Enable SGLANG_ENABLE_SPEC_V2 for nightly speculative decoding tests (#18719)

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

python/sglang/test/accuracy_test_runner.py

@@ -97,6 +97,7 @@ def _run_simple_eval(
             base_url,
             other_args=model.extra_args,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            env=model.env,
         )
 
         args = SimpleNamespace(
@@ -165,6 +166,7 @@ def _run_few_shot_eval(
             base_url,
             other_args=model.extra_args,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            env=model.env,
         )
 
         args = SimpleNamespace(

python/sglang/test/nightly_utils.py

@@ -229,6 +229,7 @@ class NightlyBenchmarkRunner:
         extra_bench_args: Optional[List[str]] = None,
         enable_profile: bool = True,
         timeout: Optional[int] = None,
+        env: Optional[dict] = None,
     ) -> Tuple[List[BenchmarkResult], bool, Optional[float]]:
         """Run a complete benchmark for a single model with server management.
 
@@ -249,6 +250,7 @@ class NightlyBenchmarkRunner:
             extra_bench_args: Extra arguments for the benchmark command
             enable_profile: Whether to enable profiling (default True for NVIDIA)
             timeout: Optional timeout for server launch (defaults to DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH)
+            env: Environment dict for subprocess
 
         Returns:
             Tuple of (list of BenchmarkResult objects, success_bool, avg_spec_accept_length or None)
@@ -265,6 +267,7 @@ class NightlyBenchmarkRunner:
             timeout=(
                 timeout if timeout is not None else DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
             ),
+            env=env,
         )
 
         try:

python/sglang/test/performance_test_runner.py

@@ -79,6 +79,7 @@ def run_performance_test(
             other_args=model.extra_args,
             variant=model.variant or "",
             extra_bench_args=extra_bench_args,
+            env=model.env,
         )
 
         if success and results:

python/sglang/test/test_utils.py

@@ -855,7 +855,9 @@ def popen_launch_server(
     if env is None:
         env = os.environ.copy()
     else:
-        env = env.copy()
+        merged = os.environ.copy()
+        merged.update(env)
+        env = merged
 
     # Store per-run marker path for potential invalidation
     per_run_marker_path = None

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",
             ),
         ]