diff --git a/engines/python/setup/djl_python/properties_manager/vllm_rb_properties.py b/engines/python/setup/djl_python/properties_manager/vllm_rb_properties.py
index f0a295ace..84318de1a 100644
--- a/engines/python/setup/djl_python/properties_manager/vllm_rb_properties.py
+++ b/engines/python/setup/djl_python/properties_manager/vllm_rb_properties.py
@@ -59,6 +59,22 @@ class VllmRbProperties(Properties):
     enable_prefix_caching: Optional[bool] = False
     disable_sliding_window: Optional[bool] = False
     limit_mm_per_prompt: Optional[Mapping[str, int]] = None
+    use_v2_block_manager: bool = False
+
+    # Speculative decoding configuration.
+    speculative_model: Optional[str] = None
+    speculative_model_quantization: Optional[str] = None
+    speculative_draft_tensor_parallel_size: Optional[int] = None
+    num_speculative_tokens: Optional[int] = None
+    speculative_max_model_len: Optional[int] = None
+    speculative_disable_by_batch_size: Optional[int] = None
+    ngram_prompt_lookup_max: Optional[int] = None
+    ngram_prompt_lookup_min: Optional[int] = None
+    spec_decoding_acceptance_method: str = 'rejection_sampler'
+    typical_acceptance_sampler_posterior_threshold: Optional[float] = None
+    typical_acceptance_sampler_posterior_alpha: Optional[float] = None
+    qlora_adapter_name_or_path: Optional[str] = None
+    disable_logprobs_during_spec_decoding: Optional[bool] = None
 
     @field_validator('engine')
     def validate_engine(cls, engine):
diff --git a/engines/python/setup/djl_python/rolling_batch/rolling_batch_vllm_utils.py b/engines/python/setup/djl_python/rolling_batch/rolling_batch_vllm_utils.py
index 0da4e322b..d0a250283 100644
--- a/engines/python/setup/djl_python/rolling_batch/rolling_batch_vllm_utils.py
+++ b/engines/python/setup/djl_python/rolling_batch/rolling_batch_vllm_utils.py
@@ -266,6 +266,27 @@ def get_engine_args_from_config(config: VllmRbProperties) -> EngineArgs:
             enable_prefix_caching=config.enable_prefix_caching,
             disable_sliding_window=config.disable_sliding_window,
             max_num_seqs=config.max_rolling_batch_size,
+            use_v2_block_manager=config.use_v2_block_manager,
+            speculative_model=config.speculative_model,
+            speculative_model_quantization=config.
+            speculative_model_quantization,
+            speculative_draft_tensor_parallel_size=config.
+            speculative_draft_tensor_parallel_size,
+            num_speculative_tokens=config.num_speculative_tokens,
+            speculative_max_model_len=config.speculative_max_model_len,
+            speculative_disable_by_batch_size=config.
+            speculative_disable_by_batch_size,
+            ngram_prompt_lookup_max=config.ngram_prompt_lookup_max,
+            ngram_prompt_lookup_min=config.ngram_prompt_lookup_min,
+            spec_decoding_acceptance_method=config.
+            spec_decoding_acceptance_method,
+            typical_acceptance_sampler_posterior_threshold=config.
+            typical_acceptance_sampler_posterior_threshold,
+            typical_acceptance_sampler_posterior_alpha=config.
+            typical_acceptance_sampler_posterior_alpha,
+            qlora_adapter_name_or_path=config.qlora_adapter_name_or_path,
+            disable_logprobs_during_spec_decoding=config.
+            disable_logprobs_during_spec_decoding,
         )
 
 
diff --git a/tests/integration/llm/client.py b/tests/integration/llm/client.py
index 2216225c8..eb14b23a8 100644
--- a/tests/integration/llm/client.py
+++ b/tests/integration/llm/client.py
@@ -463,6 +463,18 @@ def get_model_name():
         "seq_length": [256],
         "tokenizer": "tiiuae/falcon-11B"
     },
+    "llama-68m-speculative-medusa": {
+        "max_memory_per_gpu": [25.0],
+        "batch_size": [1, 4],
+        "seq_length": [256],
+        "tokenizer": "JackFram/llama-68m"
+    },
+    "llama-68m-speculative-eagle": {
+        "max_memory_per_gpu": [25.0],
+        "batch_size": [1, 4],
+        "seq_length": [256],
+        "tokenizer": "JackFram/llama-68m"
+    },
     "llama-7b-unmerged-lora": {
         "max_memory_per_gpu": [15.0, 15.0],
         "batch_size": [3],
diff --git a/tests/integration/llm/prepare.py b/tests/integration/llm/prepare.py
index 656946ea7..c07a1c38a 100644
--- a/tests/integration/llm/prepare.py
+++ b/tests/integration/llm/prepare.py
@@ -625,6 +625,24 @@
         "option.tensor_parallel_degree": 4,
         "option.enable_chunked_prefill": "true",
     },
+    "llama-68m-speculative-medusa": {
+        "option.model_id": "s3://djl-llm/llama-68m/",
+        "option.task": "text-generation",
+        "option.speculative_model": "s3://djl-llm/llama-2-tiny/",
+        "option.num_speculative_tokens": 4,
+        "option.use_v2_block_manager": True,
+        "option.tensor_parallel_degree": 1,
+        "option.max_rolling_batch_size": 4,
+    },
+    "llama-68m-speculative-eagle": {
+        "option.model_id": "s3://djl-llm/llama-68m/",
+        "option.task": "text-generation",
+        "option.speculative_model": "abhigoyal/vllm-eagle-llama-68m-random",
+        "option.num_speculative_tokens": 4,
+        "option.use_v2_block_manager": True,
+        "option.tensor_parallel_degree": 1,
+        "option.max_rolling_batch_size": 4,
+    },
     "llama-7b-unmerged-lora": {
         "option.model_id": "s3://djl-llm/huggyllama-llama-7b",
         "option.tensor_parallel_degree": "max",
diff --git a/tests/integration/tests.py b/tests/integration/tests.py
index 1f085ce38..ae9e73e24 100644
--- a/tests/integration/tests.py
+++ b/tests/integration/tests.py
@@ -604,6 +604,18 @@ def test_falcon_11b_chunked_prefill(self):
             client.run(
                 "vllm falcon-11b-chunked-prefill --in_tokens 1200".split())
 
+    def test_llama_68m_speculative_medusa(self):
+        with Runner('lmi', 'llama-68m-speculative-medusa') as r:
+            prepare.build_vllm_model("llama-68m-speculative-medusa")
+            r.launch()
+            client.run("vllm llama-68m-speculative-medusa".split())
+
+    def test_llama_68m_speculative_eagle(self):
+        with Runner('lmi', 'llama-68m-speculative-eagle') as r:
+            prepare.build_vllm_model("llama-68m-speculative-eagle")
+            r.launch()
+            client.run("vllm llama-68m-speculative-eagle".split())
+
 
 @pytest.mark.vllm
 @pytest.mark.lora