- Vendor configuration.py / modeling.py from perplexity-ai/pplx-embed-v1-0.6b
@ 2c4d510 as configuration_pplx_qwen3.py / modeling_pplx_qwen3.py. Loading
no longer executes unpinned code from another repo, works offline from a
local checkout, and drops the dynamic_module_utils internal API.
- save_pretrained also writes the tokenizer; without it, predict/mask on a
reloaded directory failed with "Couldn't instantiate the backend tokenizer".
- The lazy tokenizer load reuses the revision / token / cache_dir /
local_files_only kwargs given to from_pretrained.
- Return last_hidden_state in the encoder dtype instead of the fp32 copy.
- Drop the dead backbone.auto_map from config.json; add transformers_version.
- README: vendored encoder, truncation at max_seq_len, tokenizer in save dir.
Token logits, sensitivity and spans are bitwise identical to the previous
revision of this PR.
