v0.1.8

What's Changed

• fix: harden JSON config loader and collator negative padding shape (67dacda)
• fix: make _load_json_training_args resilient to missing optimizer/scheduler (c34666d)

---

📥 Install

pip install omnivector-embed==0.1.8

v0.1.7

What's Changed

• fix: update tests for resume_from_checkpoint removal and new MTEB categories (225aef2)
• fix: remove hardcoded resume_from_checkpoint from stage YAML configs (1147535)
• fix: run.sh — use --model-path for weight loading, not --resume (cccf156)
• style: ruff format mteb_runner.py (f43d33d)
• feat: full MTEB 56-task benchmark evaluation (NV-Embed v2 leaderboard) (d9d9b17)
• fix: run.sh — Stage 3 resume from Stage 2, evaluate all 3 stages, ONNX export final model (deecbad)
• fix: convert loss tensor to float for TensorBoard logging (fd3353a)
• chore: ruff format + add open_clip_torch to core dependencies (bfbeb87)
• fix: add grad_norm param to _maybe_log_save_evaluate — matches transformers 4.44 signature (5444171)
• fix: loss must be scalar tensor, not float — fix dict key overwrite in MRLInfoNCELoss (443e897)
• fix: handle zero negatives in EmbeddingDataCollator — skip empty torch.stack (c551679)
• fix: wire training forward pass — model accepts collated batch, computes MRL loss (1894c74)
• fix: remove prescale_gradients — incompatible with ZeRO-2 (1e0ce21)
• fix: remove optimizer/scheduler from DeepSpeed config entirely (66892e4)

---

📥 Install

pip install omnivector-embed==0.1.7

v0.1.6

What's Changed

• fix: set DeepSpeed optimizer/scheduler to auto (9041a8e)

---

📥 Install

pip install omnivector-embed==0.1.6

v0.1.5

What's Changed

• fix: remove CPU offloading from DeepSpeed config (9af4a54)
• fix: add gradient_checkpointing_enable/disable to OmniVectorModel (dfb7582)

---

📥 Install

pip install omnivector-embed==0.1.5

v0.1.4

What's Changed

• fix: set DeepSpeed batch sizes to auto for any GPU count (992b03a)
• fix: pass split/max_samples to loader constructor, not load() (739cac0)

---

📥 Install

pip install omnivector-embed==0.1.4

v0.1.3

What's Changed

• fix: pass embed_dim from backbone to LatentAttentionPooling (527126c)

---

📥 Install

pip install omnivector-embed==0.1.3

v0.1.2

What's Changed

• fix: add --local_rank arg to training.py for DeepSpeed launcher (c85cecf)

---

📥 Install

pip install omnivector-embed==0.1.2

v0.1.1

What's Changed

• use query embeddings instead of queries. (fd414ca)

---

📥 Install

pip install omnivector-embed==0.1.1

v0.1.0

What's Changed

Initial release 🎉

---

📥 Install

pip install omnivector-embed==0.1.0