Skip to content

RESUME_SUPPORT.md

Latest commit

 

History

History
41 lines (33 loc) · 2.97 KB

RESUME_SUPPORT.md

File metadata and controls

41 lines (33 loc) · 2.97 KB
Raw

Resume Phrase Support Ledger

Status date: 2026-07-25.

Exact phrase Status Evidence / blocker
Python supported Typed package, CLI, aggregation, service export, and 12 passing tests
PyTorch supported Three real torchvision pretrained models executed on CPU and CUDA
CUDA supported Tesla T4 execution with CUDA events, synchronization, and NVML memory evidence
ONNX supported All models exported and checked; corrected GPU trials verify CUDAExecutionProvider
TensorRT supported 24/24 corrected engine build/inference trials succeeded with TensorRT 11
Ray supported Integration tests execute Ray trials; Modal workers checksummed both public datasets
Built a system that automatically selects quantization, pruning, compilation, batching, and runtime configurations for a model, evaluation set, and target device. supported CPU and T4 searches exercised INT8/FP16, pruning, Inductor, batch 1/8, PyTorch, ONNX Runtime, and TensorRT on labeled public data
Implemented multi-objective optimization across accuracy, latency, throughput, and memory, then exported the best configuration as a deployable inference service. supported Pareto/constrained selection is tested; task accuracy/F1, latency, constrained throughput, RSS/GPU memory, artifacts, and service bundles are retained
Reduced median latency by [X]% and memory by [Y]% while limiting quality loss to [Z]% across [N] models and [N] hardware configurations. unsupported as written Batch-one latency is 79.616% lower and quality loss is 0.0 points across 3 models × 2 physical targets, but no positive apples-to-apples cross-target memory reduction was established
external users unsupported / not claimed No authorization or evidence
published or deployed unsupported / not claimed Explicitly not performed

Resume-safe measured alternative

Reduced median batch-one latency by 79.6% and deployable artifact size by 21.0% while limiting task-quality loss to 0.0 percentage points across 3 real pretrained models and 2 physical hardware targets.

Artifact size is named explicitly; it is not mislabeled as runtime memory.

Eligible evidence

  • Models: ResNet-18, MobileNetV3-Small, EfficientNet-B0
  • Public data: deterministic 256-image Imagenette/Imagewoof validation samples
  • Physical targets: Apple Silicon CPU and NVIDIA Tesla T4
  • Corrected GPU outcome: 60/60 successful trials, including 12/12 ONNX Runtime CUDA and 24/24 TensorRT trials
  • Ten warmups and 50 retained measurements per configuration
  • Batch-one p50 with bootstrap 95% intervals; batching reported separately as throughput under a 100 ms batch-p95 constraint
  • Artifact size, process RSS, and NVML peak GPU memory retained as separate fields
  • Final combined evidence: benchmarks/final-evidence.json

The synthetic run and first Modal pass remain preserved but are ineligible for resume metrics.