diff --git a/docs/design/attention_backends.md b/docs/design/attention_backends.md
index b665fde2cc..d70cd6fd74 100644
--- a/docs/design/attention_backends.md
+++ b/docs/design/attention_backends.md
@@ -170,8 +170,8 @@ Priority is **1 = highest** (tried first).
 | Backend | Version | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | MM Prefix | DCP | Attention Types | Compute Cap. |
 | ------- | ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | --------- | --- | --------------- | ------------ |
 | `CPU_ATTN` | | fp16, bf16, fp32 | `auto`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | 32, 64, 80, 96, 112, 128, 160, 192, 224, 256, 512 | ❌ | ❌ | ❌ | ❌ | All | N/A |
-| `FLASHINFER` | Native† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ❌ | ❌ | ❌ | ✅ | Decoder | 7.x-9.x |
-| `FLASHINFER` | TRTLLM† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `nvfp4` | 16, 32, 64 | 64, 128, 256 | ✅ | ❌ | ❌ | ✅ | Decoder | 10.x |
+| `FLASHINFER` | Native† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256, 512 | ❌ | ❌ | ❌ | ✅ | Decoder | 7.x-9.x |
+| `FLASHINFER` | TRTLLM† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `nvfp4` | 16, 32, 64 | 64, 128, 256, 512 | ✅ | ❌ | ❌ | ✅ | Decoder | 10.x |
 | `FLASH_ATTN` | FA2* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ✅ | ❌ | ✅ | All | ≥8.0 |
 | `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | 9.x |
 | `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | ≥10.0 |
diff --git a/vllm/v1/attention/backends/flashinfer.py b/vllm/v1/attention/backends/flashinfer.py
index 4abc014d1f..a81c5742c1 100755
--- a/vllm/v1/attention/backends/flashinfer.py
+++ b/vllm/v1/attention/backends/flashinfer.py
@@ -402,7 +402,7 @@ class FlashInferBackend(AttentionBackend):
     @classmethod
     def get_supported_head_sizes(cls) -> list[int]:
         # https://github.com/flashinfer-ai/flashinfer/blob/3d55c71a62052c590c130897d3a3db49b14fcc34/include/flashinfer/utils.cuh#L157
-        return [64, 128, 256]
+        return [64, 128, 256, 512]
 
     @classmethod
     def supports_compute_capability(cls, capability: DeviceCapability) -> bool:
@@ -1451,15 +1451,16 @@ class FlashInferImpl(AttentionImpl):
 
         num_actual_tokens = attn_metadata.num_actual_tokens
 
-        # The FlashInfer api requires data to be in fp8_e4m3 or fp8_e5m2
-        # to process the cache when the kv_cache_dtype is fp8
-        if self.kv_sharing_target_layer_name is None and is_quantized_kv_cache(
-            self.kv_cache_dtype
-        ):
-            torch_dtype = FlashInferBackend.get_dtype_for_flashinfer(
-                self.kv_cache_dtype
-            )
-            kv_cache = kv_cache.view(torch_dtype)
+        # FlashInfer treats uint8 KV cache as NVFP4. vLLM stores FP8 KV cache
+        # as uint8 bytes, so pass FP8 caches with their logical dtype.
+        if not self.is_kvcache_nvfp4 and kv_cache.dtype == torch.uint8:
+            fp8_view_dtype = None
+            if self.kv_cache_dtype in ("fp8", "fp8_e4m3", torch.float8_e4m3fn):
+                fp8_view_dtype = torch.float8_e4m3fn
+            elif self.kv_cache_dtype in ("fp8_e5m2", torch.float8_e5m2):
+                fp8_view_dtype = torch.float8_e5m2
+            if fp8_view_dtype is not None:
+                kv_cache = kv_cache.view(fp8_view_dtype)
 
         # Inputs and outputs may be padded for CUDA graphs
         query = query[:num_actual_tokens]