+sky/catalog/data_fetchers/fetch_lambda_cloud.py,sha256=9-_gMwErwSbitC5ByZWWOKrn1S6_KfOUIfFbfbJbXZo,5033
 sky/catalog/data_fetchers/fetch_nebius.py,sha256=PeiVTWsaw0P01cfq2xDcS1tcIhLIe4ObSPF1XLO67gc,12181
 sky/catalog/data_fetchers/fetch_vast.py,sha256=xoVDSsQVgMLzyibCFN7yDgyH1Y96gk5G53to1ZAGRyg,5017
 sky/catalog/data_fetchers/fetch_vsphere.py,sha256=Yf7tKzwJsQ_4f64IT1EAP108C1D3Rg35RUIwp7UX8KI,21438
@@ -112,19 +112,19 @@ sky/clouds/utils/azure_utils.py,sha256=NToRBnhEyuUvb-nBnsKTxjhOBRkMcrelL8LK4w6s4
 sky/clouds/utils/gcp_utils.py,sha256=09MF4Vx0EW7S-GXGpyxpl2aQlHrqeu9ioV0nyionAyk,9890
 sky/clouds/utils/oci_utils.py,sha256=TFqAqRLggg4Z0bhxrrq8nouSSomZy-ub1frHXEkud2M,7302
 sky/clouds/utils/scp_utils.py,sha256=VGuccVO5uFGr8-yolWSoYrgr11z6cIeDBGcqkBzAyOs,18409
-sky/dashboard/out/404.html,sha256=3zgZj8MoYvjd_seRSq42nzhsxQgvMcRuR3R3cINyTD8,1423
-sky/dashboard/out/clusters.html,sha256=P3iL-erjnenog559rLvVoSbdqKNFd5V4x5uc0iFelTo,1418
-sky/dashboard/out/config.html,sha256=ulAdJyD3MlL2cFKxRoIpsK09hMgimnZpdSqEn7G_8iQ,1414
+sky/dashboard/out/404.html,sha256=3Gvl6FWCjaT8L8TCCCiVT46IkRybXQJlAqZm6PI5jYA,1423
+sky/dashboard/out/clusters.html,sha256=DK23NMk9Mg8ZPXn12UX-wsn0RdgYzjS_-gg6jt86kO0,1418
+sky/dashboard/out/config.html,sha256=tc62a9w66l5sNtBf-5BeZobBvZkWKRwyiOrNjre82jI,1414
 sky/dashboard/out/favicon.ico,sha256=XilUZZglAl_1zRsg85QsbQgmQAzGPQjcUIJ-A3AzYn8,93590
-sky/dashboard/out/index.html,sha256=HbOILrCfCvtDGfcdzYxcI4y8bUVlv3Xo_is8renm-08,1407
-sky/dashboard/out/infra.html,sha256=X40H7RxFfnnDGQKak_CdbNCm08vUb9E2FPTdKmyzAbE,1412
-sky/dashboard/out/jobs.html,sha256=AIfTWtz6jUNAf2134oS-nKAm3TgWVOH4Jxt8uhvxUYU,1410
+sky/dashboard/out/index.html,sha256=__o4Clp3Xp3XGx-5n6FFJgmqGVntA_ZbS-8o4QNj0cI,1407
+sky/dashboard/out/infra.html,sha256=98bUimOMW5-Sni6kSx3wIVBiRR6js5NPN1_IiBjjYo4,1412
+sky/dashboard/out/jobs.html,sha256=r5yMieAcXZEwVea9c2b2-97Yn0aGxtxMzvHZsocLgYg,1410
 sky/dashboard/out/skypilot.svg,sha256=c0iRtlfLlaUm2p0rG9NFmo5FN0Qhf3pq5Xph-AeMPJw,5064
-sky/dashboard/out/users.html,sha256=j8tk1SCcPJPvNVnzfvuG9sVr0KM9INnVno6TnbXllEg,1412
-sky/dashboard/out/volumes.html,sha256=h5y034-4H3XQThxjD3HOj6Ffe0p6UeSLWlpKRc4KtV4,1416
-sky/dashboard/out/workspaces.html,sha256=_1A1rUYps3rz4sqiSFk1edpFqG8M9FmxJVg1LKMaru0,1422
-sky/dashboard/out/_next/static/P-Au_yHqNENhnPF3shEpK/_buildManifest.js,sha256=osxUordT3Fb32OsAF270nSx4GMgfsKF6Vc8OBBxtPgk,2428
-sky/dashboard/out/_next/static/P-Au_yHqNENhnPF3shEpK/_ssgManifest.js,sha256=Z49s4suAsf5y_GfnQSvm4qtq2ggxEbZPfEDTXjy6XgA,80
+sky/dashboard/out/users.html,sha256=_9ETCX5JkiP4a58Whkc-NHzV9MjcKEWhYq4sg4xz6f4,1412
+sky/dashboard/out/volumes.html,sha256=p3tUR7cOsCXet1rgJMhsYWsRdrbzdaPxr-uw4n_gPmU,1416
+sky/dashboard/out/workspaces.html,sha256=sfzrgzwdxfvY0Ccm6xkKivKu374k3EYcZSx9fnZzxvU,1422
+sky/dashboard/out/_next/static/TPMkEeuj85tHTmIW7Gu3S/_buildManifest.js,sha256=osxUordT3Fb32OsAF270nSx4GMgfsKF6Vc8OBBxtPgk,2428
+sky/dashboard/out/_next/static/TPMkEeuj85tHTmIW7Gu3S/_ssgManifest.js,sha256=Z49s4suAsf5y_GfnQSvm4qtq2ggxEbZPfEDTXjy6XgA,80
 sky/dashboard/out/_next/static/chunks/1121-8afcf719ea87debc.js,sha256=fLXxFyYpxIIH-GAL9X9Ew3rc2f6zqOZqg6TjrapDZUM,8554
 sky/dashboard/out/_next/static/chunks/1141-943efc7aff0f0c06.js,sha256=tUOoU0nIEShZeD5pBiOWrl8-czHc6PpnxxJilnDplHM,17330
 sky/dashboard/out/_next/static/chunks/1272-1ef0bf0237faccdb.js,sha256=VJ6y-Z6Eg2T93hQIRfWAbjAkQ7nQhglmIaVbEpKSILY,38451
@@ -179,14 +179,14 @@ sky/dashboard/out/_next/static/chunks/pages/jobs/pools/[pool]-07349868f7905d37.j
 sky/dashboard/out/_next/static/chunks/pages/workspace/new-3f88a1c7e86a3f86.js,sha256=83s5N5CZwIaRcmYMfqn2we60n2VRmgFw6Tbx18b8-e0,762
 sky/dashboard/out/_next/static/chunks/pages/workspaces/[name]-de06e613e20bc977.js,sha256=8d4XLtF8E3ahNnsbdNUQkJVbM1b9sIG9wRaoRjRwMhE,1495
 sky/dashboard/out/_next/static/css/4614e06482d7309e.css,sha256=nk6GriyGVd1aGXrLd7BcMibnN4v0z-Q_mXGxrHFWqrE,56126
-sky/dashboard/out/clusters/[cluster].html,sha256=X1D6tb-oYeVkCsmd9M4SOTBNFTt_0F2pabLeX6kdqi0,2936
-sky/dashboard/out/clusters/[cluster]/[job].html,sha256=GC_09YG2cYyXLfMfFavEdsteChdoyWgHQjnOVKlWB5s,2073
-sky/dashboard/out/infra/[context].html,sha256=aK8corxweqCa8HPl6xP87-9ONlkSuCtXfMIP8Rs4SOo,1436
-sky/dashboard/out/jobs/[job].html,sha256=pnOvaPI2uMKtAcK8v0qkC5hvodHt0GUS72NJRWjYGhY,2304
-sky/dashboard/out/jobs/pools/[pool].html,sha256=IQPvRWlwSDHH7ZIpWQkBDv2GOsy2BiJ_PZAWwWXj2uM,2142
+sky/dashboard/out/clusters/[cluster].html,sha256=N2GvGQqYXrX6ImCQ51uGGrnaJafYx7c_mk-HrrDcpZQ,2936
+sky/dashboard/out/clusters/[cluster]/[job].html,sha256=CTOhsQPFLQ-XK2pefCSn4Ywtx-7RqcHG0IaX2gi_YG8,2073
+sky/dashboard/out/infra/[context].html,sha256=DJPp07oZX0EkqIg0MPFsDBTlQ3RFViC6KQyAquq0HRs,1436
+sky/dashboard/out/jobs/[job].html,sha256=iHkradqnTT9wfFtizRvfbjNc3U1FXx-Um79YC0-PRes,2304
+sky/dashboard/out/jobs/pools/[pool].html,sha256=ijXIK3F83uLQJJusrGJxVdKERd0vXc7qDJNqrXrrbK4,2142
 sky/dashboard/out/videos/cursor-small.mp4,sha256=8tRdp1vjawOrXUar1cfjOc-nkaKmcwCPZx_LO0XlCvQ,203285
-sky/dashboard/out/workspace/new.html,sha256=fGFgyYn19fP_7sqmfo25MZbDECC3Lo4rWrseDFO90fQ,1428
-sky/dashboard/out/workspaces/[name].html,sha256=EHGka1l0lW0p1CE1ir8LL58sl1YuEWE3XiiWVNYSv6I,2759
+sky/dashboard/out/workspace/new.html,sha256=K0pQef1eo_m4Lc2XLP20d95hISfA27ceJO0L7Xk2p0o,1428
+sky/dashboard/out/workspaces/[name].html,sha256=YMnimIRnIc05ryUmLHayp5TgaKCY86Wdj1Ik_m9w_yk,2759
 sky/data/__init__.py,sha256=Nhaf1NURisXpZuwWANa2IuCyppIuc720FRwqSE2oEwY,184
 sky/data/data_transfer.py,sha256=N8b0CQebDuHieXjvEVwlYmK6DbQxUGG1RQJEyTbh3dU,12040
 sky/data/data_utils.py,sha256=AjEA_JRjo9NBMlv-Lq5iV4lBED_YZ1VqBR9pG6fGVWE,35179
@@ -261,7 +261,7 @@ sky/provision/kubernetes/constants.py,sha256=vZJQsAVjAgwsOskB48tIFSXtNw7IFnJOQE_
 sky/provision/kubernetes/instance.py,sha256=Z9c3nUIniijD9OMQ1PO4YyuiL4764w5P_3MzSWTL-Fg,67333
 sky/provision/kubernetes/network.py,sha256=Dgj8u7IQBHKHt-mSDhYzue1wfDk96FR_8fO89TwuZ2E,12846
 sky/provision/kubernetes/network_utils.py,sha256=EROwcgkdVawLQVkw6cpRaEg4k2pS0V8dEoO8vh9G9Kc,12273
-sky/provision/kubernetes/utils.py,sha256=ED3keqiKDU3FkysiTvBf_MNgE2oslv3kEgQP6yxL6xs,158275
+sky/provision/kubernetes/utils.py,sha256=NBIJtXTBxP1gMtps_j1TSyYhPH9jFnZ0XguZeQVcsOo,158411
 sky/provision/kubernetes/volume.py,sha256=mChbYJw02vrUbBWw9a2mKHWlSfCBjWS4AeWs5_6MPSw,8142
 sky/provision/kubernetes/manifests/fusermount-server-daemonset.yaml,sha256=S87GNAbDqgTrLuxF-afPAqQ0V-i41El4s_9KBZMuaag,1331
 sky/provision/lambda_cloud/__init__.py,sha256=6EEvSgtUeEiup9ivIFevHmgv0GqleroO2X0K7TRa2nE,612
@@ -333,16 +333,16 @@ sky/schemas/generated/autostopv1_pb2.py,sha256=3w4k0woHFxK0qXvYGljLtqpp3IB-oGV15
 sky/schemas/generated/autostopv1_pb2.pyi,sha256=1slXiAfgnh-Ds-yHJglvhGddatmlOPfh3AAOCheipCk,1772
 sky/schemas/generated/autostopv1_pb2_grpc.py,sha256=sUoHZNUNjxpsiYzJhSchC7p-uvigYuh8DwhhyhgfPcE,5921
 sky/serve/__init__.py,sha256=jlwErB4VeKh0wp9FkTIQoE5JQvc6KIB-vrqhtsJ2q-E,1810
-sky/serve/autoscalers.py,sha256=CWPooRr5mdtwmD2PZWl9GmcS1tifW21Hapv5mp6nB50,32971
+sky/serve/autoscalers.py,sha256=4qS01f3-UUMC4kPe7HI7ayyWNwX1oMKVF7lwW_DX2wk,48542
 sky/serve/constants.py,sha256=VYQKGuwaPI9o2xC5HzB0TGKdiMeE9gOiX82K5yGEfGg,5062
-sky/serve/controller.py,sha256=JlDvk275fBt5d_I1B2KiBHiBlz-98NlgNF-7ShvU9tE,12445
-sky/serve/load_balancer.py,sha256=ZUt1QEf8b7tpmtedg7oiWkXDapm8fb7a6YdLqDK7Tno,14204
-sky/serve/load_balancing_policies.py,sha256=Ssr0GtHKyzzAIEOUAMCLvND5rapmrpxI-OQemf3dgxU,5177
+sky/serve/controller.py,sha256=lyB6EHIfEMqKg-gshFLKUHEZqs77ufsD_ev6catz10Q,13975
+sky/serve/load_balancer.py,sha256=gQpjOwnpu2HLdymbYCeCgJodDklMEwYMGgYU9ThXHPM,15774
+sky/serve/load_balancing_policies.py,sha256=hhPrcuwT_rEJx4DMRC6iuLpUY8wOabUtXtPK7clg1F8,10148
 sky/serve/replica_managers.py,sha256=YaNyB8czSbxwlZP3oQqM4fOwxK0chIp9e7XVJYiTB4g,70830
 sky/serve/serve_state.py,sha256=k9rvkz-l8tpZZN1cL7aBtb60l_DE_3dE0DVyfT4tYzg,33328
 sky/serve/serve_utils.py,sha256=JptYHFMMlRGsHF10HjUnIiVc2MUHok393B24WJyO2Aw,69220
-sky/serve/service.py,sha256=yiqNIAJ2-whyrux-uY4CZrLDFcmNVhWZxQq2HfcFJdo,18401
-sky/serve/service_spec.py,sha256=c92moNwtjg1gCXug1ydsnoVG7mvUWNZ5esCT6GpS-K4,21689
+sky/serve/service.py,sha256=mD5bOfw6VeyyqkBG-NPiM89ZP4V7I4ACTiqg4jLQaJg,18464
+sky/serve/service_spec.py,sha256=yro9RJw62FADF-lsEQMGmyLvHancjQD9ez7B0eaPGVE,22821
 sky/serve/spot_placer.py,sha256=auRlYZNg8uIW-lxiirxpDP-iIaJtkrhesm6OGMgwPq4,11288
 sky/serve/client/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 sky/serve/client/impl.py,sha256=TMvZGnui5QeqhqGzCzrMrLXO_As8RDUFSZa67cDbJJs,9466
@@ -358,7 +358,7 @@ sky/server/config.py,sha256=XWf5Kw4am6vMO5wcyWevbQAFH-dmKb7AMEgDzD083-M,8538
 sky/server/constants.py,sha256=yjX8t73w6gj3_SDSP4vBFdNdiOqq7dnlXT2pw3yo0jM,2321
 sky/server/daemons.py,sha256=3cwp-Os2StWLaOF21Sv4jzJ1zkEYpKuiKF5HFkg86WA,8700
 sky/server/metrics.py,sha256=6H6n6dq_C5HMaU97mJlRUB9bqOEA_k205PO15wE3AWk,3648
-sky/server/rest.py,sha256=GAT1uS5tacRqBbbECbVxMsnEIdEFjjuqAryNgz0IiYA,14502
+sky/server/rest.py,sha256=6Qcn6fjypP3j9UHdKRgvt2-PU1LKz2VU2aVQEA1D6EI,14354
 sky/server/server.py,sha256=YTFhgaAKMIYvxwKdgPoAc0CP7TlVEm6AvEmrMEB8uqY,78411
 sky/server/state.py,sha256=YbVOMJ1JipQQv17gLIGyiGN7MKfnP83qlUa5MB1z0Yk,747
 sky/server/stream_utils.py,sha256=RS4RuMxQqTGqp3uxzZVtmFWzos4d49P7hMX_VklzEVU,9189
@@ -488,7 +488,7 @@ sky/utils/resource_checker.py,sha256=0rwr7yLVkYO3Qq5FZmniyPp-p66tIXmSoK5t0ZgIfso
 sky/utils/resources_utils.py,sha256=3wnzmSIldFS5NmHTx6r2viS8zaP1q20noQolgQqucUU,16722
 sky/utils/rich_console_utils.py,sha256=wPvAlshaFHuMZSjiDnaK3OSBppZLBjAn-lj7AvxNBQk,553
 sky/utils/rich_utils.py,sha256=Q-N5bZGfvqciU5cuQacInoNpldZcaMKCdBX2368KIDA,19971
-sky/utils/schemas.py,sha256=d7vWIQWCYCzGYgNNbN-jWArHYspXMmgjwLhjgNXgoU0,56352
+sky/utils/schemas.py,sha256=sqID5M6nknukfVtTlTShAp-nhptjnZJbkLBjO-i_doY,57114
 sky/utils/serialize_utils.py,sha256=nn2x-8cTZeiVr5cgaBpLOGGpSFtms62QAJFyxs_bodI,630
 sky/utils/status_lib.py,sha256=QGkd6COD1GX1h30Mk9RMUdyeUOMJs5971GkxTcFgdsU,1705
 sky/utils/subprocess_utils.py,sha256=tOpFY_1ml7JkVGAN1o473lcKPklGR95qBCW61eu8kEo,15773
@@ -533,9 +533,9 @@ sky/workspaces/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 sky/workspaces/core.py,sha256=AjwbbRwk0glzCnqICJk4sQzMoUcawixbXoQWKLB3-aQ,25372
 sky/workspaces/server.py,sha256=Box45DS54xXGHy7I3tGKGy-JP0a8G_z6IhfvGlEXtsA,3439
 sky/workspaces/utils.py,sha256=IIAiFoS6sdb2t0X5YoX9AietpTanZUQNTK8cePun-sY,2143
-skypilot_nightly-1.0.0.dev20250824.dist-info/licenses/LICENSE,sha256=emRJAvE7ngL6x0RhQvlns5wJzGI3NEQ_WMjNmd9TZc4,12170
-skypilot_nightly-1.0.0.dev20250824.dist-info/METADATA,sha256=ttUO6cdP3knZ5TtZZm6jdT6HcWiQeda4dpvBnsbicoE,20310
-skypilot_nightly-1.0.0.dev20250824.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-skypilot_nightly-1.0.0.dev20250824.dist-info/entry_points.txt,sha256=StA6HYpuHj-Y61L2Ze-hK2IcLWgLZcML5gJu8cs6nU4,36
-skypilot_nightly-1.0.0.dev20250824.dist-info/top_level.txt,sha256=qA8QuiNNb6Y1OF-pCUtPEr6sLEwy2xJX06Bd_CrtrHY,4
-skypilot_nightly-1.0.0.dev20250824.dist-info/RECORD,,
+skypilot_nightly-1.0.0.dev20250826.dist-info/licenses/LICENSE,sha256=emRJAvE7ngL6x0RhQvlns5wJzGI3NEQ_WMjNmd9TZc4,12170
+skypilot_nightly-1.0.0.dev20250826.dist-info/METADATA,sha256=jBNYOk9-nDPQ0vU2eZ2HoHuVvcWVnMinnhtS39yU4Vo,20310
+skypilot_nightly-1.0.0.dev20250826.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+skypilot_nightly-1.0.0.dev20250826.dist-info/entry_points.txt,sha256=StA6HYpuHj-Y61L2Ze-hK2IcLWgLZcML5gJu8cs6nU4,36
+skypilot_nightly-1.0.0.dev20250826.dist-info/top_level.txt,sha256=qA8QuiNNb6Y1OF-pCUtPEr6sLEwy2xJX06Bd_CrtrHY,4
+skypilot_nightly-1.0.0.dev20250826.dist-info/RECORD,,