skypilot-nightly 1.0.0.dev20241025__py3-none-any.whl → 1.0.0.dev20241027__py3-none-any.whl

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 = '057bc4b44755ac1e9dadc680e022c369e8ddff52'
+_SKYPILOT_COMMIT_SHA = 'c0c17483d1f692ad639144050f5f6fa0966e47a5'
 
 
 def _get_git_commit():
@@ -35,7 +35,7 @@ def _get_git_commit():
 
 
 __commit__ = _get_git_commit()
-__version__ = '1.0.0.dev20241025'
+__version__ = '1.0.0.dev20241027'
 __root_dir__ = os.path.dirname(os.path.abspath(__file__))
 
 

sky/backends/cloud_vm_ray_backend.py

@@ -2713,6 +2713,21 @@ class CloudVmRayBackend(backends.Backend['CloudVmRayResourceHandle']):
                     f'  Existing:\t{handle.launched_nodes}x '
                     f'{handle.launched_resources}\n'
                     f'{mismatch_str}')
+        else:
+            # For fractional acc count clusters, we round up the number of accs
+            # to 1 (sky/utils/resources_utils.py::make_ray_custom_resources_str)
+            # Here we scale the required acc count to (required / launched) * 1
+            # so the total number of accs is the same as the requested number.
+            launched_accs = launched_resources.accelerators
+            if (launched_accs is not None and
+                    valid_resource.accelerators is not None):
+                for _, count in launched_accs.items():
+                    if isinstance(count, float) and not count.is_integer():
+                        valid_resource = valid_resource.copy(
+                            accelerators={
+                                k: v / count
+                                for k, v in valid_resource.accelerators.items()
+                            })
         return valid_resource
 
     def _provision(

sky/cli.py

@@ -3519,7 +3519,7 @@ def jobs():
               default=None,
               type=str,
               hidden=True,
-              help=('Alias for --name, the name of the spot job.'))
+              help=('Alias for --name, the name of the managed job.'))
 @click.option('--job-recovery',
               default=None,
               type=str,

sky/clouds/aws.py

@@ -2,13 +2,12 @@
 import enum
 import fnmatch
 import functools
-import json
 import os
 import re
 import subprocess
 import time
 import typing
-from typing import Any, Dict, Iterator, List, Optional, Set, Tuple
+from typing import Any, Dict, Iterator, List, Optional, Set, Tuple, Union
 
 from sky import clouds
 from sky import exceptions
@@ -383,7 +382,7 @@ class AWS(clouds.Cloud):
     def get_accelerators_from_instance_type(
         cls,
         instance_type: str,
-    ) -> Optional[Dict[str, int]]:
+    ) -> Optional[Dict[str, Union[int, float]]]:
         return service_catalog.get_accelerators_from_instance_type(
             instance_type, clouds='aws')
 
@@ -411,10 +410,8 @@ class AWS(clouds.Cloud):
         r = resources
         # r.accelerators is cleared but .instance_type encodes the info.
         acc_dict = self.get_accelerators_from_instance_type(r.instance_type)
-        if acc_dict is not None:
-            custom_resources = json.dumps(acc_dict, separators=(',', ':'))
-        else:
-            custom_resources = None
+        custom_resources = resources_utils.make_ray_custom_resources_str(
+            acc_dict)
 
         if r.extract_docker_image() is not None:
             image_id_to_use = None

sky/clouds/azure.py

@@ -1,12 +1,11 @@
 """Azure."""
 import functools
-import json
 import os
 import re
 import subprocess
 import textwrap
 import typing
-from typing import Any, Dict, Iterator, List, Optional, Tuple
+from typing import Any, Dict, Iterator, List, Optional, Tuple, Union
 
 import colorama
 
@@ -39,9 +38,9 @@ _DEFAULT_AZURE_UBUNTU_HPC_IMAGE_GB = 30
 _DEFAULT_AZURE_UBUNTU_2004_IMAGE_GB = 150
 _DEFAULT_SKYPILOT_IMAGE_GB = 30
 
-_DEFAULT_CPU_IMAGE_ID = 'skypilot:gpu-ubuntu-2204'
-_DEFAULT_GPU_IMAGE_ID = 'skypilot:gpu-ubuntu-2204'
-_DEFAULT_V1_IMAGE_ID = 'skypilot:v1-ubuntu-2004'
+_DEFAULT_CPU_IMAGE_ID = 'skypilot:custom-cpu-ubuntu-v2'
+_DEFAULT_GPU_IMAGE_ID = 'skypilot:custom-gpu-ubuntu-v2'
+_DEFAULT_V1_IMAGE_ID = 'skypilot:custom-gpu-ubuntu-v1'
 _DEFAULT_GPU_K80_IMAGE_ID = 'skypilot:k80-ubuntu-2004'
 _FALLBACK_IMAGE_ID = 'skypilot:gpu-ubuntu-2204'
 
@@ -272,7 +271,7 @@ class Azure(clouds.Cloud):
     def get_accelerators_from_instance_type(
         cls,
         instance_type: str,
-    ) -> Optional[Dict[str, int]]:
+    ) -> Optional[Dict[str, Union[int, float]]]:
         return service_catalog.get_accelerators_from_instance_type(
             instance_type, clouds='azure')
 
@@ -304,10 +303,9 @@ class Azure(clouds.Cloud):
         acc_dict = self.get_accelerators_from_instance_type(r.instance_type)
         acc_count = None
         if acc_dict is not None:
-            custom_resources = json.dumps(acc_dict, separators=(',', ':'))
             acc_count = str(sum(acc_dict.values()))
-        else:
-            custom_resources = None
+        custom_resources = resources_utils.make_ray_custom_resources_str(
+            acc_dict)
 
         if (resources.image_id is None or
                 resources.extract_docker_image() is not None):

sky/clouds/cloud.py

@@ -9,8 +9,9 @@ reused across cloud object creation.
 """
 import collections
 import enum
+import math
 import typing
-from typing import Dict, Iterable, Iterator, List, Optional, Set, Tuple
+from typing import Dict, Iterable, Iterator, List, Optional, Set, Tuple, Union
 
 from sky import exceptions
 from sky import skypilot_config
@@ -306,7 +307,7 @@ class Cloud:
     def get_accelerators_from_instance_type(
         cls,
         instance_type: str,
-    ) -> Optional[Dict[str, int]]:
+    ) -> Optional[Dict[str, Union[int, float]]]:
         """Returns {acc: acc_count} held by 'instance_type', if any."""
         raise NotImplementedError
 
@@ -673,8 +674,9 @@ class Cloud:
         assert resources.is_launchable(), resources
 
         def _equal_accelerators(
-                acc_requested: Optional[Dict[str, int]],
-                acc_from_instance_type: Optional[Dict[str, int]]) -> bool:
+            acc_requested: Optional[Dict[str, Union[int, float]]],
+            acc_from_instance_type: Optional[Dict[str, Union[int,
+                                                             float]]]) -> bool:
             """Check the requested accelerators equals to the instance type
 
             Check the requested accelerators equals to the accelerators
@@ -689,12 +691,14 @@ class Cloud:
             for acc in acc_requested:
                 if acc not in acc_from_instance_type:
                     return False
-                if acc_requested[acc] != acc_from_instance_type[acc]:
+                # Avoid float point precision issue.
+                if not math.isclose(acc_requested[acc],
+                                    acc_from_instance_type[acc]):
                     return False
             return True
 
-        acc_from_instance_type = (cls.get_accelerators_from_instance_type(
-            resources.instance_type))
+        acc_from_instance_type = cls.get_accelerators_from_instance_type(
+            resources.instance_type)
         if not _equal_accelerators(resources.accelerators,
                                    acc_from_instance_type):
             with ux_utils.print_exception_no_traceback():

sky/clouds/cudo.py

@@ -1,8 +1,7 @@
 """Cudo Compute"""
-import json
 import subprocess
 import typing
-from typing import Dict, Iterator, List, Optional, Tuple
+from typing import Dict, Iterator, List, Optional, Tuple, Union
 
 from sky import clouds
 from sky.clouds import service_catalog
@@ -183,7 +182,7 @@ class Cudo(clouds.Cloud):
     def get_accelerators_from_instance_type(
         cls,
         instance_type: str,
-    ) -> Optional[Dict[str, int]]:
+    ) -> Optional[Dict[str, Union[int, float]]]:
         return service_catalog.get_accelerators_from_instance_type(
             instance_type, clouds='cudo')
 
@@ -202,10 +201,8 @@ class Cudo(clouds.Cloud):
         del zones, cluster_name  # unused
         r = resources
         acc_dict = self.get_accelerators_from_instance_type(r.instance_type)
-        if acc_dict is not None:
-            custom_resources = json.dumps(acc_dict, separators=(',', ':'))
-        else:
-            custom_resources = None
+        custom_resources = resources_utils.make_ray_custom_resources_str(
+            acc_dict)
 
         return {
             'instance_type': resources.instance_type,

sky/clouds/fluidstack.py

@@ -1,8 +1,7 @@
 """Fluidstack Cloud."""
-import json
 import os
 import typing
-from typing import Dict, Iterator, List, Optional, Tuple
+from typing import Dict, Iterator, List, Optional, Tuple, Union
 
 import requests
 
@@ -155,7 +154,7 @@ class Fluidstack(clouds.Cloud):
     def get_accelerators_from_instance_type(
         cls,
         instance_type: str,
-    ) -> Optional[Dict[str, int]]:
+    ) -> Optional[Dict[str, Union[int, float]]]:
         return service_catalog.get_accelerators_from_instance_type(
             instance_type, clouds='fluidstack')
 
@@ -184,10 +183,8 @@ class Fluidstack(clouds.Cloud):
 
         r = resources
         acc_dict = self.get_accelerators_from_instance_type(r.instance_type)
-        if acc_dict is not None:
-            custom_resources = json.dumps(acc_dict, separators=(',', ':'))
-        else:
-            custom_resources = None
+        custom_resources = resources_utils.make_ray_custom_resources_str(
+            acc_dict)
 
         return {
             'instance_type': resources.instance_type,

sky/clouds/gcp.py

@@ -7,7 +7,7 @@ import re
 import subprocess
 import time
 import typing
-from typing import Any, Dict, Iterator, List, Optional, Set, Tuple
+from typing import Any, Dict, Iterator, List, Optional, Set, Tuple, Union
 
 import colorama
 
@@ -669,7 +669,7 @@ class GCP(clouds.Cloud):
     def get_accelerators_from_instance_type(
         cls,
         instance_type: str,
-    ) -> Optional[Dict[str, int]]:
+    ) -> Optional[Dict[str, Union[int, float]]]:
         # GCP handles accelerators separately from regular instance types,
         # hence return none here.
         return None

sky/clouds/ibm.py

@@ -1,8 +1,7 @@
 """IBM Web Services."""
-import json
 import os
 import typing
-from typing import Any, Dict, Iterator, List, Optional, Tuple
+from typing import Any, Dict, Iterator, List, Optional, Tuple, Union
 
 import colorama
 
@@ -206,10 +205,8 @@ class IBM(clouds.Cloud):
             'IBM does not currently support spot instances in this framework'
 
         acc_dict = self.get_accelerators_from_instance_type(r.instance_type)
-        if acc_dict is not None:
-            custom_resources = json.dumps(acc_dict, separators=(',', ':'))
-        else:
-            custom_resources = None
+        custom_resources = resources_utils.make_ray_custom_resources_str(
+            acc_dict)
 
         instance_resources = _get_profile_resources(r.instance_type)
 
@@ -247,7 +244,7 @@ class IBM(clouds.Cloud):
     def get_accelerators_from_instance_type(
         cls,
         instance_type: str,
-    ) -> Optional[Dict[str, int]]:
+    ) -> Optional[Dict[str, Union[int, float]]]:
         """Returns {acc: acc_count} held by 'instance_type', if any."""
         return service_catalog.get_accelerators_from_instance_type(
             instance_type, clouds='ibm')

sky/clouds/kubernetes.py

@@ -1,10 +1,9 @@
 """Kubernetes."""
 import functools
-import json
 import os
 import re
 import typing
-from typing import Dict, Iterator, List, Optional, Tuple
+from typing import Dict, Iterator, List, Optional, Tuple, Union
 
 from sky import clouds
 from sky import sky_logging
@@ -271,7 +270,7 @@ class Kubernetes(clouds.Cloud):
     def get_accelerators_from_instance_type(
         cls,
         instance_type: str,
-    ) -> Optional[Dict[str, int]]:
+    ) -> Optional[Dict[str, Union[int, float]]]:
         inst = kubernetes_utils.KubernetesInstanceType.from_instance_type(
             instance_type)
         return {
@@ -328,10 +327,8 @@ class Kubernetes(clouds.Cloud):
 
         r = resources
         acc_dict = self.get_accelerators_from_instance_type(r.instance_type)
-        if acc_dict is not None:
-            custom_resources = json.dumps(acc_dict, separators=(',', ':'))
-        else:
-            custom_resources = None
+        custom_resources = resources_utils.make_ray_custom_resources_str(
+            acc_dict)
 
         # resources.memory and cpus are None if they are not explicitly set.
         # We fetch the default values for the instance type in that case.

sky/clouds/lambda_cloud.py

@@ -1,7 +1,6 @@
 """Lambda Cloud."""
-import json
 import typing
-from typing import Dict, Iterator, List, Optional, Tuple
+from typing import Dict, Iterator, List, Optional, Tuple, Union
 
 import requests
 
@@ -136,7 +135,7 @@ class Lambda(clouds.Cloud):
     def get_accelerators_from_instance_type(
         cls,
         instance_type: str,
-    ) -> Optional[Dict[str, int]]:
+    ) -> Optional[Dict[str, Union[int, float]]]:
         return service_catalog.get_accelerators_from_instance_type(
             instance_type, clouds='lambda')
 
@@ -164,10 +163,8 @@ class Lambda(clouds.Cloud):
 
         r = resources
         acc_dict = self.get_accelerators_from_instance_type(r.instance_type)
-        if acc_dict is not None:
-            custom_resources = json.dumps(acc_dict, separators=(',', ':'))
-        else:
-            custom_resources = None
+        custom_resources = resources_utils.make_ray_custom_resources_str(
+            acc_dict)
 
         resources_vars = {
             'instance_type': resources.instance_type,

sky/clouds/oci.py

@@ -20,11 +20,10 @@ History:
  - Hysun He (hysun.he@oracle.com) @ Oct 13, 2024:
    Support more OS types additional to ubuntu for OCI resources.
 """
-import json
 import logging
 import os
 import typing
-from typing import Dict, Iterator, List, Optional, Tuple
+from typing import Dict, Iterator, List, Optional, Tuple, Union
 
 from sky import clouds
 from sky import exceptions
@@ -193,7 +192,7 @@ class OCI(clouds.Cloud):
     def get_accelerators_from_instance_type(
         cls,
         instance_type: str,
-    ) -> Optional[Dict[str, int]]:
+    ) -> Optional[Dict[str, Union[int, float]]]:
         return service_catalog.get_accelerators_from_instance_type(
             instance_type, clouds='oci')
 
@@ -213,10 +212,8 @@ class OCI(clouds.Cloud):
 
         acc_dict = self.get_accelerators_from_instance_type(
             resources.instance_type)
-        if acc_dict is not None:
-            custom_resources = json.dumps(acc_dict, separators=(',', ':'))
-        else:
-            custom_resources = None
+        custom_resources = resources_utils.make_ray_custom_resources_str(
+            acc_dict)
 
         image_str = self._get_image_id(resources.image_id, region.name,
                                        resources.instance_type)
@@ -468,7 +465,11 @@ class OCI(clouds.Cloud):
             api_key_file = oci_cfg[
                 'key_file'] if 'key_file' in oci_cfg else 'BadConf'
             sky_cfg_file = oci_utils.oci_config.get_sky_user_config_file()
-        except (ImportError, oci_adaptor.oci.exceptions.ConfigFileNotFound):
+        # Must catch ImportError before any oci_adaptor.oci.exceptions
+        # because oci_adaptor.oci.exceptions can throw ImportError.
+        except ImportError:
+            return {}
+        except oci_adaptor.oci.exceptions.ConfigFileNotFound:
             return {}
 
         # OCI config and API key file are mandatory

sky/clouds/paperspace.py

@@ -1,8 +1,7 @@
 """ Paperspace Cloud. """
 
-import json
 import typing
-from typing import Dict, Iterator, List, Optional, Tuple
+from typing import Dict, Iterator, List, Optional, Tuple, Union
 
 import requests
 
@@ -162,7 +161,7 @@ class Paperspace(clouds.Cloud):
 
     @classmethod
     def get_accelerators_from_instance_type(
-            cls, instance_type: str) -> Optional[Dict[str, int]]:
+            cls, instance_type: str) -> Optional[Dict[str, Union[int, float]]]:
         return service_catalog.get_accelerators_from_instance_type(
             instance_type, clouds='paperspace')
 
@@ -181,10 +180,8 @@ class Paperspace(clouds.Cloud):
 
         r = resources
         acc_dict = self.get_accelerators_from_instance_type(r.instance_type)
-        if acc_dict is not None:
-            custom_resources = json.dumps(acc_dict, separators=(',', ':'))
-        else:
-            custom_resources = None
+        custom_resources = resources_utils.make_ray_custom_resources_str(
+            acc_dict)
 
         return {
             'instance_type': resources.instance_type,

sky/clouds/runpod.py

@@ -1,8 +1,7 @@
 """ RunPod Cloud. """
 
-import json
 import typing
-from typing import Dict, Iterator, List, Optional, Tuple
+from typing import Dict, Iterator, List, Optional, Tuple, Union
 
 from sky import clouds
 from sky.clouds import service_catalog
@@ -147,7 +146,7 @@ class RunPod(clouds.Cloud):
 
     @classmethod
     def get_accelerators_from_instance_type(
-            cls, instance_type: str) -> Optional[Dict[str, int]]:
+            cls, instance_type: str) -> Optional[Dict[str, Union[int, float]]]:
         return service_catalog.get_accelerators_from_instance_type(
             instance_type, clouds='runpod')
 
@@ -166,10 +165,8 @@ class RunPod(clouds.Cloud):
 
         r = resources
         acc_dict = self.get_accelerators_from_instance_type(r.instance_type)
-        if acc_dict is not None:
-            custom_resources = json.dumps(acc_dict, separators=(',', ':'))
-        else:
-            custom_resources = None
+        custom_resources = resources_utils.make_ray_custom_resources_str(
+            acc_dict)
 
         if r.image_id is None:
             image_id = 'runpod/base:0.0.2'

sky/clouds/scp.py

@@ -4,9 +4,8 @@ This module includes the set of functions
 to access the SCP catalog and check credentials for the SCP access.
 """
 
-import json
 import typing
-from typing import Dict, Iterator, List, Optional, Tuple
+from typing import Dict, Iterator, List, Optional, Tuple, Union
 
 from sky import clouds
 from sky import exceptions
@@ -160,7 +159,7 @@ class SCP(clouds.Cloud):
     def get_accelerators_from_instance_type(
         cls,
         instance_type: str,
-    ) -> Optional[Dict[str, int]]:
+    ) -> Optional[Dict[str, Union[int, float]]]:
         return service_catalog.get_accelerators_from_instance_type(
             instance_type, clouds='scp')
 
@@ -188,11 +187,9 @@ class SCP(clouds.Cloud):
 
         r = resources
         acc_dict = self.get_accelerators_from_instance_type(r.instance_type)
+        custom_resources = resources_utils.make_ray_custom_resources_str(
+            acc_dict)
 
-        if acc_dict is not None:
-            custom_resources = json.dumps(acc_dict, separators=(',', ':'))
-        else:
-            custom_resources = None
         image_id = self._get_image_id(r.image_id, region.name, r.instance_type)
         return {
             'instance_type': resources.instance_type,

sky/clouds/service_catalog/__init__.py

@@ -238,7 +238,7 @@ def get_default_instance_type(cpus: Optional[str] = None,
 
 def get_accelerators_from_instance_type(
         instance_type: str,
-        clouds: CloudFilter = None) -> Optional[Dict[str, int]]:
+        clouds: CloudFilter = None) -> Optional[Dict[str, Union[int, float]]]:
     """Returns the accelerators from a instance type."""
     return _map_clouds_catalog(clouds, 'get_accelerators_from_instance_type',
                                instance_type)

sky/clouds/service_catalog/aws_catalog.py

@@ -8,7 +8,7 @@ import hashlib
 import os
 import threading
 import typing
-from typing import Dict, List, Optional, Tuple
+from typing import Dict, List, Optional, Tuple, Union
 
 from sky import exceptions
 from sky import sky_logging
@@ -243,7 +243,7 @@ def get_default_instance_type(
 
 
 def get_accelerators_from_instance_type(
-        instance_type: str) -> Optional[Dict[str, int]]:
+        instance_type: str) -> Optional[Dict[str, Union[int, float]]]:
     return common.get_accelerators_from_instance_type_impl(
         _get_df(), instance_type)
 

sky/clouds/service_catalog/azure_catalog.py

@@ -4,14 +4,17 @@ This module loads the service catalog file and can be used to query
 instance types and pricing information for Azure.
 """
 import re
-from typing import Dict, List, Optional, Tuple
+from typing import Dict, List, Optional, Tuple, Union
 
 from sky import clouds as cloud_lib
+from sky import sky_logging
 from sky.clouds import Azure
 from sky.clouds.service_catalog import common
 from sky.utils import resources_utils
 from sky.utils import ux_utils
 
+logger = sky_logging.init_logger(__name__)
+
 # This list should match the list of regions in
 # skypilot image generation Packer script's replication_regions
 # sky/clouds/service_catalog/images/skypilot-azure-cpu-ubuntu.pkr.hcl
@@ -134,7 +137,7 @@ def get_default_instance_type(
 
 
 def get_accelerators_from_instance_type(
-        instance_type: str) -> Optional[Dict[str, int]]:
+        instance_type: str) -> Optional[Dict[str, Union[int, float]]]:
     return common.get_accelerators_from_instance_type_impl(_df, instance_type)
 
 
@@ -154,6 +157,7 @@ def get_instance_type_for_accelerator(
     if zone is not None:
         with ux_utils.print_exception_no_traceback():
             raise ValueError('Azure does not support zones.')
+
     return common.get_instance_type_for_accelerator_impl(df=_df,
                                                          acc_name=acc_name,
                                                          acc_count=acc_count,
@@ -191,9 +195,16 @@ def list_accelerators(
 
 def get_image_id_from_tag(tag: str, region: Optional[str]) -> Optional[str]:
     """Returns the image id from the tag."""
-    # Azure images are not region-specific.
-    del region  # Unused.
-    return common.get_image_id_from_tag_impl(_image_df, tag, None)
+    global _image_df
+    image_id = common.get_image_id_from_tag_impl(_image_df, tag, region)
+    if image_id is None:
+        # Refresh the image catalog and try again, if the image tag is not
+        # found.
+        logger.debug('Refreshing the image catalog and trying again.')
+        _image_df = common.read_catalog('azure/images.csv',
+                                        pull_frequency_hours=0)
+        image_id = common.get_image_id_from_tag_impl(_image_df, tag, region)
+    return image_id
 
 
 def is_image_tag_valid(tag: str, region: Optional[str]) -> bool:

sky/clouds/service_catalog/common.py

@@ -5,7 +5,7 @@ import hashlib
 import os
 import time
 import typing
-from typing import Callable, Dict, List, NamedTuple, Optional, Tuple
+from typing import Callable, Dict, List, NamedTuple, Optional, Tuple, Union
 
 import filelock
 import requests
@@ -481,7 +481,7 @@ def get_instance_type_for_cpus_mem_impl(
 def get_accelerators_from_instance_type_impl(
     df: 'pd.DataFrame',
     instance_type: str,
-) -> Optional[Dict[str, int]]:
+) -> Optional[Dict[str, Union[int, float]]]:
     df = _get_instance_type(df, instance_type, None)
     if len(df) == 0:
         with ux_utils.print_exception_no_traceback():
@@ -490,13 +490,19 @@ def get_accelerators_from_instance_type_impl(
     acc_name, acc_count = row['AcceleratorName'], row['AcceleratorCount']
     if pd.isnull(acc_name):
         return None
-    return {acc_name: int(acc_count)}
+
+    def _convert(value):
+        if int(value) == value:
+            return int(value)
+        return float(value)
+
+    return {acc_name: _convert(acc_count)}
 
 
 def get_instance_type_for_accelerator_impl(
     df: 'pd.DataFrame',
     acc_name: str,
-    acc_count: int,
+    acc_count: Union[int, float],
     cpus: Optional[str] = None,
     memory: Optional[str] = None,
     use_spot: bool = False,
@@ -509,7 +515,7 @@ def get_instance_type_for_accelerator_impl(
     accelerators with sorted prices and a list of candidates with fuzzy search.
     """
     result = df[(df['AcceleratorName'].str.fullmatch(acc_name, case=False)) &
-                (df['AcceleratorCount'] == acc_count)]
+                (abs(df['AcceleratorCount'] - acc_count) <= 0.01)]
     result = _filter_region_zone(result, region, zone)
     if len(result) == 0:
         fuzzy_result = df[
@@ -522,8 +528,11 @@ def get_instance_type_for_accelerator_impl(
         fuzzy_candidate_list = []
         if len(fuzzy_result) > 0:
             for _, row in fuzzy_result.iterrows():
+                acc_cnt = float(row['AcceleratorCount'])
+                acc_count_display = (int(acc_cnt) if acc_cnt.is_integer() else
+                                     f'{acc_cnt:.2f}')
                 fuzzy_candidate_list.append(f'{row["AcceleratorName"]}:'
-                                            f'{int(row["AcceleratorCount"])}')
+                                            f'{acc_count_display}')
         return (None, fuzzy_candidate_list)
 
     result = _filter_with_cpus(result, cpus)