Merge branch 'features/refactor/base' into features/refactor/working

Signed-off-by: Jared O'Connell <[email protected]>

Author: jaredoconnell

pyproject.toml

@@ -10,6 +10,9 @@ include = ["*"]
 [tool.setuptools.package-data]
 "guidellm.data" = ["*.gz"]
 
+[tool.pdm]
+distribution = true
+
 
 # ************************************************
 # ********** Project Metadata **********
@@ -70,6 +73,10 @@ perf = [
     "msgspec",
     "uvloop",
 ]
+recommended = [
+    "tiktoken>=0.11.0",  # For OpenAI tokenizer
+    "blobfile>=3.1.0",   # For OpenAI tokenizer
+]
 dev = [
     # build
     "build>=1.0.0",
@@ -110,6 +117,9 @@ dev = [
     "mkdocs-linkcheck~=1.0.6",
 ]
 
+[dependency-groups]
+dev = [ "guidellm[dev]" ]
+
 [project.urls]
 homepage = "https://github.com/vllm-project/guidellm"
 source = "https://github.com/vllm-project/guidellm"

src/guidellm/benchmark/aggregator.py

@@ -441,7 +441,7 @@ def __call__(
 
     def compile(
         self, state: AggregatorState, scheduler_state: SchedulerState
-    ) -> dict[Literal["scheduler_stats"], BenchmarkSchedulerStats]:
+    ) -> dict[Literal["run_stats"], BenchmarkSchedulerStats]:
         """
         Compile scheduler timing metrics into benchmark statistics.
 
@@ -473,7 +473,7 @@ def compile(
                     key="worker_resolve_time", type_="avg", default=0.0
                 ),
                 worker_resolve_end_delay_avg=state.get_metric(
-                    key="worker_resolve_end_delay", type_="avg"
+                    key="worker_resolve_end_delay", type_="avg", default=0.0
                 ),
                 finalized_delay_avg=state.get_metric(
                     key="finalized_delay", type_="avg", default=0.0

src/guidellm/benchmark/profile.py

@@ -680,6 +680,8 @@ def next_strategy(
             self.throughput_rate = (
                 prev_benchmark.metrics.requests_per_second.successful.mean
             )
+            if self.synchronous_rate <= 0 and self.throughput_rate <= 0:
+                raise RuntimeError("Invalid rates in sweep; aborting. Were there any successful requests?")
             self.measured_rates = list(
                 np.linspace(
                     self.synchronous_rate,
@@ -698,7 +700,6 @@ def next_strategy(
                 if strat.type_ == self.strategy_type
             ]
         )
-
         if self.strategy_type == "constant":
             return AsyncConstantStrategy(
                 rate=self.measured_rates[next_rate_index],