Clean deprecated DeepSeek V4 Environs (#34926)

test/registered/amd/test_deepseek_v4_flash_fp8_tbo.py

@@ -53,7 +53,6 @@ COMMON_ENV_VARS = {
     "SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
     "SGLANG_USE_AITER": "1",
     "SGLANG_USE_ROCM700A": "0",
-    "SGLANG_OPT_USE_FUSED_COMPRESS": "true",
     "SGLANG_HACK_FLASHMLA_BACKEND": FLASHMLA_BACKEND,
     "SGLANG_OPT_FP8_WO_A_GEMM": "false",
     "SGLANG_OPT_USE_JIT_INDEXER_METADATA": "false",
@@ -63,7 +62,6 @@ COMMON_ENV_VARS = {
     "SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
     "SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
     "SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
-    "SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
     "SGLANG_OPT_USE_MULTI_STREAM_OVERLAP": "false",
     "SGLANG_ROCM_USE_MULTI_STREAM": "false",
     "AITER_BF16_FP8_MOE_BOUND": "0",

test/registered/dcp/test_kimi_linear_dcp_dspark4.py

@@ -168,7 +168,6 @@ class TestKimiLinearDCPDSpark4(CustomTestCase):
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 8,
             other_args=other_args,
             env={
-                "SGLANG_PREP_IN_CUDA_GRAPH": "1",
                 "SGLANG_RAGGED_VERIFY_MODE": "static",
             },
         )

test/registered/models_e2e/test_deepseek_v4_flash_fp8_h200.py

@@ -140,9 +140,6 @@ class TestDSV4FlashFP8H200MegaMoE(
             ],
             env={
                 "SGLANG_DSV4_FP4_EXPERTS": "0",
-                "SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE": "1",
-                "SGLANG_OPT_FIX_MEGA_MOE_MEMORY": "1",
-                "SGLANG_OPT_USE_JIT_EP_ACTIVATION": "1",
                 # INVARIANT: this per-rank cap MUST equal
                 # chunked_prefill_size / dp_size (= 8192 / 4 = 2048), the per-rank
                 # prefill bound under --enable-dp-attention. If you change

test/registered/unit/layers/test_dsv4_nonpaged_indexer.py

@@ -157,10 +157,7 @@ class TestDSV4NonPagedIndexer(CustomTestCase):
         threshold = envs.SGLANG_OPT_DSV4_NONPAGED_INDEXER_MIN_QUERY_TOKENS
         with threshold.override(threshold.default):
             self.assertIsNone(build_plan())
-        with (
-            threshold.override(query_rows),
-            envs.SGLANG_TOPK_TRANSFORM_512_TORCH.override(False),
-        ):
+        with threshold.override(query_rows):
             plan = build_plan()
         self.assertEqual(
             (plan.seq_len_sum, plan.max_seqlen_k, plan.query_rows),

test/registered/unit/test_model_overrides.py

@@ -2123,7 +2123,6 @@ class TestGoldenModelOverrides(_IsolatedPublish):
     def test_data_parallelism_and_a2a_passes(self):
         from sglang.srt.arg_groups.overrides import (
             ResolvedView,
-            _a2a_backend_overrides,
             _a2a_ep_size,
             _data_parallelism_defaults,
         )
@@ -2141,27 +2140,6 @@ class TestGoldenModelOverrides(_IsolatedPublish):
             {},
         )
 
-        with patch("sglang.srt.environ.envs.SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE") as e:
-            e.get.return_value = False
-            self.assertEqual(
-                _a2a_backend_overrides(
-                    ResolvedView(
-                        SimpleNamespace(enable_waterfill=True, moe_a2a_backend="none")
-                    )
-                ),
-                {"moe_a2a_backend": "deepep"},
-            )
-            e.get.return_value = True
-            # megamoe env wins over the waterfill override (chained, last write)
-            self.assertEqual(
-                _a2a_backend_overrides(
-                    ResolvedView(
-                        SimpleNamespace(enable_waterfill=True, moe_a2a_backend="none")
-                    )
-                ),
-                {"moe_a2a_backend": "megamoe"},
-            )
-
         self.assertEqual(
             _a2a_ep_size(
                 ResolvedView(