lynae-1219 commited on
Commit
0fbd016
·
verified ·
1 Parent(s): e2a84f8

Remove predictions and roofline_quadrant configs, keep 3 configs only

Browse files
Files changed (1) hide show
  1. README.md +2 -128
README.md CHANGED
@@ -28,14 +28,6 @@ configs:
28
  data_files:
29
  - split: train
30
  path: kernel_profiles/kernels_labeled.parquet
31
- - config_name: roofline_quadrant
32
- data_files:
33
- - split: train
34
- path: kernel_profiles/roofline_quadrant.parquet
35
- - config_name: predictions
36
- data_files:
37
- - split: train
38
- path: predictions/serving_predictions.parquet
39
  dataset_info:
40
  - config_name: trace_replay
41
  features:
@@ -226,121 +218,11 @@ dataset_info:
226
  splits:
227
  - name: train
228
  num_examples: 148077
229
- - config_name: roofline_quadrant
230
- features:
231
- - name: model
232
- dtype: string
233
- - name: profile
234
- dtype: string
235
- - name: concurrency
236
- dtype: int64
237
- - name: engine
238
- dtype: string
239
- - name: hardware
240
- dtype: string
241
- - name: oi
242
- dtype: float64
243
- - name: cf_gb
244
- dtype: float64
245
- - name: output_tput
246
- dtype: float64
247
- - name: tpot_ms
248
- dtype: float64
249
- - name: ttft_ms
250
- dtype: float64
251
- splits:
252
- - name: train
253
- num_examples: 2163
254
- - config_name: predictions
255
- features:
256
- - name: hardware_config
257
- dtype: string
258
- - name: model
259
- dtype: string
260
- - name: backend
261
- dtype: string
262
- - name: profile
263
- dtype: string
264
- - name: data_scope
265
- dtype: string
266
- - name: concurrency
267
- dtype: int64
268
- - name: isl
269
- dtype: int64
270
- - name: osl
271
- dtype: int64
272
- - name: calibration_status
273
- dtype: string
274
- - name: ttft_validation_scope
275
- dtype: string
276
- - name: ttft_kernel_ms
277
- dtype: float64
278
- - name: ttft_base_ms
279
- dtype: float64
280
- - name: ttft_floor_ms
281
- dtype: float64
282
- - name: ttft_first_decode_ms
283
- dtype: float64
284
- - name: ttft_queue_ms
285
- dtype: float64
286
- - name: itl_meas
287
- dtype: float64
288
- - name: total_context_tokens
289
- dtype: int64
290
- - name: new_prefill_tokens
291
- dtype: int64
292
- - name: cached_context_tokens
293
- dtype: int64
294
- - name: cache_hit_rate
295
- dtype: float64
296
- - name: cache_aware_applied
297
- dtype: bool
298
- - name: cache_feature_source
299
- dtype: string
300
- - name: cache_prediction_regime
301
- dtype: string
302
- - name: ttft_prediction_supported
303
- dtype: bool
304
- - name: multiturn_prediction_mode
305
- dtype: string
306
- - name: predicted_turn_count
307
- dtype: float64
308
- - name: total_successful_turn_requests
309
- dtype: float64
310
- - name: mean_predicted_turn_ttft_ms
311
- dtype: float64
312
- - name: mean_predicted_turn_tpot_ms
313
- dtype: float64
314
- - name: multiturn_turn_predictions
315
- dtype: string
316
- - name: ttft_pred
317
- dtype: float64
318
- - name: ttft_meas
319
- dtype: float64
320
- - name: ttft_err
321
- dtype: float64
322
- - name: tpot_pred
323
- dtype: float64
324
- - name: tpot_meas
325
- dtype: float64
326
- - name: tpot_err
327
- dtype: float64
328
- - name: e2el_pred
329
- dtype: float64
330
- - name: e2el_meas
331
- dtype: float64
332
- - name: e2el_err
333
- dtype: float64
334
- - name: measurement_semantics_warning
335
- dtype: string
336
- splits:
337
- - name: train
338
- num_examples: 4715
339
  ---
340
 
341
  # AgentPerfBench
342
 
343
- LLM inference benchmark: 3,392 serving runs, 148,077 per-kernel NCU profiles, and 4,715 latency predictions across 9 models, 14 GPU configurations, and 2 serving engines (vLLM 0.19.0, SGLang 0.5.9). All models served in BF16 except gpt-oss, which uses mxfp4 for projection weights.
344
 
345
  ## Dataset configurations
346
 
@@ -360,14 +242,6 @@ ISL/OSL sampled from lognormal fits to real workload statistics. 42 combinations
360
 
361
  Per-kernel Nsight Compute (ncu) profiles across 4 GPUs (A100, H100, RTX 3090, RTX 2080 Ti) and 13 model/sweep sources.
362
 
363
- ### roofline_quadrant (2,163 rows)
364
-
365
- Operational intensity and achieved throughput per kernel, for roofline analysis. H100 reference hardware (989 peak TFLOPS, 3.35 TB/s HBM).
366
-
367
- ### predictions (4,715 rows)
368
-
369
- Predicted vs. measured TTFT, TPOT, and E2EL for each serving configuration, with cache-aware prediction metadata. 14 hardware configs.
370
-
371
  ### Concurrency filtering
372
 
373
  Concurrency is controlled by a fixed-size connection pool. Trace replay uses levels {1, 5, 10, 20, 40, 80}; distributional uses {1, 5, 10, 40, 80, 200, 320}.
@@ -449,7 +323,7 @@ Each row in `summary.parquet` (both configs):
449
  from datasets import load_dataset
450
 
451
  ds = load_dataset("agent-perf-bench/AgentPerfBench", "trace_replay")
452
- # or "distributional", "kernels_labeled", "roofline_quadrant", "predictions"
453
  ```
454
 
455
  ## Benchmark methodology
 
28
  data_files:
29
  - split: train
30
  path: kernel_profiles/kernels_labeled.parquet
 
 
 
 
 
 
 
 
31
  dataset_info:
32
  - config_name: trace_replay
33
  features:
 
218
  splits:
219
  - name: train
220
  num_examples: 148077
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
221
  ---
222
 
223
  # AgentPerfBench
224
 
225
+ LLM inference benchmark: 3,392 serving runs and 148,077 per-kernel NCU profiles across 9 models, 14 GPU configurations, and 2 serving engines (vLLM 0.19.0, SGLang 0.5.9). All models served in BF16 except gpt-oss, which uses mxfp4 for projection weights.
226
 
227
  ## Dataset configurations
228
 
 
242
 
243
  Per-kernel Nsight Compute (ncu) profiles across 4 GPUs (A100, H100, RTX 3090, RTX 2080 Ti) and 13 model/sweep sources.
244
 
 
 
 
 
 
 
 
 
245
  ### Concurrency filtering
246
 
247
  Concurrency is controlled by a fixed-size connection pool. Trace replay uses levels {1, 5, 10, 20, 40, 80}; distributional uses {1, 5, 10, 40, 80, 200, 320}.
 
323
  from datasets import load_dataset
324
 
325
  ds = load_dataset("agent-perf-bench/AgentPerfBench", "trace_replay")
326
+ # or "distributional", "kernels_labeled"
327
  ```
328
 
329
  ## Benchmark methodology