From a77c9d42987dc5bc5970407e648296df0254cb49 Mon Sep 17 00:00:00 2001
From: Sunny-bot1 <592045536@qq.com>
Date: Wed, 1 Apr 2026 17:21:25 +0800
Subject: [PATCH 1/2] rl gate fp32

---
 fastdeploy/engine/args_utils.py | 2 ++
 fastdeploy/envs.py              | 2 ++
 2 files changed, 4 insertions(+)

diff --git a/fastdeploy/engine/args_utils.py b/fastdeploy/engine/args_utils.py
index ff0965c56bb..26157d7d23d 100644
--- a/fastdeploy/engine/args_utils.py
+++ b/fastdeploy/engine/args_utils.py
@@ -624,6 +624,8 @@ def __post_init__(self):
             raise NotImplementedError(
                 f"not support model_impl: '{self.model_impl}'. " f"Must be one of: {', '.join(valid_model_impls)}"
             )
+        if envs.FD_ENABLE_RL == 1:
+            self.moe_gate_fp32 = True
 
         self.post_init_all_ports()
 
diff --git a/fastdeploy/envs.py b/fastdeploy/envs.py
index 0c7ac3e22b1..5534871fcb3 100644
--- a/fastdeploy/envs.py
+++ b/fastdeploy/envs.py
@@ -266,6 +266,8 @@ def _validate_split_kv_size(value: int) -> int:
     "FD_SAVE_OUTPUT_CACHE_FOR_PREEMPTED_REQUEST": lambda: bool(
         int(os.getenv("FD_SAVE_OUTPUT_CACHE_FOR_PREEMPTED_REQUEST", "1"))
     ),
+    # Whether to align RoPE and moe gate precision with training
+    "FD_ENABLE_RL": lambda: int(os.getenv("FD_ENABLE_RL", "0")),
 }
 
 

From c3891cc85e30c9d09b412e9522439ca6a69abee5 Mon Sep 17 00:00:00 2001
From: Sunny-bot1 <592045536@qq.com>
Date: Wed, 1 Apr 2026 19:10:16 +0800
Subject: [PATCH 2/2] clean

---
 fastdeploy/model_executor/models/glm4_moe.py | 4 +---
 fastdeploy/model_executor/models/qwen3moe.py | 4 +---
 2 files changed, 2 insertions(+), 6 deletions(-)

diff --git a/fastdeploy/model_executor/models/glm4_moe.py b/fastdeploy/model_executor/models/glm4_moe.py
index 3f45e9df614..64b82229d37 100644
--- a/fastdeploy/model_executor/models/glm4_moe.py
+++ b/fastdeploy/model_executor/models/glm4_moe.py
@@ -150,9 +150,7 @@ def __init__(
             output_size=fd_config.model_config.n_routed_experts,
             with_bias=False,
             skip_quant=True,
-            weight_dtype=(
-                "float32" if fd_config.load_config.dynamic_load_weight or fd_config.model_config.moe_gate_fp32 else ""
-            ),
+            weight_dtype=("float32" if fd_config.model_config.moe_gate_fp32 else ""),
         )
         self.gate.e_score_correction_bias = self.create_parameter(
             shape=[1, fd_config.model_config.n_routed_experts],
diff --git a/fastdeploy/model_executor/models/qwen3moe.py b/fastdeploy/model_executor/models/qwen3moe.py
index 74ca37ab695..6c443d68bcc 100644
--- a/fastdeploy/model_executor/models/qwen3moe.py
+++ b/fastdeploy/model_executor/models/qwen3moe.py
@@ -77,9 +77,7 @@ def __init__(
             output_size=fd_config.model_config.num_experts,
             with_bias=False,
             skip_quant=True,
-            weight_dtype=(
-                "float32" if fd_config.load_config.dynamic_load_weight or fd_config.model_config.moe_gate_fp32 else ""
-            ),
+            weight_dtype=("float32" if fd_config.model_config.moe_gate_fp32 else ""),
         )
 
     def forward(self, x, forward_meta):