[Bugfix] Correct prompt lengths for timed_traces benchmark (#45423)

Signed-off-by: Stanislaw Wozniak <stw@zurich.ibm.com>
Co-authored-by: Thomas Parnell <tpa@zurich.ibm.com>
This commit is contained in:
Stan Wozniak
2026-08-13 21:51:24 +00:00
committed by GitHub
co-authored by Thomas Parnell
parent b652dedd0c
commit 63550515b4
3 changed files with 8 additions and 4 deletions
+2 -3
View File
@@ -79,7 +79,7 @@ class SampleRequest:
Represents a single inference request for benchmarking.
"""
prompt: str | list[str] | list[dict]
prompt: str | list[str] | list[int] | list[dict]
prompt_len: int
expected_output_len: int = 0
multi_modal_data: MultiModalDataDict | dict | list[dict] | None = None
@@ -1572,7 +1572,6 @@ class TimedTrace(BenchmarkDataset):
prompt_ids = self._expand_prompt(
entry.get(self.label_hash_ids, []), input_length, tokenizer
)
prompt = tokenizer.decode(prompt_ids)
# Get timestamp with proper error handling
ts_value = entry.get(self.label_ts)
@@ -1588,7 +1587,7 @@ class TimedTrace(BenchmarkDataset):
samples.append(
SampleRequest(
prompt=prompt,
prompt=prompt_ids,
prompt_len=prompt_len,
expected_output_len=new_output_len,
lora_request=None,
+1 -1
View File
@@ -66,7 +66,7 @@ class StreamedResponseHandler:
class RequestFuncInput:
"""The input for the request function."""
prompt: str | list[str] | list[dict[str, Any]]
prompt: str | list[str] | list[int] | list[dict[str, Any]]
api_url: str
prompt_len: int
output_len: int
+5
View File
@@ -2097,6 +2097,11 @@ async def main_async(args: argparse.Namespace) -> dict[str, Any]:
args.ignore_eos = True
if args.dataset_name == "timed_trace":
if args.backend not in ("vllm", "openai"):
raise ValueError(
"timed_trace dataset passes pre-tokenized prompts (list[int])"
" and requires a completions backend ('vllm' or 'openai')."
)
# timed_trace carries per-request timestamps;
# ignore EOS so generation runs to the trace's specified output length,
# and default to using those timestamps for scheduling unless the user