Adapt for iluvatar gpu (#2684)

fastdeploy/model_executor/layers/sample/ops/apply_penalty_multi_scores.py

@@ -64,6 +64,21 @@ def apply_penalty_multi_scores(
             min_dec_lens,
             eos_token_ids,
         )
+    elif current_platform.is_iluvatar():
+        from fastdeploy.model_executor.ops.iluvatar import \
+            get_token_penalty_multi_scores
+        logits = get_token_penalty_multi_scores(
+            pre_token_ids,
+            logits,
+            repetition_penalties,
+            frequency_penalties,
+            presence_penalties,
+            temperature,
+            bad_words_token_ids,
+            step_idx,
+            min_dec_lens,
+            eos_token_ids,
+        )
     else:
         raise NotImplementedError()
 

fastdeploy/model_executor/layers/sample/sampler.py

@@ -170,7 +170,8 @@ class Sampler(nn.Layer):
         """
         """
         super().__init__()
-        if current_platform.is_cuda() or current_platform.is_xpu():
+        if current_platform.is_cuda() or current_platform.is_xpu(
+        ) or current_platform.is_iluvatar():
             self.forward = self.forward_cuda
         else:
             raise NotImplementedError()