Issues

perf(triton): ocr_pipeline runs one instance, so OCR is capped at about 21 images/s (2 instances 42, 4 instances 80)

davidamacey/OpenProcessor#240 · opened

on Oct 10, 2026

· v0.6.0

fix(api): oversize upload returns 500 instead of 413, and the 50 MB cap counts the whole multipart request

davidamacey/OpenProcessor#239 · opened

on Oct 10, 2026

· v0.6.0

auto_label job stuck in queued when the worker is down is never repaired or surfaced (found in #234)

davidamacey/OpenProcessor#236 · opened

on Oct 10, 2026

· v0.6.0

cluster-refresh: persist the per-project growth state across daemon restarts (follow-up to #212)

davidamacey/OpenProcessor#235 · opened

on Oct 10, 2026

· v0.6.0

perf: scale ladder 10k, 100k, 1M ingest on the full server (WP-S7)

davidamacey/OpenProcessor#232 · opened

on Oct 10, 2026

· v0.6.0

perf(serving): ortloom-serve vs Triton head-to-head on ingest-path models (WP-S6)

davidamacey/OpenProcessor#231 · opened

on Oct 10, 2026

· v0.6.0

1

perf(opensearch): datastore write ceiling and cluster pass at 100k-1M scale (WP-S5)

davidamacey/OpenProcessor#230 · opened

on Oct 10, 2026

· v0.6.0

perf(curation): bulk ingest runner with prefetch, multi-GPU sharding and resume (WP-S4)

davidamacey/OpenProcessor#229 · opened

on Oct 10, 2026

· v0.6.0

1

perf(triton): data-parallel PE across both A6000s and the 12 GB card, NUMA-pinned (WP-S3)

davidamacey/OpenProcessor#228 · opened

on Oct 10, 2026

· v0.6.0

perf(models): PE TensorRT engine efficiency pass, fused attention and profile shape (WP-S2)

davidamacey/OpenProcessor#227 · opened

on Oct 10, 2026

· v0.6.0

perf(triton): Triton and client concurrency sweep on the FP16 ingest stack (WP-S1)

davidamacey/OpenProcessor#226 · opened

on Oct 10, 2026

· v0.6.0

MobileCLIP image encoder FP16 build fails on TensorRT 11.1 (reparam conv), exporter ships FP32

davidamacey/OpenProcessor#224 · opened

on Oct 10, 2026

perf(search): vector database evaluation for the embedding index (WP-V)

davidamacey/OpenProcessor#221 · opened

on Oct 10, 2026

· v0.6.0

plan: faces and OCR on resident crops (WP-2.5)

davidamacey/OpenProcessor#220 · opened

on Oct 10, 2026

· v0.6.0

perf(curation): work crops at model size for the VLM and region consumers (WP-4.1)

davidamacey/OpenProcessor#219 · opened

on Oct 10, 2026

· v0.6.0

perf(curation): clustering benchmark harness, GPU IVF without the cuML gate, incremental refresh (WP-3.1)

davidamacey/OpenProcessor#218 · opened

on Oct 10, 2026

· v0.6.0

perf(triton): GPU decode once and a GPU-resident ingest pipeline (WP-2.4)

davidamacey/OpenProcessor#217 · opened

on Oct 10, 2026

· v0.6.0

perf(curation): cross-image PE batching in batch ingest and embed-missing (WP-2.3)

davidamacey/OpenProcessor#216 · opened

on Oct 10, 2026

· v0.6.0

perf(triton): system shared memory between the API and Triton (WP-2.2)

davidamacey/OpenProcessor#215 · opened

on Oct 10, 2026

· v0.6.0

perf(triton): uint8 detector and PE inputs with in-graph normalization (WP-2.1)

davidamacey/OpenProcessor#214 · opened

on Oct 10, 2026

· v0.6.0

perf(models): measured INT8 evaluation of the PE image encoder on Ampere (WP-1.6)

davidamacey/OpenProcessor#213 · opened

on Oct 10, 2026

· v0.6.0

perf(api): remove API CPU waste on the ingest path (WP-1.3)

davidamacey/OpenProcessor#211 · opened

on Oct 10, 2026

· v0.6.0

perf(api): attribute API CPU outside the timed ingest stages (WP-1.2)

davidamacey/OpenProcessor#210 · opened

on Oct 10, 2026

· v0.6.0

perf(models): rebuild stale PE engines and stamp every TensorRT plan with its build facts (WP-1.1)

davidamacey/OpenProcessor#209 · opened

on Oct 10, 2026

· v0.6.0

perf(bench): parity capture and compare tooling for the v0.6.0 waves (WP-0.1)

davidamacey/OpenProcessor#208 · opened

on Oct 10, 2026

· v0.6.0

Navigation Menu

Repository navigation

Issues

perf(triton): ocr_pipeline runs one instance, so OCR is capped at about 21 images/s (2 instances 42, 4 instances 80)

fix(api): oversize upload returns 500 instead of 413, and the 50 MB cap counts the whole multipart request

auto_label job stuck in queued when the worker is down is never repaired or surfaced (found in #234)

cluster-refresh: persist the per-project growth state across daemon restarts (follow-up to #212)

perf: scale ladder 10k, 100k, 1M ingest on the full server (WP-S7)

perf(serving): ortloom-serve vs Triton head-to-head on ingest-path models (WP-S6)

perf(opensearch): datastore write ceiling and cluster pass at 100k-1M scale (WP-S5)

perf(curation): bulk ingest runner with prefetch, multi-GPU sharding and resume (WP-S4)

perf(triton): data-parallel PE across both A6000s and the 12 GB card, NUMA-pinned (WP-S3)

perf(models): PE TensorRT engine efficiency pass, fused attention and profile shape (WP-S2)

perf(triton): Triton and client concurrency sweep on the FP16 ingest stack (WP-S1)

MobileCLIP image encoder FP16 build fails on TensorRT 11.1 (reparam conv), exporter ships FP32

perf(search): vector database evaluation for the embedding index (WP-V)

plan: faces and OCR on resident crops (WP-2.5)

perf(curation): work crops at model size for the VLM and region consumers (WP-4.1)

perf(curation): clustering benchmark harness, GPU IVF without the cuML gate, incremental refresh (WP-3.1)

perf(triton): GPU decode once and a GPU-resident ingest pipeline (WP-2.4)

perf(curation): cross-image PE batching in batch ingest and embed-missing (WP-2.3)

perf(triton): system shared memory between the API and Triton (WP-2.2)

perf(triton): uint8 detector and PE inputs with in-graph normalization (WP-2.1)

perf(models): measured INT8 evaluation of the PE image encoder on Ampere (WP-1.6)

perf(api): remove API CPU waste on the ingest path (WP-1.3)

perf(api): attribute API CPU outside the timed ingest stages (WP-1.2)

perf(models): rebuild stale PE engines and stamp every TensorRT plan with its build facts (WP-1.1)

perf(bench): parity capture and compare tooling for the v0.6.0 waves (WP-0.1)