llama : support attention bias on LLaMA architecture (#4283)

* Support attention_bias on LLaMA architecture

QKVO bias, should fix InternLM (https://github.com/ggerganov/llama.cpp/issues/3133) and works for LLaMAfied Qwen models (https://github.com/ggerganov/llama.cpp/pull/3743#issuecomment-1825923608).

* check existence of qkvo bias while loading llama models

Tested on LLaMA2, CUDA and CPU.

* Update llama.cpp

diff --git a/llama.cpp b/llama.cpp
index ca21cffab34f2a9126bf29051e9230fecb878666..15e52ad36a313325c7d036d7f740ef1a94396916 100644 (file)
--- a/llama.cpp
+++ b/llama.cpp
@@ -1266,6 +1266,9 @@ struct llama_layer {
     struct ggml_tensor * wqkv;
 
     // attention bias
+    struct ggml_tensor * bq;
+    struct ggml_tensor * bk;
+    struct ggml_tensor * bv;
     struct ggml_tensor * bo;
     struct ggml_tensor * bqkv;
 
@@ -2809,6 +2812,30 @@ static void llm_load_tensors(
                         layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, backend_split);
                         layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd},     backend_split);
 
+                        try {
+                            layer.bq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, backend);
+                        } catch (const std::runtime_error& e) {
+                            if (std::string(e.what()).find("not found") != std::string::npos) layer.bq = NULL; else throw;
+                        }
+
+                        try {
+                            layer.bk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, backend);
+                        } catch (const std::runtime_error& e) {
+                            if (std::string(e.what()).find("not found") != std::string::npos) layer.bk = NULL; else throw;
+                        }
+
+                        try {
+                            layer.bv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, backend);
+                        } catch (const std::runtime_error& e) {
+                            if (std::string(e.what()).find("not found") != std::string::npos) layer.bv = NULL; else throw;
+                        }
+
+                        try {
+                            layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend);
+                        } catch (const std::runtime_error& e) {
+                            if (std::string(e.what()).find("not found") != std::string::npos) layer.bo = NULL; else throw;
+                        }
+
                         layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
 
                         layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, backend_split);
@@ -2817,9 +2844,14 @@ static void llm_load_tensors(
 
                         if (backend == GGML_BACKEND_GPU) {
                             vram_weights +=
-                                ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq)       + ggml_nbytes(layer.wk)       +
-                                ggml_nbytes(layer.wv)        + ggml_nbytes(layer.wo)       + ggml_nbytes(layer.ffn_norm) +
-                                ggml_nbytes(layer.ffn_gate)  + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up);
+                                ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) +
+                                ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) +
+                                (layer.bq ? ggml_nbytes(layer.bq) : 0) +
+                                (layer.bk ? ggml_nbytes(layer.bk) : 0) +
+                                (layer.bv ? ggml_nbytes(layer.bv) : 0) +
+                                (layer.bo ? ggml_nbytes(layer.bo) : 0) +
+                                ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_gate) +
+                                ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up);
                         }
                     }
                 } break;
@@ -3983,12 +4015,24 @@ struct llm_build_context {
                 // compute Q and K and RoPE them
                 struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
                 cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
 
                 struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
                 cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
 
                 struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
                 cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
 
                 Qcur = ggml_rope_custom(
                     ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
@@ -4007,7 +4051,7 @@ struct llm_build_context {
                 llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
 
                 cur = llm_build_kqv(ctx0, hparams, kv_self,
-                        model.layers[il].wo, NULL,
+                        model.layers[il].wo, model.layers[il].bo,
                         Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
                 cb(cur, "kqv_out", il);
             }