diff --git a/nemo_deploy/llm/inference/inference_base.py b/nemo_deploy/llm/inference/inference_base.py
index 129344179..6d72ef948 100644
--- a/nemo_deploy/llm/inference/inference_base.py
+++ b/nemo_deploy/llm/inference/inference_base.py
@@ -536,7 +536,11 @@ def create_mcore_engine(
         buffer_size_gb=int(buffer_size_gb),
         max_requests=max_batch_size,
         block_size_tokens=block_size_tokens,
-        materialize_only_last_token_logits=True,
+        # Coordinator mode constructs one DynamicInferenceEngine per rank independently, so a
+        # per-request runtime toggle (as _infer_fn used to do) only ever reaches the primary
+        # rank's engine. Prompt log probs require this to be False on every rank, so it's set
+        # here unconditionally at construction time instead.
+        materialize_only_last_token_logits=False,
     )
 
     coordinator_host = os.environ.get("MASTER_ADDR") or _default_coordinator_host()
diff --git a/nemo_deploy/llm/megatronllm_deployable.py b/nemo_deploy/llm/megatronllm_deployable.py
index bec928698..0fc2d003b 100755
--- a/nemo_deploy/llm/megatronllm_deployable.py
+++ b/nemo_deploy/llm/megatronllm_deployable.py
@@ -205,12 +205,6 @@ def generate_other_ranks(self):
                     stop_words=stop_words,
                 )
 
-                if log_probs:
-                    dynamic_engine = getattr(self.mcore_engine, "engine", None)
-                    if dynamic_engine is not None:
-                        dynamic_engine.materialize_only_last_token_logits = False
-                        dynamic_engine.context.config.materialize_only_last_token_logits = False
-
                 self.generate(prompts, inference_params)
             else:
                 return
@@ -400,22 +394,6 @@ def _infer_fn(
         if apply_chat_template:
             prompts = [self.apply_chat_template(prompt) for prompt in prompts]
 
-        if torch.distributed.is_initialized():
-            if torch.distributed.get_world_size() > 1:
-                torch.distributed.broadcast(torch.tensor([0], dtype=torch.long, device="cuda"), src=0)
-                broadcast_list(prompts, src=0)
-                broadcast_list(
-                    data=[
-                        temperature,
-                        top_k,
-                        top_p,
-                        num_tokens_to_generate,
-                        log_probs,
-                        stop_words,
-                    ],
-                    src=0,
-                )
-
         # cast top_k,top_p to native int, float since typecheck assert statements added in MCore0.13 error otherwise
         # skip_prompt_log_probs=False (default) includes prompt tokens in top-N logprobs when top_logprobs>0.
         inference_params = SamplingParams(
@@ -429,17 +407,6 @@ def _infer_fn(
             stop_words=stop_words,
         )
 
-        # Mcore's dynamic inference engine defaults materialize_only_last_token_logits=True for
-        # performance, but prompt log probs require all token logits to be materialized
-        # (prompt log probs are required for logprob eval benchmarks).
-        # Toggle it on both the engine and the context config (controls the
-        # model forward pass and log prob calculations).
-        dynamic_engine = getattr(self.mcore_engine, "engine", None)
-        needs_all_logits = log_probs or bool(top_logprobs)
-        if dynamic_engine is not None and needs_all_logits:
-            dynamic_engine.materialize_only_last_token_logits = False
-            dynamic_engine.context.config.materialize_only_last_token_logits = False
-
         results = self.generate(prompts, inference_params)
         # Handle DynamicInferenceRequestRecord objects by merging them into a single request
         results = [
diff --git a/nemo_deploy/llm/megatronllm_deployable_ray.py b/nemo_deploy/llm/megatronllm_deployable_ray.py
index 3d2202aa9..8e4d326c4 100644
--- a/nemo_deploy/llm/megatronllm_deployable_ray.py
+++ b/nemo_deploy/llm/megatronllm_deployable_ray.py
@@ -104,8 +104,6 @@ def __init__(
                 legacy_model_format=legacy_model_format,
                 **model_config_kwargs,
             )
-            if rank != 0:
-                self.model.generate_other_ranks()
         except Exception as e:
             LOGGER.error(f"Replica {replica_id} - Failed to initialize model for rank {rank}: {str(e)}")
             raise