From 403fbacbbcd3650abcc53e84f5e6bf88e51dd6f5 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Mon, 31 Mar 2025 16:36:25 +0200 Subject: [PATCH] convert : Qwerky : use lora_rank_tokenshift and lora_rank_decay if present (#12667) --- convert_hf_to_gguf.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 0919cd3f..37d6cc26 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -3557,8 +3557,8 @@ class RWKV6Qwen2Model(Rwkv6Model): head_size = hidden_size // num_attention_heads rms_norm_eps = self.hparams["rms_norm_eps"] intermediate_size = self.hparams["intermediate_size"] - time_mix_extra_dim = 64 if hidden_size >= 4096 else 32 - time_decay_extra_dim = 128 if hidden_size >= 4096 else 64 + time_mix_extra_dim = self.hparams.get("lora_rank_tokenshift", 64 if hidden_size >= 4096 else 32) + time_decay_extra_dim = self.hparams.get("lora_rank_decay", 128 if hidden_size >= 4096 else 64) # RWKV isn't context limited self.gguf_writer.add_context_length(1048576)