diff --git a/requirements.txt b/requirements.txt
index 475490aa87e..d5a3a8ecb96 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -63,3 +63,4 @@ llguidance==0.7.29
 soundfile
 deep_gemm @ git+https://github.com/zongfeijing/DeepGEMM.git@a9d538ef4dff0326fe521c6ca0bfde115703b56a
 triton==3.3.1; platform_machine == "x86_64"
+blobfile
diff --git a/tensorrt_llm/llmapi/tokenizer.py b/tensorrt_llm/llmapi/tokenizer.py
index 858f98289cd..6e5f7bbcee0 100644
--- a/tensorrt_llm/llmapi/tokenizer.py
+++ b/tensorrt_llm/llmapi/tokenizer.py
@@ -160,7 +160,8 @@ def decode_incrementally(
         # HF incremental detokenization implementation is faster than TRTLLM when stream_interval is smaller.
         if (TLLM_INCREMENTAL_DETOKENIZATION_BACKEND == "TRTLLM"
                 or stream_interval >= TLLM_STREAM_INTERVAL_THRESHOLD
-                or spaces_between_special_tokens is False):
+                or spaces_between_special_tokens is False
+                or not hasattr(self.tokenizer, "_tokenizer")):
             return self.trtllm_decode_incrementally(
                 token_ids,
                 prev_text,
diff --git a/tests/integration/defs/accuracy/references/gsm8k.yaml b/tests/integration/defs/accuracy/references/gsm8k.yaml
index 353a0f71404..f69f02eaeb5 100644
--- a/tests/integration/defs/accuracy/references/gsm8k.yaml
+++ b/tests/integration/defs/accuracy/references/gsm8k.yaml
@@ -99,6 +99,9 @@ Qwen3/Qwen3-235B-A22B:
     quant_algo: NVFP4
     kv_cache_quant_algo: FP8
     accuracy: 85.78
+moonshotai/Kimi-K2-Instruct:
+  - quant_algo: FP8_BLOCK_SCALES
+    accuracy: 94.84
 nvidia/Llama-3_3-Nemotron-Super-49B-v1:
   - accuracy: 92.57
   - quant_algo: FP8
diff --git a/tests/integration/defs/accuracy/references/mmlu.yaml b/tests/integration/defs/accuracy/references/mmlu.yaml
index b4b76a8161b..485ad7c0295 100644
--- a/tests/integration/defs/accuracy/references/mmlu.yaml
+++ b/tests/integration/defs/accuracy/references/mmlu.yaml
@@ -187,6 +187,9 @@ Qwen3/Qwen3-235B-A22B:
     quant_algo: NVFP4
     kv_cache_quant_algo: FP8
     accuracy: 86
+moonshotai/Kimi-K2-Instruct:
+  - quant_algo: FP8_BLOCK_SCALES
+    accuracy: 87.65
 nvidia/Llama-3_3-Nemotron-Super-49B-v1:
   - accuracy: 79.43
   - quant_algo: FP8
diff --git a/tests/integration/defs/accuracy/test_llm_api_pytorch.py b/tests/integration/defs/accuracy/test_llm_api_pytorch.py
index a7a3ecd57e0..b4adb178498 100644
--- a/tests/integration/defs/accuracy/test_llm_api_pytorch.py
+++ b/tests/integration/defs/accuracy/test_llm_api_pytorch.py
@@ -1638,6 +1638,49 @@ def test_fp8_blockscale(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
             task.evaluate(llm)
 
 
+@pytest.mark.timeout(7200)
+@pytest.mark.skip_less_device_memory(100000)
+class TestKimiK2(LlmapiAccuracyTestHarness):
+    MODEL_NAME = "moonshotai/Kimi-K2-Instruct"
+    MODEL_PATH = f"{llm_models_root()}/Kimi-K2-Instruct"
+
+    @pytest.mark.skip_less_mpi_world_size(8)
+    @skip_pre_hopper
+    @pytest.mark.parametrize(
+        "tp_size,pp_size,ep_size,fp8kv,attention_dp,cuda_graph,overlap_scheduler,max_batch_size",
+        [(8, 1, 8, False, False, True, True, 16)],
+        ids=["latency"])
+    def test_fp8_blockscale(self, tp_size, pp_size, ep_size, fp8kv,
+                            attention_dp, cuda_graph, overlap_scheduler,
+                            max_batch_size):
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
+        pytorch_config = dict(
+            disable_overlap_scheduler=not overlap_scheduler,
+            cuda_graph_config=CudaGraphConfig() if cuda_graph else None,
+        )
+
+        if fp8kv:
+            kv_cache_config.dtype = "fp8"
+
+        mtp_config = None
+        with LLM(f"{llm_models_root()}/Kimi-K2-Instruct",
+                 max_batch_size=max_batch_size,
+                 tensor_parallel_size=tp_size,
+                 pipeline_parallel_size=pp_size,
+                 moe_expert_parallel_size=ep_size,
+                 trust_remote_code=True,
+                 kv_cache_config=kv_cache_config,
+                 **pytorch_config,
+                 enable_attention_dp=attention_dp,
+                 speculative_config=mtp_config) as llm:
+            assert llm.args.quant_config.quant_algo == QuantAlgo.FP8_BLOCK_SCALES
+
+            task = MMLU(self.MODEL_NAME)
+            task.evaluate(llm)
+            task = GSM8K(self.MODEL_NAME)
+            task.evaluate(llm)
+
+
 class TestMinitron4BBaseInstruct(LlmapiAccuracyTestHarness):
     MODEL_NAME = "nvidia/Nemotron-Mini-4B-Instruct"
     MODEL_PATH = f"{llm_models_root()}/nemotron/nemotron-mini-4b-instruct_vfp8-fp8-bf16-export"
diff --git a/tests/integration/test_lists/qa/llm_function_sanity.txt b/tests/integration/test_lists/qa/llm_function_sanity.txt
index 717bf2579df..16606a07958 100644
--- a/tests/integration/test_lists/qa/llm_function_sanity.txt
+++ b/tests/integration/test_lists/qa/llm_function_sanity.txt
@@ -47,6 +47,7 @@ accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4_4gpus_online_ep
 accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4_4gpus_online_eplb[fp8kv=True]
 accuracy/test_llm_api_pytorch.py::TestGemma3_1BInstruct::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestKanana_Instruct::test_auto_dtype
+accuracy/test_llm_api_pytorch.py::TestKimiK2::test_fp8_blockscale[latency]
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8B::test_nvfp4
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_chunked_prefill[attn_backend=FLASHINFER]
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_eagle3[eagle3_one_model=False-overlap_scheduler=False]