RWKV v6: Add time_mix_decay_w1/w2 in quant exclusion list (ggml-org#9387)

Signed-off-by: Molly Sophia <[email protected]>

Author: MollySophia

convert_hf_to_gguf.py

@@ -302,6 +302,8 @@ def prepare_tensors(self):
                             gguf.MODEL_TENSOR.TIME_MIX_FIRST,
                             gguf.MODEL_TENSOR.TIME_MIX_W1,
                             gguf.MODEL_TENSOR.TIME_MIX_W2,
+                            gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1,
+                            gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2,
                         )
                     )
                     or not new_name.endswith(".weight")

src/llama.cpp

@@ -17530,6 +17530,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         quantize &= name.find("time_mix_first.weight") == std::string::npos;
         quantize &= name.find("time_mix_w1.weight") == std::string::npos;
         quantize &= name.find("time_mix_w2.weight") == std::string::npos;
+        quantize &= name.find("time_mix_decay_w1.weight") == std::string::npos;
+        quantize &= name.find("time_mix_decay_w2.weight") == std::string::npos;
 
         // do not quantize relative position bias (T5)
         quantize &= name.find("attn_rel_b.weight") == std::string::npos;