LLM inference in C/C++
C++ ggml
models : appropriately flag noscan ssm_a tensors (#28121)
Sigbjørn Skjæretf28493c78347
diff --git a/src/models/bailingmoe3.cpp b/src/models/bailingmoe3.cppindex 1583d9be34..5ebedaecbb 100644--- a/src/models/bailingmoe3.cpp+++ b/src/models/bailingmoe3.cpp@@ -86,7 +86,7 @@ void llama_model_bailingmoe3::load_arch_tensors(llama_model_loader & ml) {create_tensor_qkv(layer, il, n_embd, d_inner, d_inner, d_inner, trunk_flags);layer.ssm_f_a = create_tensor(tn(LLM_TENSOR_SSM_F_A, "weight", il), { n_embd, d_inner }, trunk_flags);layer.ssm_beta = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", il), { n_embd, n_head }, trunk_flags);- layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, il), { 1, n_head }, trunk_flags);+ layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, il), { 1, n_head }, trunk_flags);layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", il), { d_inner }, trunk_flags);layer.ssm_g_a = create_tensor(tn(LLM_TENSOR_SSM_G_A, "weight", il), { n_embd, d_inner }, trunk_flags);layer.ssm_o_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", il), { head_dim }, trunk_flags);diff --git a/src/models/kimi-k3.cpp b/src/models/kimi-k3.cppindex d952d72cdf..7b46bccdb0 100644--- a/src/models/kimi-k3.cpp+++ b/src/models/kimi-k3.cpp@@ -94,7 +94,7 @@ void llama_model_kimi_k3::load_arch_tensors(llama_model_loader &) {layer.ssm_beta = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", i), {n_embd, n_head}, 0);// K3's A_log is a plain 1-D [n_head] tensor (kimi-linear's is padded)- layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {n_head}, 0);+ layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, i), {n_head}, 0);layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {d_inner}, 0);// K3 uses a single full-rank gate instead of kimi-linear's g_a/g_b pairdiff --git a/src/models/kimi-linear.cpp b/src/models/kimi-linear.cppindex 367f6990d1..bda3cd9b04 100644--- a/src/models/kimi-linear.cpp+++ b/src/models/kimi-linear.cpp@@ -84,9 +84,9 @@ void llama_model_kimi_linear::load_arch_tensors(llama_model_loader &) {layer.ssm_beta = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", i), {n_embd, n_head}, 0);// A_log - Shape in GGUF: [1, num_heads, 1, 1] (4D) or [1, num_heads] (2D after quantization) Note: -exp(A_log) is applied in convert_hf_to_gguf.py- layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_head, 1, 1}, TENSOR_NOT_REQUIRED);+ layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, i), {1, n_head, 1, 1}, TENSOR_NOT_REQUIRED);if (!layer.ssm_a) {- layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_head}, 0);+ layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, i), {1, n_head}, 0);}// dt_bias - shape [n_embd_head_k_kda * n_head] = [4096]