Fix for random sampler recompilations for incomplete batches

vllm/worker/hpu_model_runner.py

@@ -1228,10 +1228,18 @@ def prepare_input_tensors(
         batch_size_padded = self.bucketing_ctx.get_padded_batch_size(
             real_batch_size, is_prompt)
         batch_size_padding = batch_size_padded - real_batch_size
+        if all([
+                seq_group_metadata.sampling_params.temperature
+                for seq_group_metadata in seq_group_metadata_list
+        ]):
+            temperature = 1.0
+        else:
+            temperature = 0
+
         seq_group_metadata_list = seq_group_metadata_list.copy()
         if batch_size_padding > 0:
             dummy_seq_group_metadata = self.create_dummy_seq_group_metadata(
-                0, 0, is_prompt)
+                0, 0, is_prompt, temperature=temperature)
             seq_group_metadata_list.extend(dummy_seq_group_metadata
                                            for _ in range(batch_size_padding))