diff --git a/benchmarks/kernels/benchmark_flydsl_moe_w4a16.py b/benchmarks/kernels/benchmark_flydsl_moe_w4a16.py
index 9e4f4157a8a..09a01b301be 100644
--- a/benchmarks/kernels/benchmark_flydsl_moe_w4a16.py
+++ b/benchmarks/kernels/benchmark_flydsl_moe_w4a16.py
@@ -17,7 +17,7 @@ from vllm.model_executor.layers.fused_moe.fused_flydsl_moe import fused_flydsl_m
 from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe import (  # noqa: E501
     compressed_tensors_moe_w4a16_flydsl,
 )
-from vllm.platforms import current_platform
+from vllm.utils.platform_utils import get_device_name_as_file_name
 
 RoutingBuffers = tuple[
     torch.Tensor,  # sorted_token_ids
@@ -259,7 +259,7 @@ def tune_flydsl_moe_w4a16(
                             )
                             us_best = us
                             tuned_config[str(num_tokens)] = tile_config
-            device_name = current_platform.get_device_name().replace(" ", "_")
+            device_name = get_device_name_as_file_name()
             tuned_config_file_name = (
                 f"E={num_experts},N={inter_dim},device_name={device_name},"
                 f"dtype=int4_w4a16,backend=flydsl.json"
diff --git a/benchmarks/kernels/benchmark_w8a8_block_fp8.py b/benchmarks/kernels/benchmark_w8a8_block_fp8.py
index 36dce1b6388..590d4cfdc6d 100644
--- a/benchmarks/kernels/benchmark_w8a8_block_fp8.py
+++ b/benchmarks/kernels/benchmark_w8a8_block_fp8.py
@@ -19,6 +19,7 @@ from vllm.model_executor.layers.quantization.utils.fp8_utils import (
 from vllm.platforms import current_platform
 from vllm.triton_utils import triton
 from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.utils.platform_utils import get_device_name_as_file_name
 
 mp.set_start_method("spawn", force=True)
 
@@ -264,7 +265,7 @@ def save_configs(
     input_type="fp8",
 ) -> None:
     os.makedirs(save_path, exist_ok=True)
-    device_name = current_platform.get_device_name().replace(" ", "_")
+    device_name = get_device_name_as_file_name()
     json_file_name = (
         f"N={N},K={K},device_name={device_name},dtype={input_type}_w8a8,"
         f"block_shape=[{block_n},{block_k}].json"
diff --git a/vllm/model_executor/layers/fused_moe/fused_flydsl_moe.py b/vllm/model_executor/layers/fused_moe/fused_flydsl_moe.py
index cf49e01e628..b1588c4a2a4 100644
--- a/vllm/model_executor/layers/fused_moe/fused_flydsl_moe.py
+++ b/vllm/model_executor/layers/fused_moe/fused_flydsl_moe.py
@@ -15,7 +15,7 @@ from aiter.ops.flydsl.kernels.moe_gemm_2stage import (
 )
 
 from vllm.logger import init_logger
-from vllm.platforms import current_platform
+from vllm.utils.platform_utils import get_device_name_as_file_name
 from vllm.utils.torch_utils import direct_register_custom_op
 
 logger = init_logger(__name__)
@@ -106,7 +106,7 @@ def build_routing_buffers(
 
 @functools.lru_cache
 def try_get_optimal_config(num_experts, inter_dim):
-    device_name = current_platform.get_device_name().replace(" ", "_")
+    device_name = get_device_name_as_file_name()
     json_file_name = (
         f"E={num_experts},N={inter_dim},device_name={device_name},"
         "dtype=int4_w4a16,backend=flydsl.json"
diff --git a/vllm/model_executor/layers/fused_moe/fused_moe.py b/vllm/model_executor/layers/fused_moe/fused_moe.py
index 269b6e3da0b..62cf12ea24e 100644
--- a/vllm/model_executor/layers/fused_moe/fused_moe.py
+++ b/vllm/model_executor/layers/fused_moe/fused_moe.py
@@ -31,6 +31,7 @@ from vllm.model_executor.layers.fused_moe.utils import (
 )
 from vllm.platforms import current_platform
 from vllm.triton_utils import tl, triton
+from vllm.utils.platform_utils import get_device_name_as_file_name
 from vllm.utils.torch_utils import direct_register_custom_op
 
 logger = init_logger(__name__)
@@ -1034,7 +1035,7 @@ def zero_experts_compute_triton(
 def get_config_file_name(
     E: int, N: int, dtype: str | None, block_shape: list[int] | None = None
 ) -> str:
-    device_name = current_platform.get_device_name().replace(" ", "_")
+    device_name = get_device_name_as_file_name()
     # Set device_name to H200 if a device from the H200 family is detected
     if "H200" in device_name.split("_"):
         device_name = "NVIDIA_H200"
diff --git a/vllm/model_executor/layers/mamba/ops/mamba_ssm.py b/vllm/model_executor/layers/mamba/ops/mamba_ssm.py
index 8c5a6355803..d348defcc76 100644
--- a/vllm/model_executor/layers/mamba/ops/mamba_ssm.py
+++ b/vllm/model_executor/layers/mamba/ops/mamba_ssm.py
@@ -19,6 +19,7 @@ from vllm.logger import init_logger
 from vllm.model_executor.layers.mamba.ops.triton_helpers import fast_exp
 from vllm.platforms import current_platform
 from vllm.triton_utils import HAS_TRITON, tl, triton
+from vllm.utils.platform_utils import get_device_name_as_file_name
 from vllm.v1.attention.backends.utils import NULL_BLOCK_ID
 
 if current_platform.is_xpu():
@@ -53,7 +54,7 @@ def get_ssm_config_file_name(
 
 
 def get_ssm_device_name() -> str:
-    return current_platform.get_device_name().replace(" ", "_")
+    return get_device_name_as_file_name()
 
 
 def _canonical_cache_dtype(cache_dtype: str) -> str:
diff --git a/vllm/model_executor/layers/quantization/utils/fp8_utils.py b/vllm/model_executor/layers/quantization/utils/fp8_utils.py
index 32a2d86899c..49eadc5a152 100644
--- a/vllm/model_executor/layers/quantization/utils/fp8_utils.py
+++ b/vllm/model_executor/layers/quantization/utils/fp8_utils.py
@@ -33,6 +33,7 @@ from vllm.utils.deep_gemm import (
     is_deep_gemm_e8m0_used,
     transform_sf_into_required_layout,
 )
+from vllm.utils.platform_utils import get_device_name_as_file_name
 from vllm.utils.torch_utils import direct_register_custom_op
 
 logger = init_logger(__name__)
@@ -864,7 +865,7 @@ def get_w8a8_block_fp8_configs(
 
     # First look up if an optimized configuration is available in the configs
     # directory
-    device_name = current_platform.get_device_name().replace(" ", "_")
+    device_name = get_device_name_as_file_name()
     json_file_name = f"N={N},K={K},device_name={device_name},dtype=fp8_w8a8,block_shape=[{block_n},{block_k}].json"  # noqa: E501
 
     config_file_path = os.path.join(
diff --git a/vllm/model_executor/layers/quantization/utils/int8_utils.py b/vllm/model_executor/layers/quantization/utils/int8_utils.py
index e0db2526948..4f624cf4963 100644
--- a/vllm/model_executor/layers/quantization/utils/int8_utils.py
+++ b/vllm/model_executor/layers/quantization/utils/int8_utils.py
@@ -12,6 +12,7 @@ import torch
 
 from vllm.platforms import current_platform
 from vllm.triton_utils import tl, triton
+from vllm.utils.platform_utils import get_device_name_as_file_name
 
 logger = logging.getLogger(__name__)
 
@@ -329,7 +330,7 @@ def get_w8a8_block_int8_configs(
 
     # First look up if an optimized configuration is available in the configs
     # directory
-    device_name = current_platform.get_device_name().replace(" ", "_")
+    device_name = get_device_name_as_file_name()
     json_file_name = f"N={N},K={K},device_name={device_name},dtype=int8_w8a8,block_shape=[{block_n}, {block_k}].json"  # noqa: E501
 
     config_file_path = os.path.join(
diff --git a/vllm/utils/platform_utils.py b/vllm/utils/platform_utils.py
index cc69d9a241c..5d7fed3c990 100644
--- a/vllm/utils/platform_utils.py
+++ b/vllm/utils/platform_utils.py
@@ -7,6 +7,7 @@ from concurrent.futures.process import ProcessPoolExecutor
 from functools import cache
 from typing import Any
 
+import regex as re
 import torch
 
 
@@ -62,3 +63,12 @@ def num_compute_units(device_id: int = 0) -> int:
     from vllm.platforms import current_platform
 
     return current_platform.num_compute_units(device_id)
+
+
+@cache
+def get_device_name_as_file_name(device_id: int = 0) -> str:
+    from vllm.platforms import current_platform
+
+    name = current_platform.get_device_name(device_id)
+    name = re.sub(r"[\s/]+", "_", name)
+    return name