[Feature] block sparse attention (#3668)

* 支持稀疏attn

* fix bug

* code style

* fix moba attn get kv shape

* 修复a100编译

* codestyle

* code style

* code style

* code style

* fix conflict

* 增加单侧

* code style

* 增加eblite 加载时间

* fix bug

* for ci

* for ci

* for ci

* for ci

* 支持mlp block size 128

* 增加小算子单测

* fix 单测 mlp

* 将环境变量加入到config里面

* fix rollout config

* 修复显存

* add test server

* add test server

* fix mlp  最后一层使用full attn

fastdeploy/worker/worker_process.py

@@ -34,6 +34,7 @@ from fastdeploy.config import (
     FDConfig,
     GraphOptimizationConfig,
     LoadConfig,
+    MobaAttentionConfig,
     ModelConfig,
     ParallelConfig,
     SpeculativeConfig,
@@ -553,6 +554,12 @@ def parse_args():
         default=None,
         help="Configation of Graph optimization backend.",
     )
+    parser.add_argument(
+        "--moba_attention_config",
+        type=json.loads,
+        default=None,
+        help="Configation of moba attention.",
+    )
     parser.add_argument(
         "--guided_decoding_backend",
         type=str,
@@ -658,6 +665,8 @@ def initialize_fd_config(args, ranks: int = 1, local_rank: int = 0) -> FDConfig:
 
     graph_opt_config = GraphOptimizationConfig(args.graph_optimization_config)
 
+    moba_attention_config = MobaAttentionConfig(args.moba_attention_config)
+
     early_stop_config = EarlyStopConfig(args.early_stop_config)
 
     # Note(tangbinhan): used for load_checkpoint
@@ -739,6 +748,7 @@ def initialize_fd_config(args, ranks: int = 1, local_rank: int = 0) -> FDConfig:
         cache_config=cache_config,
         engine_worker_queue_port=args.engine_worker_queue_port,
         ips=args.ips,
+        moba_attention_config=moba_attention_config,
     )
     update_fd_config_for_mm(fd_config)