diff --git a/tensorrt_llm/_torch/pyexecutor/py_executor.py b/tensorrt_llm/_torch/pyexecutor/py_executor.py
index d04f9a25352..eb8619308b3 100644
--- a/tensorrt_llm/_torch/pyexecutor/py_executor.py
+++ b/tensorrt_llm/_torch/pyexecutor/py_executor.py
@@ -870,7 +870,7 @@ def _executor_loop(self):
                     self.resource_manager.prepare_resources(scheduled_batch)
                     if self.drafter is not None:
                         self.drafter.prepare_draft_tokens(
-                            scheduled_batch, self.resource_manager)
+                            scheduled_batch, self.resource_manager, iter_stats)
 
                     batch_outputs = self._forward_step(scheduled_batch)
 
diff --git a/tensorrt_llm/_torch/speculative/drafter.py b/tensorrt_llm/_torch/speculative/drafter.py
index e08044cbb4f..abaa05985f1 100644
--- a/tensorrt_llm/_torch/speculative/drafter.py
+++ b/tensorrt_llm/_torch/speculative/drafter.py
@@ -1,6 +1,8 @@
 from abc import ABC, abstractmethod
 from typing import Optional
 
+from tensorrt_llm.bindings.executor import IterationStats
+
 from ..pyexecutor.resource_manager import ResourceManager
 from ..pyexecutor.scheduler import ScheduledRequests
 
@@ -13,6 +15,7 @@ def prepare_draft_tokens(
         self,
         scheduled_requests: ScheduledRequests,
         resource_manager: Optional[ResourceManager] = None,
+        iter_stats: IterationStats = None,
     ) -> None:
         """
         Prepare the drafter tokens for the forward computation this step.
diff --git a/tensorrt_llm/_torch/speculative/model_drafter.py b/tensorrt_llm/_torch/speculative/model_drafter.py
index 53d7af3d360..38707d0e47e 100644
--- a/tensorrt_llm/_torch/speculative/model_drafter.py
+++ b/tensorrt_llm/_torch/speculative/model_drafter.py
@@ -6,6 +6,7 @@
 import torch
 
 from tensorrt_llm._utils import nvtx_range
+from tensorrt_llm.bindings.executor import IterationStats
 from tensorrt_llm.logger import logger
 
 from ..pyexecutor.llm_request import LlmRequest, LlmRequestState, SamplingConfig
@@ -297,6 +298,7 @@ def prepare_draft_tokens(
         self,
         scheduled_requests: ScheduledRequests,
         resource_manager: Optional[ResourceManager] = None,
+        iter_stats: IterationStats = None,
     ) -> None:
         """
         Prepare draft tokens for the scheduled requests.
diff --git a/tensorrt_llm/_torch/speculative/ngram.py b/tensorrt_llm/_torch/speculative/ngram.py
index 9113900ef94..473e257bb1d 100644
--- a/tensorrt_llm/_torch/speculative/ngram.py
+++ b/tensorrt_llm/_torch/speculative/ngram.py
@@ -1,7 +1,10 @@
+import time
 from itertools import chain
+from typing import Dict, Tuple
 
 from ordered_set import OrderedSet
 
+from tensorrt_llm.bindings.executor import IterationStats
 from tensorrt_llm.logger import logger
 
 from ..pyexecutor.llm_request import *
@@ -174,6 +177,21 @@ def prepare_draft_tokens(
         self,
         scheduled_requests: ScheduledRequests,
         resource_manager: Optional[ResourceManager] = None,
+        iter_stats: IterationStats = None,
+    ) -> None:
+
+        if iter_stats is not None:
+            start_time = time.time()
+
+        self._prepare_draft_tokens(scheduled_requests)
+
+        if iter_stats is not None:
+            self._update_ngram_iter_stats(scheduled_requests, iter_stats,
+                                          start_time)
+
+    def _prepare_draft_tokens(
+        self,
+        scheduled_requests: ScheduledRequests,
     ) -> None:
         # Sort by request_id when py_batch_idx is None as a fallback.
         # This happens in the disagg case: for a set of new requests, we draft
@@ -195,6 +213,48 @@ def prepare_draft_tokens(
             )
             # Pad length to `self.max_draft_len`
             if len(draft_tokens) > 0:
-                pad_length = self.max_draft_len - len(draft_tokens)
+                draft_length = len(draft_tokens)
+                pad_length = self.max_draft_len - draft_length
                 draft_tokens.extend([request.py_end_id] * pad_length)
+            else:
+                draft_length = 0
             request.py_draft_tokens = draft_tokens
+            request.py_draft_length = draft_length
+
+    def _update_ngram_iter_stats(
+        self,
+        scheduled_requests: ScheduledRequests,
+        iter_stats: IterationStats,
+        start_time: float,
+    ) -> Tuple[ScheduledRequests, Dict[int, LlmRequest]]:
+        """
+        Get statistic information from the draft tokens in NGram drafter
+        """
+        now_time = time.time()
+
+        total_num_draft_tokens = 0
+        total_num_accepted_tokens = 0
+        num_requests_with_draft_tokens = 0
+        for request in scheduled_requests.generation_requests:
+            if request.py_last_draft_tokens is not None:
+                total_num_draft_tokens += request.py_draft_length
+                total_num_accepted_tokens += request.py_num_accepted_draft_tokens
+                num_requests_with_draft_tokens += 1
+
+        if num_requests_with_draft_tokens > 0:
+            iter_stats.specdec_stats.iter_latency_ms = (now_time -
+                                                        start_time) * 1e3
+            iter_stats.specdec_stats.num_draft_tokens = total_num_draft_tokens
+            iter_stats.specdec_stats.num_accepted_tokens = total_num_accepted_tokens
+            iter_stats.specdec_stats.num_requests_with_draft_tokens = num_requests_with_draft_tokens
+            iter_stats.specdec_stats.acceptance_length = (
+                total_num_accepted_tokens +
+                num_requests_with_draft_tokens) / num_requests_with_draft_tokens
+        else:
+            iter_stats.specdec_stats.iter_latency_ms = 0.0
+            iter_stats.specdec_stats.num_draft_tokens = 0
+            iter_stats.specdec_stats.num_accepted_tokens = 0
+            iter_stats.specdec_stats.num_requests_with_draft_tokens = 0
+            iter_stats.specdec_stats.acceptance_length = 1.0
+
+        return