From 3f460a2b723c8b936ac29ecfd02f244b3adeba55 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 25 Aug 2023 11:55:59 +0300 Subject: cuda : add RoPE kernel for mode == 2 (NeoX) (#2760) * cuda : add RoPE kernel for mode == 2 (NeoX) * falcon : do not offload the embeddings layer --- llama.cpp | 22 +++++++++++++++++++++- 1 file changed, 21 insertions(+), 1 deletion(-) (limited to 'llama.cpp') diff --git a/llama.cpp b/llama.cpp index 67319396..52ba31d7 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1958,6 +1958,14 @@ static void llm_load_tensors( model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + vram_weights += ggml_nbytes(model.output_norm_b); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } } const uint32_t n_ff = hparams.n_ff; @@ -1967,7 +1975,7 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT + const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT auto & layer = model.layers[i]; @@ -1978,6 +1986,11 @@ static void llm_load_tensors( if (gguf_find_tensor(ml.ctx_gguf, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i).c_str()) >= 0) { layer.attn_norm_2 = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, backend); layer.attn_norm_2_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, backend); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(layer.attn_norm_2); + vram_weights += ggml_nbytes(layer.attn_norm_2_b); + } } layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); @@ -1985,6 +1998,13 @@ static void llm_load_tensors( layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += + ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) + + ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.wo) + + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3); + } } } break; default: -- cgit v1.2.3