Skip to content

LLM inference in C/C++

C++ ggml

dflash : load optional FC input scale (#30945)

Hrishith Thadicherla4912ac24f12a
diff --git a/src/llama-model.h b/src/llama-model.h
index 9fe9d77a23..cdf060db29 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -779,8 +779,9 @@ struct llama_model {
struct ggml_tensor * per_layer_proj_norm = nullptr;
// eagle3 / dflash feature fusion layer
- struct ggml_tensor * fc = nullptr;
- struct ggml_tensor * fc_s = nullptr;
+ struct ggml_tensor * fc = nullptr;
+ struct ggml_tensor * fc_s = nullptr;
+ struct ggml_tensor * fc_in_s = nullptr;
struct ggml_tensor * d2t = nullptr; // draft to target vocabulary mapping
// dspark
diff --git a/src/models/dflash.cpp b/src/models/dflash.cpp
index 6f28d2f1db..e15af6afdc 100644
--- a/src/models/dflash.cpp
+++ b/src/models/dflash.cpp
@@ -238,6 +238,7 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
fc = create_tensor(tn(LLM_TENSOR_FC, "weight"), { n_embd_inp, n_embd }, 0);
fc_s = create_tensor(tn(LLM_TENSOR_FC, "scale"), { 1 }, TENSOR_NOT_REQUIRED);
+ fc_in_s = create_tensor(tn(LLM_TENSOR_FC, "input_scale"), { 1 }, TENSOR_NOT_REQUIRED);
output_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), { n_embd }, 0); // encoder hidden_norm (after fc)
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0); // decoder final norm