diff --git a/vllm/models/kimi_k3/nvidia/mla.py b/vllm/models/kimi_k3/nvidia/mla.py
index 654f83b452..8f70574142 100644
--- a/vllm/models/kimi_k3/nvidia/mla.py
+++ b/vllm/models/kimi_k3/nvidia/mla.py
@@ -664,14 +664,10 @@ class MultiHeadLatentAttention(nn.Module, AttentionLayerBase):
                 cos_sin_cache=cos_sin_cache,
             )
         if is_quantized_kv_cache(self.kv_cache_dtype):
-            assert self.impl.supports_quant_query_input, (  # type: ignore[attr-defined]
-                "Kimi-K3 fp8 KV cache decode requires a backend that accepts an "
-                "fp8 (quantized) query input."
-            )
             cache = self.kv_cache
             if cache.dtype != torch.float8_e4m3fn:
                 cache = cache.view(torch.float8_e4m3fn)
-            return fused_mla_decode_q_concat_kv_cache_insert(
+            mqa_q = fused_mla_decode_q_concat_kv_cache_insert(
                 ql_nope,
                 q_pe,
                 kv_c_normed,
@@ -683,6 +679,12 @@ class MultiHeadLatentAttention(nn.Module, AttentionLayerBase):
                 positions=positions,
                 cos_sin_cache=cos_sin_cache,
             )
+            if not self.impl.supports_quant_query_input:  # type: ignore[attr-defined]
+                # Backend dequantizes fp8 KV on load and takes a bf16 query
+                # (e.g. TRITON_MLA, the DSpark draft); undo the query
+                # quantization.
+                mqa_q = (mqa_q.to(torch.float32) * self._q_scale).to(ql_nope.dtype)
+            return mqa_q
         return fused_mla_decode_q_concat_kv_cache_insert(
             ql_nope,
             q_pe,