fixing the eagle serving

Author: richardhuo-nv

components/backends/trtllm/engine_configs/llama4/eagle_one_model/eagle_agg.yml

@@ -30,11 +30,10 @@ speculative_config:
 
 kv_cache_config:
   free_gpu_memory_fraction: 0.5
-  enable_block_reuse: true # true when target and draft are same kv dtype
+  enable_block_reuse: false # true when target and draft are same kv dtype
 
 cuda_graph_config:
   padding_enabled: true
   max_batch_size: 8
-  dtype: fp8
 
 print_iter_log: true

components/backends/trtllm/engine_configs/llama4/eagle_one_model/eagle_prefill.yaml

@@ -33,7 +33,6 @@ speculative_config:
 kv_cache_config:
   free_gpu_memory_fraction: 0.5
   enable_block_reuse: true
-  dtype: fp8
 
 cache_transceiver_config:
   backend: default