Register missing korean/el/th PP-OCRv5_mobile_rec for HPI (gpu_cuda12/paddle311)

[ddoron9]

Summary

gpu_cuda12/paddle311 in paddlex/inference/models/runners/hpi/hpi_model_info_collection.json is missing three PP-OCRv5 multilingual recognition models — korean_PP-OCRv5_mobile_rec, el_PP-OCRv5_mobile_rec, th_PP-OCRv5_mobile_rec — that are present in every other environment/version bucket. As a result, HPI backend selection fails for these models on CUDA 12 with PaddlePaddle ≥ 3.1.0.

Symptom

Building an HPI predictor for korean_PP-OCRv5_mobile_rec on a GPU + CUDA 12 environment with PaddlePaddle ≥ 3.1.0 raises:

RuntimeError: No inference backend and configuration could be suggested.
Reason: 'korean_PP-OCRv5_mobile_rec' is not a known model.

Disabling HPI works, so the model itself is fine — it is simply not registered in the HPI prior-knowledge collection for this bucket.

Root cause

suggest_inference_backend_and_config() in paddlex/inference/models/runners/hpi/backend.py normalizes any Paddle ≥ 3.1.0 to the paddle311 knowledge bucket:

if paddle_version[:3] >= (3, 1, 0):
    # "...The prior knowledge of Paddle 3.1.1 will be used."
    paddle_version = (3, 1, 1, None)   # -> key "paddle311"

It then looks up collection[env]["paddle311"][model_name]. For env = gpu_cuda12, the paddle311 bucket is missing the three models above — they exist in gpu_cuda12/paddle30 and gpu_cuda12/paddle31, and in every cpu_x64 / gpu_cuda11 bucket — so the lookup hits the "is not a known model" branch.

Per-bucket model count (and presence of the korean/el/th PP-OCRv5 mobile rec models):

env	paddle30	paddle31	paddle311
cpu_x64	281 ✓	281 ✓	281 ✓
gpu_cuda11	281 ✓	281 ✓	281 ✓
gpu_cuda12	281 ✓	281 ✓	278 ✗ (korean / el / th missing)

Fix

Add the three missing entries to gpu_cuda12/paddle311, mirroring gpu_cuda12/paddle31 (backends ["onnxruntime", "paddle"]). This restores exact key parity between gpu_cuda12/paddle311 and gpu_cuda12/paddle31.

Verification

• json.load() succeeds on the edited file.
• set(gpu_cuda12/paddle311) == set(gpu_cuda12/paddle31) after the change (278 → 281 models).
• No other env/version bucket is modified — the diff is +12 lines (three entries).

Related

• 'PP-OCRv6_small_det' is not a known model for hpi #5162 — 'PP-OCRv6_small_det' is not a known model for hpi (same registration-gap class).
• 'latin_PP-OCRv5_mobile_rec' is not a known model when enable_hpi=True PaddleOCR#16450 — 'latin_PP-OCRv5_mobile_rec' is not a known model when enable_hpi=True (same class).

[CLAassistant]

All committers have signed the CLA.

[paddle-bot]

Thanks for your contribution!