[Feature] dyc8 support prefixcache (#5125)

* dyc8 support prefixcache

* fix cache_trans test case

* update code

fastdeploy/worker/gpu_model_runner.py

@@ -1450,8 +1450,10 @@ class GPUModelRunner(ModelRunnerBase):
         for i in range(self.model_config.num_hidden_layers):
             # init key cache
             key_cache_name = f"key_caches_{i}_rank{local_rank}.device{self.device_id}"
+            key_cache_scales_name = f"key_cache_scales_{i}_rank{local_rank}.device{self.device}"
             if value_cache_shape:
                 val_cache_name = f"value_caches_{i}_rank{local_rank}.device{self.device_id}"
+                value_cache_scales_name = f"value_cache_scales_{i}_rank{local_rank}.device{self.device}"
             if create_cache_tensor:
                 logger.info(f"..creating kv cache for layer {i}: key:{key_cache_shape}, value:{value_cache_shape}")
                 key_cache = paddle.full(shape=key_cache_shape, fill_value=0, dtype=cache_type)
@@ -1477,12 +1479,25 @@ class GPUModelRunner(ModelRunnerBase):
                 logger.info(f"..attaching kv cache for layer {i}: key:{key_cache_shape}, value:{value_cache_shape}")
                 key_cache = paddle.empty(shape=[], dtype=cache_type)
                 key_cache = share_external_data(key_cache, key_cache_name, key_cache_shape)
+                if kv_cache_quant_type == "block_wise_fp8":
+                    key_cache_scales = paddle.empty(shape=[], dtype=paddle.get_default_dtype())
+                    key_cache_scales = share_external_data(
+                        key_cache_scales, key_cache_scales_name, kv_cache_scale_shape
+                    )
                 if value_cache_shape:
                     val_cache = paddle.empty(shape=[], dtype=cache_type)
                     val_cache = share_external_data(val_cache, val_cache_name, value_cache_shape)
                     cache_kvs_list.extend([key_cache, val_cache])
+                    if kv_cache_quant_type == "block_wise_fp8":
+                        val_cache_scales = paddle.empty(shape=[], dtype=paddle.get_default_dtype())
+                        val_cache_scales = share_external_data(
+                            val_cache_scales, value_cache_scales_name, kv_cache_scale_shape
+                        )
+                        cache_kvs_list.extend([key_cache_scales, val_cache_scales])
                 else:
                     cache_kvs_list.extend([key_cache])
+                    if kv_cache_quant_type == "block_wise_fp8":
+                        cache_kvs_list.extend([key_cache_scales])
 
         self.share_inputs["caches"] = cache_kvs_list