diff --git a/vllm/config/speculative.py b/vllm/config/speculative.py
index b9c612621982..47536695b190 100644
--- a/vllm/config/speculative.py
+++ b/vllm/config/speculative.py
@@ -10,6 +10,7 @@
 from typing_extensions import Self
 
 from vllm.config import LoadConfig
+from vllm.config.cache import CacheDType
 from vllm.config.kernel import MoEBackend
 from vllm.config.model import HfOverrides, ModelConfig
 from vllm.config.parallel import ParallelConfig
@@ -118,6 +119,9 @@ class SpeculativeConfig:
     """Attention backend to use for the draft model. When `None`, the backend is
     automatically selected. Useful when the drafter requires a different attention
     backend (e.g. DFlash needs a non-causal-capable backend like FLASH_ATTN)."""
+    kv_cache_dtype: CacheDType | None = None
+    """KV cache dtype for the draft model. When `None`, the draft inherits the
+    target model's `--kv-cache-dtype`."""
     max_model_len: int | None = Field(default=None, ge=1)
     """The maximum model length of the draft model. Used when testing the
     ability to skip speculation for some sequences."""
diff --git a/vllm/engine/arg_utils.py b/vllm/engine/arg_utils.py
index 742d62ac3698..07877b0812c7 100644
--- a/vllm/engine/arg_utils.py
+++ b/vllm/engine/arg_utils.py
@@ -1754,6 +1754,10 @@ def create_speculative_config(
         if self.speculative_config is None:
             return None
 
+        self.speculative_config = {
+            k.replace("-", "_"): v for k, v in self.speculative_config.items()
+        }
+
         # Note(Shangming): These parameters are not obtained from the cli arg
         # '--speculative-config' and must be passed in when creating the engine
         # config.
diff --git a/vllm/v1/spec_decode/llm_base_proposer.py b/vllm/v1/spec_decode/llm_base_proposer.py
index 756c5f3b3717..f8b52d079a89 100644
--- a/vllm/v1/spec_decode/llm_base_proposer.py
+++ b/vllm/v1/spec_decode/llm_base_proposer.py
@@ -1300,6 +1300,15 @@ def _create_draft_vllm_config(self) -> VllmConfig:
             ),
         )
 
+        if spec_cfg.kv_cache_dtype is not None:
+            base = replace(
+                base,
+                cache_config=replace(
+                    base.cache_config,
+                    cache_dtype=spec_cfg.kv_cache_dtype,
+                ),
+            )
+
         return base
 
     def _get_model(self) -> nn.Module:
diff --git a/vllm/v1/worker/gpu/spec_decode/dflash/utils.py b/vllm/v1/worker/gpu/spec_decode/dflash/utils.py
index 998a28b6aee0..478274a05671 100644
--- a/vllm/v1/worker/gpu/spec_decode/dflash/utils.py
+++ b/vllm/v1/worker/gpu/spec_decode/dflash/utils.py
@@ -27,6 +27,14 @@ def load_dflash_model(target_model: nn.Module, vllm_config: VllmConfig) -> nn.Mo
             use_non_causal=dflash_has_any_non_causal(draft_model_config.hf_config),
             backend=speculative_config.attention_backend,
         ),
+        cache_config=(
+            replace(
+                vllm_config.cache_config,
+                cache_dtype=speculative_config.kv_cache_dtype,
+            )
+            if speculative_config.kv_cache_dtype is not None
+            else vllm_config.cache_config
+        ),
     )
     with set_model_tag("dflash_head"):
         dflash_model = get_model(
diff --git a/vllm/v1/worker/gpu/spec_decode/dspark/utils.py b/vllm/v1/worker/gpu/spec_decode/dspark/utils.py
index bbee614a1aeb..9ea2ff0f7363 100644
--- a/vllm/v1/worker/gpu/spec_decode/dspark/utils.py
+++ b/vllm/v1/worker/gpu/spec_decode/dspark/utils.py
@@ -27,6 +27,14 @@ def load_dspark_model(target_model: nn.Module, vllm_config: VllmConfig) -> nn.Mo
             use_non_causal=dflash_has_any_non_causal(draft_model_config.hf_config),
             backend=speculative_config.attention_backend,
         ),
+        cache_config=(
+            replace(
+                vllm_config.cache_config,
+                cache_dtype=speculative_config.kv_cache_dtype,
+            )
+            if speculative_config.kv_cache_dtype is not None
+            else vllm_config.cache_config
+        ),
     )
 
     with set_model_tag("dspark_head"):
diff --git a/vllm/v1/worker/gpu/spec_decode/eagle/utils.py b/vllm/v1/worker/gpu/spec_decode/eagle/utils.py
index bdd588e5786d..579652bc7d64 100644
--- a/vllm/v1/worker/gpu/spec_decode/eagle/utils.py
+++ b/vllm/v1/worker/gpu/spec_decode/eagle/utils.py
@@ -3,7 +3,7 @@
 import torch
 import torch.nn as nn
 
-from vllm.config import VllmConfig
+from vllm.config import VllmConfig, replace
 from vllm.distributed.parallel_state import get_pp_group
 from vllm.lora.layers.base import BaseLayerWithLoRA
 from vllm.model_executor.model_loader import get_model
@@ -39,6 +39,14 @@ def load_eagle_model(target_model: nn.Module, vllm_config: VllmConfig) -> nn.Mod
     speculative_config = vllm_config.speculative_config
     assert speculative_config is not None
     draft_model_config = speculative_config.draft_model_config
+    if speculative_config.kv_cache_dtype is not None:
+        vllm_config = replace(
+            vllm_config,
+            cache_config=replace(
+                vllm_config.cache_config,
+                cache_dtype=speculative_config.kv_cache_dtype,
+            ),
+        )
     with set_model_tag("eagle_head"):
         eagle_model = get_model(
             vllm_config=vllm_config, model_config=draft_model_config