1 parent
d1044bc827
commit
9309bf1bcf
20 files changed
+1112
-2
No files matched your search
@@ -361,6 +361,16 @@ static bool tensor_allows_quantization(const llama_model_quantize_params * param
|
||||
quantize &= name.find("indexer.k_proj.weight") == std::string::npos;
|
||||
quantize &= name.find("indexer.q_proj.weight") == std::string::npos;
|
||||
|
||||
// metis16: small control tensors that steer every layer; the stream and hc_ tables (ne[0] = 4) would fall back to F16
|
||||
if (arch == LLM_ARCH_METIS16) {
|
||||
quantize &= name.rfind("depth_mem_", 0) == std::string::npos;
|
||||
quantize &= name.rfind("cont_", 0) == std::string::npos;
|
||||
quantize &= name.find("stream_embd") == std::string::npos;
|
||||
quantize &= name.find("attn_pass_lora") == std::string::npos;
|
||||
quantize &= name.find("ffn_exp_embd") == std::string::npos;
|
||||
quantize &= name.find("hc_") == std::string::npos;
|
||||
}
|
||||
|
||||
// glm5-next
|
||||
if (arch == LLM_ARCH_GLM5_NEXT) {
|
||||
quantize &= name.find("hc_") == std::string::npos;
|
||||
|
||||
Reference in new issue
Block a user