Datasets:
Update README metadata template
Browse files- generate_prefixbench.py +38 -1
generate_prefixbench.py
CHANGED
|
@@ -554,7 +554,44 @@ def build_eviction_rows(count: int, max_tokens: int, seed: int) -> list[dict[str
|
|
| 554 |
|
| 555 |
|
| 556 |
def write_readme(output_dir: Path) -> None:
|
| 557 |
-
content = """
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 558 |
|
| 559 |
These datasets generate deterministic prompts for testing KV prefix caching behavior in LLM inference servers such as vLLM and SGLang. The prompts use controlled shared prefixes plus small unique suffixes so benchmark clients can compare cache reuse, latency, and throughput across server configurations.
|
| 560 |
|
|
|
|
| 554 |
|
| 555 |
|
| 556 |
def write_readme(output_dir: Path) -> None:
|
| 557 |
+
content = """---
|
| 558 |
+
pretty_name: PrefixBench
|
| 559 |
+
language:
|
| 560 |
+
- en
|
| 561 |
+
task_categories:
|
| 562 |
+
- text-generation
|
| 563 |
+
tags:
|
| 564 |
+
- text
|
| 565 |
+
- synthetic
|
| 566 |
+
- benchmark
|
| 567 |
+
- llm-inference
|
| 568 |
+
- prefix-caching
|
| 569 |
+
- kv-cache
|
| 570 |
+
- vllm
|
| 571 |
+
- sglang
|
| 572 |
+
- radix-attention
|
| 573 |
+
size_categories:
|
| 574 |
+
- 1K<n<10K
|
| 575 |
+
configs:
|
| 576 |
+
- config_name: shared_schema_1k
|
| 577 |
+
data_files:
|
| 578 |
+
- split: test
|
| 579 |
+
path: shared_schema_1k.jsonl
|
| 580 |
+
- config_name: same_document_multi_query_8k
|
| 581 |
+
data_files:
|
| 582 |
+
- split: test
|
| 583 |
+
path: same_document_multi_query_8k.jsonl
|
| 584 |
+
- config_name: multiturn_agent_branching
|
| 585 |
+
data_files:
|
| 586 |
+
- split: test
|
| 587 |
+
path: multiturn_agent_branching.jsonl
|
| 588 |
+
- config_name: eviction_pressure_4k
|
| 589 |
+
data_files:
|
| 590 |
+
- split: test
|
| 591 |
+
path: eviction_pressure_4k.jsonl
|
| 592 |
+
---
|
| 593 |
+
|
| 594 |
+
# PrefixBench JSONL Datasets
|
| 595 |
|
| 596 |
These datasets generate deterministic prompts for testing KV prefix caching behavior in LLM inference servers such as vLLM and SGLang. The prompts use controlled shared prefixes plus small unique suffixes so benchmark clients can compare cache reuse, latency, and throughput across server configurations.
|
| 597 |
|