[CP]Glm45 air 2.2 (#4073)

* [Feature] Support zai-org/GLM-4.5-Air BF16 model (#3928)

* support glm45_air

* [Feature] GLM-45-AIR Support Mix Quantization(Dense wfp8afp8 and wint8 triton_moe_backend) (#4051)

* check

* fix v1 load for mix and wint8

* check --quantizations 'None'

* check

* support RL rollout

* check v1 loader

* check glm rollout_model, change wfp8afp8 per_token_cast_to_fp8 to native impl

* check rollout moe gate begin layer_id

* check rollout e_score_correction_bias

* delete infer_to_train_mapping={}

* code check

fastdeploy/worker/worker_process.py

@@ -44,7 +44,7 @@ from fastdeploy.inter_communicator import EngineWorkerQueue as TaskQueue
 from fastdeploy.inter_communicator import IPCSignal
 from fastdeploy.model_executor.layers.quantization import get_quantization_config
 from fastdeploy.platforms import current_platform
-from fastdeploy.utils import get_logger
+from fastdeploy.utils import get_logger, parse_quantization
 from fastdeploy.worker.worker_base import WorkerBase
 
 logger = get_logger("worker_process", "worker_process.log")
@@ -545,9 +545,9 @@ def parse_args():
 
     parser.add_argument(
         "--quantization",
-        type=str,
-        default="None",
-        help="Quantization name for the model, currentlly support "
+        type=json.loads,
+        default=None,
+        help="Quantization name for the model, currently support "
         "'wint4', 'wint8',"
         "default is None. The priority of this configuration "
         "is lower than that of the config file. "
@@ -635,6 +635,9 @@ def initialize_fd_config(args, ranks: int = 1, local_rank: int = 0) -> FDConfig:
     Returns:
         FDConfig: Initialized FastDeploy configuration object
     """
+    # RL rollout
+    if args.quantization is not None and isinstance(args.quantization, str):
+        args.quantization = parse_quantization(args.quantization)
     paddle.set_default_dtype(args.dtype)
     model_config = ModelConfig(vars(args))
     device_config = DeviceConfig(vars(args))
@@ -704,12 +707,16 @@ def initialize_fd_config(args, ranks: int = 1, local_rank: int = 0) -> FDConfig:
 
     if quantization_config is not None:
         quant_config_name = quantization_config["quantization"]
-    elif args.quantization != "None":
+    elif args.quantization is not None:
         quantization_config = {}
-        quant_config_name = args.quantization
-        quantization_config["quantization"] = quant_config_name
+        try:
+            quantization_config.update(args.quantization)
+            quant_config_name = quantization_config["quantization"]
+        except:
+            quant_config_name = args.quantization["quantization"]
+            quantization_config["quantization"] = quant_config_name
         # Only v1 loader sets is_checkpoint_bf16=True during dynamic quantization.
-        if load_config.load_choices == "default_v1":
+        if load_config.load_choices == "default_v1" and not load_config.dynamic_load_weight:
             quantization_config["is_checkpoint_bf16"] = True
         # Special handling for Ernie models
         is_ernie = ErnieArchitectures.contains_ernie_arch(model_config.architectures)