Fix fdconfig bugs (#3528)

* fix config

* fix parallel

* fix ips

* fix rl

* open code

.github/workflows/pr_build_and_test.yml

@@ -84,12 +84,12 @@ jobs:
       FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
       MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
 
-  # stable_test:
-  #   name: Run Stable Tests
-  #   needs: [clone,build]
-  #   uses: ./.github/workflows/_stable_test.yml
-  #   with:
-  #     DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-  #     FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-  #     FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
-  #     MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
+  stable_test:
+    name: Run Stable Tests
+    needs: [clone,build]
+    uses: ./.github/workflows/_stable_test.yml
+    with:
+      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
+      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
+      FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
+      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"

fastdeploy/engine/engine.py

@@ -1108,6 +1108,7 @@ class LLMEngine:
             f" --load_strategy {self.cfg.load_config.load_strategy}"
             f" --early_stop_config '{self.cfg.early_stop_config.to_json_string()}'"
             f" --load_choices {self.cfg.load_config.load_choices}"
+            f" --ips {self.cfg.ips}"
         )
 
         worker_append_flag = {
@@ -1292,7 +1293,7 @@ class LLMEngine:
                     )
                 )
                 llm_logger.info(
-                    f"Engine is initialized successfully with {self.cfg.tensor_parallel_size}"
+                    f"Engine is initialized successfully with {self.cfg.parallel_config.tensor_parallel_size}"
                     + f" data parallel id {i}"
                 )
                 self.dp_processed[-1].start()

fastdeploy/engine/sched/resource_manager_v1.py

@@ -132,7 +132,7 @@ class ResourceManagerV1(ResourceManager):
         num_new_tokens = request.need_prefill_tokens - request.num_computed_tokens
         num_new_tokens = min(num_new_tokens, token_budget)
 
-        if not self.config.enable_mm:
+        if not self.config.model_config.enable_mm:
             return num_new_tokens
 
         inputs = request.multimodal_inputs
@@ -290,7 +290,7 @@ class ResourceManagerV1(ResourceManager):
                 while self.waiting and token_budget > 0:
                     if len(self.running) == self.max_num_seqs:
                         break
-                    if self.config.enable_mm and self.exist_prefill(scheduled_reqs):
+                    if self.config.model_config.enable_mm and self.exist_prefill(scheduled_reqs):
                         break
                     request = self.waiting[0]
                     if request.status == RequestStatus.WAITING:

fastdeploy/rl/rollout_config.py

@@ -102,6 +102,7 @@ class RolloutModelConfig:
         self.graph_optimization_config = graph_optimization_config
         self.local_rank = local_rank
         self.early_stop_config = early_stop_config
+        self.ips = None
 
     def __str__(self):
         return "\n".join(f"{k}: {v}" for k, v in self.__dict__.items())

fastdeploy/worker/worker_process.py

@@ -606,6 +606,13 @@ def parse_args():
         help="The format of the model weights to load. default/new_loader.",
     )
 
+    parser.add_argument(
+        "--ips",
+        type=str,
+        default=None,
+        help="The ips of multinode deployment.",
+    )
+
     args = parser.parse_args()
     return args
 
@@ -721,6 +728,7 @@ def initialize_fd_config(args, ranks: int = 1, local_rank: int = 0) -> FDConfig:
         graph_opt_config=graph_opt_config,
         early_stop_config=early_stop_config,
         cache_config=cache_config,
+        ips=args.ips,
     )
     update_fd_config_for_mm(fd_config)