[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/model_executor/layers/utils.py

@@ -77,6 +77,17 @@ def per_block_cast_to_fp8(x: Tensor, block_size: list = [128, 128]) -> Tuple[Ten
     )
 
 
+def per_token_cast_to_fp8(x: Tensor) -> Tuple[Tensor, Tensor]:
+    """
+    Per token cast to float8_e4m3fn used in wfp8apf8
+    """
+    x_abs = paddle.abs(x).astype(paddle.float32)
+    x_max = x_abs.max(axis=-1, keepdim=True).clip_(min=1e-4)
+    x_s = x_max / 448.0
+    x_q = paddle.clip(x / x_s, -448.0, 448.0).astype(paddle.float8_e4m3fn)
+    return x_q, x_s
+
+
 # for distributed tensor model parallel
 def _set_var_distributed(var: Tensor, split_axis: int):
     """