asdwb/cara_latency_prediction

CARA Latency Prediction Dataset

A large-scale dataset for training latency predictors in heterogeneous LLM serving systems. Contains 450,000 request-level records from 18 model instances across 4 GPU types, collected under 10 different load levels.

Dataset Description

Each record captures the instance state at scheduling time and the actual end-to-end latency after request completion. This enables training learned latency predictors that map observable system state to request latency — analogous to learned index structures in databases.

Key Statistics

Split	Records	Prompts Source	Sweeps	QPS Levels	File Size
Train	400,000	cara_model_estimator train split	2	9-36 (10 levels)	423 MB
Test	50,000	cara_model_estimator test split	1	9-36 (10 levels)	53 MB

Cluster Configuration

18 instances across 4 GPU types serving 4 Qwen2.5 model sizes:

Model	GPU	Instances	Tensor Parallel	vLLM Version
Qwen2.5-72B	A100 40GB	2	4	v1 (dev11708)
Qwen2.5-14B	V100 16GB	3	4	v1 (dev11708)
Qwen2.5-7B	A30 24GB	5	1	v1 (dev11708)
Qwen2.5-3B	A30 24GB	3	1	v1 (dev11708)
Qwen2.5-3B	P100 12GB	5	1	v0 (dev5353)

Full instance configurations are in instance_metadata.json.

Files

File	Description
train.jsonl	Training records with aggregate schedule state features (for XGBoost, analytical models)
test.jsonl	Test records (same schema as train)
train_queue_details.jsonl	Training records with per-request queue details (for LSTM, simulators)
test_queue_details.jsonl	Test records with queue details
instance_metadata.json	Anonymized instance configurations
train.meta.json	Training set statistics
test.meta.json	Test set statistics
scripts/prepare_latency_dataset.py	Post-processing script for reproducibility

Schema

Flat records (train.jsonl, test.jsonl) — 35 fields

Identifiers:

• request_id (str): Unique request identifier
• instance_id (str): Anonymized instance ID (e.g., "instance_0"), maps to instance_metadata.json
• sweep (int): Sweep number (1 or 2)
• qps_level (int): Cluster-level queries per second (9, 12, 15, 18, 21, 24, 27, 30, 33, 36)

Request features:

• num_prompt_tokens (int): Input prompt length in tokens (range: 20-717)
• actual_output_tokens (int): Actual generated output tokens (range: 1-1024)

Schedule state (at scheduling time):

• num_running (int): Currently running requests on this instance
• num_waiting (int): Queued requests
• num_active_decode_seqs (int): Actively decoding sequences
• decode_ctx_p50/p95/max (float): Decode context length distribution
• pending_prefill_tokens (int): Total prefill tokens in queue
• pending_decode_tokens (int): Total decode tokens in queue
• kv_cache_utilization (float): KV cache usage fraction
• kv_free_blocks (int): Available KV cache blocks
• token_budget_per_iter (int): Scheduler token budget
• prefill_chunk_size (int): Chunked prefill size
• max_num_seqs (int): Max concurrent sequences
• num_preempted (int): Preempted sequence count
• ema_decode_tok_per_s (float): EMA decode throughput
• ema_prefill_tok_per_s (float): EMA prefill throughput
• ema_decode_iter_ms (float): EMA per-iteration decode latency
• kv_evictions_per_s (float): KV cache eviction rate

Derived features:

• running_requests_count (int): Count of running requests
• waiting_requests_count (int): Count of waiting requests
• total_pending_tokens (int): Sum of pending prefill + decode tokens
• queue_depth (int): num_running + num_waiting

Overhead:

• probe_latency_ms (float): Sidecar probe latency
• prediction_latency_ms (float): Prediction overhead

Timestamps:

• prediction_timestamp (float): Unix timestamp at scheduling
• completion_timestamp (float): Unix timestamp at completion

Target labels:

• actual_e2e_latency (float): End-to-end latency in seconds
• actual_ttft (float): Time to first token in seconds
• actual_tpot (float): Time per output token in seconds

Queue details (train_queue_details.jsonl) — 13 fields

Same identifiers, request features, timestamps, and target labels as flat records, plus:

• running_requests (list): Per-request details for each running request:
  • num_prompt_tokens (int)
  • num_computed_tokens (int): Tokens already processed
  • num_output_tokens (int): Output tokens generated so far
  • total_num_tokens (int): Total tokens so far
  • expected_output_tokens (int): Ground-truth output length (from length predictor in production)
• waiting_requests (list): Same schema as running_requests

Collection Method

1. Real prompts sampled from the cara_model_estimator dataset
2. Sent through a coordinator with random scheduling to 18 heterogeneous instances
3. Sidecar on each instance captures (state, latency) pairs
4. Collected at 10 QPS levels (9-36) covering idle to saturated conditions
5. Train/test split aligned with cara_model_estimator splits (no prompt overlap)

Intended Use

• Training latency predictors for LLM serving schedulers (XGBoost, LSTM, neural networks)
• Benchmarking simulation-based vs learned latency prediction
• Studying heterogeneous LLM serving behavior across GPU types
• Evaluating scheduling policies under different load conditions

Citation

If you use this dataset, please cite:

@dataset{cara_latency_2026,
  title={CARA Latency Prediction Dataset},
  author={Wei Da},
  year={2026},
  url={https://huggingface.co/datasets/asdwb/cara_latency_prediction}
}

Related

• cara_model_estimator — Multi-model quality and length estimation dataset
• CARA: Context-Aware Resource Allocation for heterogeneous LLM serving