llama系列最小上下文长度提到16k （早期的llama上下文太小了）

ztxz16 · Jun 27, 2024 · 73b6d27 · 73b6d27
1 parent b1e6c8e
commit 73b6d27
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/src/models/llama.cpp b/src/models/llama.cpp
@@ -92,7 +92,7 @@ namespace fastllm {
         if (this->weight.dicts.find("rope_scaling.factor") != this->weight.dicts.end()) {
             rope_factor = atof(this->weight.dicts["rope_scaling.factor"].c_str());
         }
-        std::pair<std::vector<float>, std::vector<float>> &&pair = this->UpdateRotaryPosEmb(rope_base, rope_factor);
+        std::pair<std::vector<float>, std::vector<float>> &&pair = this->UpdateRotaryPosEmb(rope_base, rope_factor, std::max(max_positions, 16384));
         sinData.ToDevice(DataDevice::CPU);
         cosData.ToDevice(DataDevice::CPU);
         sinData.CopyFrom(Data(DataType::FLOAT32, { (int)this->sin.size(), (int)this->sin[0].size() }, pair.first));