Enable SGLANG_ENABLE_SPEC_V2 for nightly speculative decoding tests (#18719)

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

test/registered/8-gpu-models/test_deepseek_v31.py

@@ -54,6 +54,7 @@ class TestDeepseekV31(unittest.TestCase):
                 DEEPSEEK_V31_MODEL_PATH,
                 tp_size=8,
                 extra_args=base_args + mtp_args,
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
                 variant="TP8+MTP",
             ),
         ]

test/registered/8-gpu-models/test_deepseek_v32.py

@@ -68,6 +68,7 @@ class TestDeepseekV32(unittest.TestCase):
                 DEEPSEEK_V32_MODEL_PATH,
                 tp_size=8,
                 extra_args=BASE_ARGS + DP_ARGS + TOOL_CALL_ARGS + MTP_ARGS,
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
                 variant="DP8+MTP",
             ),
             # Variant: "tp" - Pure TP=8 only
@@ -82,6 +83,7 @@ class TestDeepseekV32(unittest.TestCase):
                 DEEPSEEK_V32_MODEL_PATH,
                 tp_size=8,
                 extra_args=BASE_ARGS + TP_ARGS + TOOL_CALL_ARGS + MTP_ARGS,
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
                 variant="TP8+MTP",
             ),
         ]

test/registered/8-gpu-models/test_deepseek_v32_cp_single_node.py

@@ -65,6 +65,7 @@ class TestDeepseekV32CPSingleNode(unittest.TestCase):
                 DEEPSEEK_V32_EXP_MODEL_PATH,
                 tp_size=8,
                 extra_args=BASE_ARGS + DP_ARGS + MTP_ARGS + CP_IN_SEQ_SPLIT_ARGS,
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
                 variant="CP-in-seq-split",
             ),
             # Variant: round-robin-split CP mode (TP only, no DP)
@@ -72,6 +73,7 @@ class TestDeepseekV32CPSingleNode(unittest.TestCase):
                 DEEPSEEK_V32_EXP_MODEL_PATH,
                 tp_size=8,
                 extra_args=BASE_ARGS + MTP_ARGS + CP_ROUND_ROBIN_ARGS,
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
                 variant="CP-round-robin-split",
             ),
         ]

test/registered/8-gpu-models/test_glm_46_fp8.py

@@ -45,6 +45,7 @@ class TestGLM46FP8(unittest.TestCase):
                 GLM_4_6_FP8_MODEL_PATH,
                 tp_size=8,
                 extra_args=base_args + mtp_args,
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
                 variant="TP8+MTP",
             ),
         ]

test/registered/8-gpu-models/test_mistral_large3.py

@@ -76,6 +76,7 @@ class TestMistralLarge3(unittest.TestCase):
                 MISTRAL_LARGE3_FP8_MODEL_PATH,
                 tp_size=8,
                 extra_args=base_args + eagle_args,
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
                 variant="TP8+MTP",
             ),
             # Variant: "nvfp4" - NVFP4 model + TP=8 + trtllm_mla backend

test/registered/8-gpu-models/test_qwen3_235b.py

@@ -55,6 +55,7 @@ class TestQwen3235BFP8(unittest.TestCase):
                 QWEN3_235B_FP8_MODEL_PATH,
                 tp_size=8,
                 extra_args=base_args + eagle3_args,
+                env={"SGLANG_ENABLE_SPEC_V2": "1"},
                 variant="TP8+EP2+EAGLE3",
             ),
         ]