perplexity-ai

/

pplx-embed-context-v1-4b

like 37

Follow

Perplexity 2.67k

modeling.py built the bidirectional mask with a hardcoded create_causal_mask(input_embeds=..., cache_position=...) call. That signature drifted across transformers 5.x:

• input_embeds was renamed to inputs_embeds (~5.2),
• cache_position became optional (5.6) then was dropped (5.9).

So the old code broke on transformers 5.9+ (unexpected keyword argument 'input_embeds'). Detect the accepted kwargs once at import time and build the call accordingly — the same approach already shipped in pplx-embed-v1-0.6b. PPLXQwen3ContextualModel / encode() are unchanged.

Validated by loading the model across transformers 5.0.0, 5.8.1, 5.9.0 and 5.12.0 — it loads, runs a forward pass, and produces identical output on all of them. Scope: transformers 5.x.