30 days ago
-
Card: how to load with transformers, and why float32 matters
-
Load in float32 by default: bfloat16 weights change the router's expert selection and collapse the output
-
Load with transformers: trust_remote_code modelling code, a tokenizers-format tokenizer.json, and auto_map in config.json (weights untouched)
-
Card: drop the reference to a scheduled task that no longer exists
-
Tokenizer: rename the format id nanosarvam-bpe -> anulm-bpe (merges unchanged; older readers still accept both)
-
AnuLM v0.1.0
-
initial commit