Increase default live decode budget

demo_controller.py

@@ -11,7 +11,7 @@ from engines.compare import build_summary_sentence, compute_agreement, compute_m
 from engines.dense_runner import run_dense_live
 from engines.dotcache_runner import run_dotcache_live
 from engines.fixture_builder import FIXTURE_VERSION, build_fixture_result
-from engines.live_request import resolve_live_runtime_settings
+from engines.live_request import DEFAULT_LIVE_DECODE_STEPS, resolve_live_runtime_settings
 from engines.presets import (
     BIT_WIDTHS,
     CONTEXT_LENGTHS,
@@ -250,7 +250,7 @@ def _execute_live_request(request: DemoRequest, cache: ResultCache) -> Controlle
 def _validate_live_request(request_payload: dict[str, Any]) -> None:
     resolve_live_runtime_settings(
         request_payload,
-        decode_steps=int(os.getenv("DOTCACHE_SPACE_LIVE_DECODE_STEPS", "24")),
+        decode_steps=int(os.getenv("DOTCACHE_SPACE_LIVE_DECODE_STEPS", str(DEFAULT_LIVE_DECODE_STEPS))),
         max_live_context=int(os.getenv("DOTCACHE_SPACE_MAX_LIVE_CONTEXT", "4096")),
     )
 

engines/live_request.py

@@ -14,6 +14,8 @@ SHORTLIST_TOP_K = {
     "aggressive": 2,
 }
 
+DEFAULT_LIVE_DECODE_STEPS = 48
+
 SELECTIVE_EXACT_K_LAYERS = (19, 23)
 
 

scripts/space_dense_runner.py

@@ -10,7 +10,7 @@ if str(REPO_ROOT) not in sys.path:
     sys.path.insert(0, str(REPO_ROOT))
 
 from dotcache.integrations.qwen35 import Qwen35TextHarness  # noqa: E402
-from engines.live_request import resolve_live_runtime_settings  # noqa: E402
+from engines.live_request import DEFAULT_LIVE_DECODE_STEPS, resolve_live_runtime_settings  # noqa: E402
 from scripts.space_runner_common import (  # noqa: E402
     configure_model_cache_env,
     decode_generated_text,
@@ -51,7 +51,7 @@ def main() -> int:
     request = load_request_from_stdin()
     settings = resolve_live_runtime_settings(
         request,
-        decode_steps=int(os.getenv("DOTCACHE_SPACE_LIVE_DECODE_STEPS", "24")),
+        decode_steps=int(os.getenv("DOTCACHE_SPACE_LIVE_DECODE_STEPS", str(DEFAULT_LIVE_DECODE_STEPS))),
         max_live_context=int(os.getenv("DOTCACHE_SPACE_MAX_LIVE_CONTEXT", "4096")),
     )
 

scripts/space_dotcache_runner.py

@@ -11,7 +11,11 @@ if str(REPO_ROOT) not in sys.path:
 
 from benchmarks.bench_qwen35_attention_subset_dotcache_serving import _build_dotcache_config  # noqa: E402
 from dotcache.integrations.qwen35 import Qwen35AttentionSubsetDotCacheHarness  # noqa: E402
-from engines.live_request import resolve_live_runtime_settings, selective_exact_k_overrides  # noqa: E402
+from engines.live_request import (  # noqa: E402
+    DEFAULT_LIVE_DECODE_STEPS,
+    resolve_live_runtime_settings,
+    selective_exact_k_overrides,
+)
 from scripts.space_runner_common import (  # noqa: E402
     configure_model_cache_env,
     decode_generated_text,
@@ -149,7 +153,7 @@ def main() -> int:
     request = load_request_from_stdin()
     settings = resolve_live_runtime_settings(
         request,
-        decode_steps=int(os.getenv("DOTCACHE_SPACE_LIVE_DECODE_STEPS", "24")),
+        decode_steps=int(os.getenv("DOTCACHE_SPACE_LIVE_DECODE_STEPS", str(DEFAULT_LIVE_DECODE_STEPS))),
         max_live_context=int(os.getenv("DOTCACHE_SPACE_MAX_LIVE_CONTEXT", "4096")),
     )