Revert sampler.py and mtp.py to upstream version

These changes (XPU guard removal, num_cpu_blocks condition) were
debugging artifacts unrelated to MiniMax-M2.5 SM80 FP8 MoE support.

Author: ZhijunLStudio

fastdeploy/model_executor/layers/sample/sampler.py

@@ -77,7 +77,10 @@ def padding_sampling_params(top_p, top_k, infer_seed, seq_lens_this_time, seq_le
     top_k_padding = paddle.repeat_interleave(top_k[:real_bsz], repeats).unsqueeze(1)
     topp_seed = paddle.repeat_interleave(infer_seed[:real_bsz], repeats).unsqueeze(1)
 
-    MAX_INFER_SEED = 9223372036854775806
+    if current_platform.is_xpu():
+        MAX_INFER_SEED = 2147483646
+    else:
+        MAX_INFER_SEED = 9223372036854775806
 
     token_lens = paddle.where(
         seq_lens_encoder[:real_bsz] == 0,
@@ -97,7 +100,7 @@ def padding_sampling_params(top_p, top_k, infer_seed, seq_lens_this_time, seq_le
 
     offsets = paddle.where(
         is_decoder,
-        local_pos * 4,
+        local_pos * (32 if current_platform.is_xpu() else 4),
         paddle.zeros_like(local_pos),
     )
 

fastdeploy/spec_decode/mtp.py

@@ -428,8 +428,7 @@ def clear_mtp_cache(self, profile=False):
         Clear allocated cacheKV
         """
         create_cache_tensor = profile or not (
-            self.fd_config.cache_config.num_cpu_blocks > 0
-            or self.fd_config.cache_config.kvcache_storage_backend
+            self.fd_config.cache_config.kvcache_storage_backend
             or self.fd_config.scheduler_config.splitwise_role != "mixed"
         )
         if not create_cache_tensor: