[Refactor] Refactor DeepEP dispatcher (#22822)

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>

test/manual/layers/moe/test_moe_runners_4gpu.py

@@ -55,7 +55,6 @@ class TestMoERunner4GPU(CustomTestCase):
         },
         "moe_runner_cutlass_w4a8_deepep_ll": {
             "model": "tencent/DeepSeek-V3.1-Terminus-W4AFP8",  # FP8 W8A8 MoE model
-            "env_overrides": {"SGLANG_DEEPEP_BF16_DISPATCH": "1"},
             "other_args": [
                 "--trust-remote-code",
                 "--moe-runner-backend",
@@ -64,6 +63,8 @@ class TestMoERunner4GPU(CustomTestCase):
                 "deepep",
                 "--deepep-mode",
                 "low_latency",
+                "--deepep-dispatcher-output-dtype",
+                "bf16",
                 "--attention-backend",
                 "triton",
                 "--sampling-backend",

test/manual/test_w4a8_deepseek_v3.py

@@ -192,6 +192,8 @@ class TestDeepseekV3W4Afp8DeepepAutoMtp(CustomTestCase):
             "deepep",
             "--deepep-mode",
             "auto",
+            "--deepep-dispatcher-output-dtype",
+            "bf16",
             "--dp",
             "8",
             "--enable-dp-attention",
@@ -215,7 +217,6 @@ class TestDeepseekV3W4Afp8DeepepAutoMtp(CustomTestCase):
             other_args=other_args,
             env={
                 **os.environ,
-                "SGLANG_DEEPEP_BF16_DISPATCH": "1",
                 "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "256",
             },
         )