Keep Minimax's indexer tensors at F32 for speed and accuracy (#26144)

* Keep Minimax's indexer tensors at F32 for speed and accuracy * name -> new_name
2026-07-26 18:02:56 -04:00
parent 55b7d6c4c7
commit 7657a6c26a
2 changed files with 9 additions and 0 deletions
@@ -326,6 +326,10 @@ static bool tensor_allows_quantization(const llama_model_quantize_params * param
    quantize &= name.find("ssm_conv1d") == std::string::npos;
    quantize &= name.find("shortconv.conv.weight") == std::string::npos;

+    // do not quantize MiniMax's indexer projection weights, they are tiny
+    quantize &= name.find("indexer.k_proj.weight") == std::string::npos;
+    quantize &= name.find("indexer.q_proj.weight") == std::string::npos;
+
    // do not quantize RWKV's small yet 2D weights
    quantize &= name.find("time_mix_first.weight") == std::string::npos;
    quantize &= name.find("time_mix_w0.weight") == std::string::npos;