Enable SGLANG_ENABLE_SPEC_V2 for nightly speculative decoding tests (#18719)

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Kangyan-Zhou
2026-02-14 23:00:33 +08:00
committed by GitHub
parent 92cdd398cd
commit ae95869292
10 changed files with 17 additions and 1 deletions

View File

@@ -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(

View File

@@ -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:

View File

@@ -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:

View File

@@ -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