galthran-wq / clip-web-service / 26038482344

78%

Build Type

push

github

Committed by web-flow

Commit Message

refactor: drop onnx-clip, load CLIP ONNX directly from HuggingFace (#3)

* refactor: drop onnx-clip, load CLIP ONNX directly from HuggingFace

The onnx-clip package downloaded its ONNX checkpoint from an S3 bucket
that has since expired, breaking startup. Replace it with direct loading
from openai/clip-vit-base-patch32 on HuggingFace Hub, pinned to commit
12b36594d.

- pyproject.toml: drop onnx-clip; add onnxruntime, huggingface-hub,
  tokenizers, pillow, numpy as direct dependencies.
- src/services/clip_service.py: rewrite to use onnxruntime.InferenceSession
  on the combined model.onnx (reads logits_per_image directly), with
  CLIP-specific image preprocessing (bicubic shorter-side resize +
  center crop + CLIP mean/std) and the standalone tokenizers library
  for prompts. Public API (CLIPService.__init__/decode_image/classify)
  and HTTP /classify/batch shape are unchanged.
- src/config.py: add CLIP_MODEL_REPO, CLIP_MODEL_REVISION,
  CLIP_ONNX_FILENAME, CLIP_TOKENIZER_FILENAME knobs.
- Dockerfile, Dockerfile.gpu: pre-pull the ONNX and tokenizer into
  HF_HOME=/app/.hf_cache at build time and set HF_HUB_OFFLINE=1 at
  runtime, so the service never reaches the network at startup.
- CLAUDE.md: refresh architecture description + config table.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(clip): restore batch_size chunking + clarify HF cache constraint

Review noted that the rewrite silently dropped the chunking semantics
of OnnxClip(batch_size=...) — all images were run in a single
session.run, which spikes GPU/CPU memory at CLIP_MAX_BATCH_SIZE=64.

Restore the original behavior: classify() now loops session.run over
image chunks of size batch_size, concatenates the per-chunk
logits_per_image, then applies softmax. Verified that chunked and
single-shot results match exactly (7 images at batch_size=3 → 3
session.run calls → identical probabilities to a 1-image call).

Also clarify in CLAUDE.m... (continued)

Coverage Stats

16 of 51 new or added lines in 2 files covered. (31.37%)

1 existing line in 1 file now uncovered.

181 of 231 relevant lines covered (78.35%)

0.78 hits per line