diff --git a/python/sglang/srt/layers/attention/dual_chunk_flashattention_backend.py b/python/sglang/srt/layers/attention/dual_chunk_flashattention_backend.py
index ffcd06131..4c9c8ea46 100644
--- a/python/sglang/srt/layers/attention/dual_chunk_flashattention_backend.py
+++ b/python/sglang/srt/layers/attention/dual_chunk_flashattention_backend.py
@@ -126,11 +126,7 @@ class DualChunkFlashAttentionBackend(AttentionBackend):
         self.req_to_token = model_runner.req_to_token_pool.req_to_token
         self.kv_cache_dtype = model_runner.kv_cache_dtype
 
-        self.kv_cache_dtype_str = getattr(
-            model_runner,
-            "kv_cache_dtype_str",
-            model_runner.server_args.kv_cache_dtype,
-        )
+        self.kv_cache_dtype_str = model_runner.kv_cache_dtype_str
         self.page_size = model_runner.page_size
 
         assert self.num_heads % self.num_kv_heads == 0
diff --git a/python/sglang/srt/layers/attention/flashattention_backend.py b/python/sglang/srt/layers/attention/flashattention_backend.py
index 973b2a2dd..ac440dbab 100644
--- a/python/sglang/srt/layers/attention/flashattention_backend.py
+++ b/python/sglang/srt/layers/attention/flashattention_backend.py
@@ -166,9 +166,7 @@ class FlashAttentionBackend(AttentionBackend):
         self.token_to_kv_pool = model_runner.token_to_kv_pool
         self.req_to_token = model_runner.req_to_token_pool.req_to_token
         self.kv_cache_dtype = model_runner.kv_cache_dtype
-        from sglang.srt.runtime_context import get_model
-
-        self.kv_cache_dtype_str = get_model().kv_cache_dtype
+        self.kv_cache_dtype_str = model_runner.kv_cache_dtype_str
         self.kv_cache_is_mxfp8 = self.kv_cache_dtype_str == "mxfp8"
         self.page_size = model_runner.page_size
         # Static page-table width (upper bound). The device-side page-table build
diff --git a/python/sglang/srt/layers/attention/linear/lightning_backend.py b/python/sglang/srt/layers/attention/linear/lightning_backend.py
index 2a46ae1a1..6cd3c629e 100644
--- a/python/sglang/srt/layers/attention/linear/lightning_backend.py
+++ b/python/sglang/srt/layers/attention/linear/lightning_backend.py
@@ -65,11 +65,7 @@ class LightningAttentionBackend(MambaAttnBackendBase):
         self.decode_cuda_graph_metadata = {}
         self.kv_cache_dtype = model_runner.kv_cache_dtype
 
-        self.kv_cache_dtype_str = getattr(
-            model_runner,
-            "kv_cache_dtype_str",
-            model_runner.server_args.kv_cache_dtype,
-        )
+        self.kv_cache_dtype_str = model_runner.kv_cache_dtype_str
         self.BLOCK = (
             model_runner.model_config.block
             if hasattr(model_runner.model_config, "block")
diff --git a/python/sglang/srt/layers/attention/xpu_backend.py b/python/sglang/srt/layers/attention/xpu_backend.py
index 27ad57ce5..2b6b469ca 100644
--- a/python/sglang/srt/layers/attention/xpu_backend.py
+++ b/python/sglang/srt/layers/attention/xpu_backend.py
@@ -69,9 +69,7 @@ class XPUAttentionBackend(AttentionBackend):
         self.token_to_kv_pool = model_runner.token_to_kv_pool
         self.req_to_token = model_runner.req_to_token_pool.req_to_token
         self.kv_cache_dtype = model_runner.kv_cache_dtype
-        from sglang.srt.runtime_context import get_model
-
-        self.kv_cache_dtype_str = get_model().kv_cache_dtype
+        self.kv_cache_dtype_str = model_runner.kv_cache_dtype_str
         self.page_size = model_runner.page_size
         self.use_mla = model_runner.model_config.attention_arch == AttentionArch.MLA
         self.skip_prefill = skip_prefill
diff --git a/python/sglang/test/kits/attention_unittest/attention_methods/dense_attention.py b/python/sglang/test/kits/attention_unittest/attention_methods/dense_attention.py
index 2c48ae75e..a69f07d06 100644
--- a/python/sglang/test/kits/attention_unittest/attention_methods/dense_attention.py
+++ b/python/sglang/test/kits/attention_unittest/attention_methods/dense_attention.py
@@ -318,6 +318,7 @@ class MockModelRunner(ModelRunner):
         self.device = device
         self.dtype = dtype
         self.kv_cache_dtype = dtype
+        self.kv_cache_dtype_str = "auto"
         self.gpu_id = 0
         self.canary_manager = None
         self.page_size = case.page_size
diff --git a/python/sglang/test/kits/attention_unittest/attention_methods/dsa_attention.py b/python/sglang/test/kits/attention_unittest/attention_methods/dsa_attention.py
index 62b6042a0..7fc5f7cb6 100644
--- a/python/sglang/test/kits/attention_unittest/attention_methods/dsa_attention.py
+++ b/python/sglang/test/kits/attention_unittest/attention_methods/dsa_attention.py
@@ -289,6 +289,7 @@ class DSAMockModelRunner(ModelRunner):
         # 656 bytes/token while the model still projects K/V in BF16;
         # `set_mla_kv_buffer` does the quantize on the way in.
         self.kv_cache_dtype = torch.float8_e4m3fn if fp8_kv_cache else dtype
+        self.kv_cache_dtype_str = "auto"
         # For TARGET_VERIFY / DRAFT_EXTEND, the DSA backend uses
         # `self.speculative_num_draft_tokens` to size `seqlens_expanded`
         # (`dsa_backend.py:482-486,510-515`). When zero, deep_gemm's
diff --git a/python/sglang/test/kits/attention_unittest/attention_methods/dsv4_attention.py b/python/sglang/test/kits/attention_unittest/attention_methods/dsv4_attention.py
index 0d41fef66..626b6776c 100644
--- a/python/sglang/test/kits/attention_unittest/attention_methods/dsv4_attention.py
+++ b/python/sglang/test/kits/attention_unittest/attention_methods/dsv4_attention.py
@@ -328,6 +328,7 @@ class MockDSV4ModelRunner:
         self.device = device
         self.dtype = dtype
         self.kv_cache_dtype = dtype
+        self.kv_cache_dtype_str = "auto"
         self.gpu_id = 0
         self.canary_manager = None
         self.page_size = case.page_size
diff --git a/python/sglang/test/kits/attention_unittest/attention_methods/dual_chunk_attention.py b/python/sglang/test/kits/attention_unittest/attention_methods/dual_chunk_attention.py
index a0846f6c8..7802aef15 100644
--- a/python/sglang/test/kits/attention_unittest/attention_methods/dual_chunk_attention.py
+++ b/python/sglang/test/kits/attention_unittest/attention_methods/dual_chunk_attention.py
@@ -318,6 +318,7 @@ class DualChunkMockModelRunner(ModelRunner):
         self.device = device
         self.dtype = dtype
         self.kv_cache_dtype = dtype
+        self.kv_cache_dtype_str = "auto"
         self.gpu_id = 0
         self.canary_manager = None
         self.page_size = case.page_size
diff --git a/python/sglang/test/kits/attention_unittest/attention_methods/gdn_attention.py b/python/sglang/test/kits/attention_unittest/attention_methods/gdn_attention.py
index f42b26a83..409a92767 100644
--- a/python/sglang/test/kits/attention_unittest/attention_methods/gdn_attention.py
+++ b/python/sglang/test/kits/attention_unittest/attention_methods/gdn_attention.py
@@ -212,6 +212,7 @@ class MockGDNModelRunner(ModelRunner):
         self.device = device
         self.dtype = dtype
         self.kv_cache_dtype = dtype
+        self.kv_cache_dtype_str = "auto"
         self.gpu_id = 0
         self.ps = ParallelState.trivial()
         self.canary_manager = None
diff --git a/python/sglang/test/kits/attention_unittest/attention_methods/kda_attention.py b/python/sglang/test/kits/attention_unittest/attention_methods/kda_attention.py
index 4799286f0..481e724a6 100644
--- a/python/sglang/test/kits/attention_unittest/attention_methods/kda_attention.py
+++ b/python/sglang/test/kits/attention_unittest/attention_methods/kda_attention.py
@@ -218,6 +218,7 @@ class MockKDAModelRunner(ModelRunner):
         self.device = device
         self.dtype = dtype
         self.kv_cache_dtype = dtype
+        self.kv_cache_dtype_str = "auto"
         self.gpu_id = 0
         self.ps = ParallelState.trivial()
         self.canary_manager = None
diff --git a/python/sglang/test/kits/attention_unittest/attention_methods/lightning_attention.py b/python/sglang/test/kits/attention_unittest/attention_methods/lightning_attention.py
index 2f1de2d33..bbb0235e9 100644
--- a/python/sglang/test/kits/attention_unittest/attention_methods/lightning_attention.py
+++ b/python/sglang/test/kits/attention_unittest/attention_methods/lightning_attention.py
@@ -226,6 +226,7 @@ class MockLightningModelRunner(ModelRunner):
         self.device = device
         self.dtype = dtype
         self.kv_cache_dtype = dtype
+        self.kv_cache_dtype_str = "auto"
         self.gpu_id = 0
         self.ps = ParallelState.trivial()
         self.canary_manager = None
diff --git a/python/sglang/test/kits/attention_unittest/attention_methods/mamba2_attention.py b/python/sglang/test/kits/attention_unittest/attention_methods/mamba2_attention.py
index 6a98a1fcb..d1f91810d 100644
--- a/python/sglang/test/kits/attention_unittest/attention_methods/mamba2_attention.py
+++ b/python/sglang/test/kits/attention_unittest/attention_methods/mamba2_attention.py
@@ -311,6 +311,7 @@ class MockMamba2ModelRunner(ModelRunner):
         self.device = device
         self.dtype = dtype
         self.kv_cache_dtype = dtype
+        self.kv_cache_dtype_str = "auto"
         self.gpu_id = 0
         self.ps = ParallelState.trivial()
         self.canary_manager = None
diff --git a/python/sglang/test/kits/attention_unittest/attention_methods/mla_attention.py b/python/sglang/test/kits/attention_unittest/attention_methods/mla_attention.py
index cfe81c0b0..3fd845321 100644
--- a/python/sglang/test/kits/attention_unittest/attention_methods/mla_attention.py
+++ b/python/sglang/test/kits/attention_unittest/attention_methods/mla_attention.py
@@ -229,6 +229,7 @@ class MockMLAModelRunner(ModelRunner):
         # while the model still projects K/V in bf16; `set_mla_kv_buffer`
         # does the BF16->FP8 cast on the way in.
         self.kv_cache_dtype = torch.float8_e4m3fn if fp8_kv_cache else dtype
+        self.kv_cache_dtype_str = "fp8_e4m3" if fp8_kv_cache else "auto"
         self.gpu_id = 0
         self.canary_manager = None
         self.page_size = case.page_size
diff --git a/test/manual/attention/test_flashattn_backend.py b/test/manual/attention/test_flashattn_backend.py
index 51c04968d..0122ef1f2 100644
--- a/test/manual/attention/test_flashattn_backend.py
+++ b/test/manual/attention/test_flashattn_backend.py
@@ -50,13 +50,13 @@ class MockModelRunner:
         self.kv_cache_dtype = (
             self.dtype
         )  # torch dtype, required by FlashAttentionBackend
+        self.kv_cache_dtype_str = "auto"
 
-        # server_args is still needed for string-based config (kv_cache_dtype_str)
         self.server_args = type(
             "ServerArgs",
             (),
             {
-                "kv_cache_dtype": "auto",  # string version for kv_cache_dtype_str
+                "kv_cache_dtype": "auto",
                 "speculative_eagle_topk": None,
                 "speculative_num_draft_tokens": 0,
                 "enable_deterministic_inference": False,
diff --git a/test/manual/attention/test_flashattn_mla_backend.py b/test/manual/attention/test_flashattn_mla_backend.py
index d19f992b6..7bc8dd29e 100644
--- a/test/manual/attention/test_flashattn_mla_backend.py
+++ b/test/manual/attention/test_flashattn_mla_backend.py
@@ -49,6 +49,7 @@ class MockModelRunner:
             },
         )
         self.kv_cache_dtype = self.server_args.kv_cache_dtype
+        self.kv_cache_dtype_str = "auto"
 
         batch_size = 160
         # Create a proper req_to_token_pool with the req_to_token attribute