diff --git a/vllm/model_executor/layers/sampler.py b/vllm/model_executor/layers/sampler.py
index 9dcfa42e2a52..013b44060d1a 100644
--- a/vllm/model_executor/layers/sampler.py
+++ b/vllm/model_executor/layers/sampler.py
@@ -82,8 +82,9 @@ class Sampler(nn.Module):
         # We use float32 for probabilities and log probabilities.
         # Compute the probabilities.
         probs = torch.softmax(logits, dim=-1, dtype=torch.float)
-        # Compute the log probabilities (before applying top-p and top-k).
-        logprobs = torch.log(probs)
+        # Compute the log probabilities.
+        # Use log_softmax to ensure numerical stability.
+        logprobs = torch.log_softmax(logits, dim=-1, dtype=torch.float)
 
         # Sample the next tokens.
         return _sample(probs, logprobs, input_metadata)
diff --git a/vllm/sequence.py b/vllm/sequence.py
index 795397a3731b..eac3af2823a1 100644
--- a/vllm/sequence.py
+++ b/vllm/sequence.py
@@ -350,7 +350,7 @@ class SequenceOutputs:
 
     def __eq__(self, other: object) -> bool:
         if not isinstance(other, SequenceOutputs):
-            return NotImplementedError()
+            raise NotImplementedError()
         return (self.parent_seq_id == other.parent_seq_id
                 and self.output_token == other.output_token
                 and self.logprobs == other.logprobs)