[ROCm][CI] fix test_common.py

tests/conftest.py

@@ -642,10 +642,12 @@ def generate(
 
         outputs: list[tuple[list[list[int]], list[str]]] = []
         for inputs in all_inputs:
+            generate_kwargs = dict(kwargs)
+            generate_kwargs.setdefault("tokenizer", self.tokenizer)
             output_ids: torch.Tensor = self.model.generate(
                 **self.wrap_device(inputs),
                 use_cache=True,
-                **kwargs,
+                **generate_kwargs,
             )
             if self.processor is None:
                 raise RuntimeError(
@@ -724,14 +726,16 @@ def generate_greedy_logprobs(
 
         all_logprobs: list[list[torch.Tensor]] = []
         for inputs in all_inputs:
+            generate_kwargs = dict(kwargs)
+            generate_kwargs.setdefault("tokenizer", self.tokenizer)
             output: "GenerateOutput" = self.model.generate(
                 **self.wrap_device(inputs),
                 use_cache=True,
                 do_sample=False,
                 max_new_tokens=max_tokens,
                 output_hidden_states=True,
                 return_dict_in_generate=True,
-                **kwargs,
+                **generate_kwargs,
             )
             seq_logprobs = self._hidden_states_to_seq_logprobs(output.hidden_states)
             all_logprobs.append(seq_logprobs)
@@ -812,14 +816,16 @@ def generate_greedy_logprobs_limit(
         all_output_strs: list[str] = []
 
         for inputs in all_inputs:
+            generate_kwargs = dict(kwargs)
+            generate_kwargs.setdefault("tokenizer", self.tokenizer)
             output: "GenerateOutput" = self.model.generate(
                 **self.wrap_device(inputs),
                 use_cache=use_cache,
                 do_sample=False,
                 max_new_tokens=max_tokens,
                 output_hidden_states=True,
                 return_dict_in_generate=True,
-                **kwargs,
+                **generate_kwargs,
             )
 
             # Encoder-decoder models return decoder_hidden_states instead of

vllm/model_executor/models/hyperclovax.py

@@ -459,7 +459,7 @@ def forward(  # type: ignore[override]
         input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         *,
-        intermediate_tensors: IntermediateTensors | None,
+        intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
     ) -> torch.Tensor | IntermediateTensors:
         model_output = self.model(