From a07d070b6c339564b533534b2667c62004fd2b43 Mon Sep 17 00:00:00 2001 From: oobabooga <112222186+oobabooga@users.noreply.github.com> Date: Mon, 24 Jul 2023 16:37:03 -0300 Subject: [PATCH] Add llama-2-70b GGML support (#3285) --- README.md | 2 ++ modules/llamacpp_hf.py | 2 ++ modules/llamacpp_model.py | 2 ++ modules/loaders.py | 4 ++++ modules/shared.py | 2 ++ modules/ui.py | 2 ++ requirements.txt | 8 ++++---- server.py | 2 ++ 8 files changed, 20 insertions(+), 4 deletions(-) diff --git a/README.md b/README.md index 225ed6c4..1a54b914 100644 --- a/README.md +++ b/README.md @@ -247,6 +247,8 @@ Optionally, you can use the following command-line flags: | `--n-gpu-layers N_GPU_LAYERS` | Number of layers to offload to the GPU. Only works if llama-cpp-python was compiled with BLAS. Set this to 1000000000 to offload all layers to the GPU. | | `--n_ctx N_CTX` | Size of the prompt context. | | `--llama_cpp_seed SEED` | Seed for llama-cpp models. Default 0 (random). | +| `--n_gqa N_GQA` | grouped-query attention. Must be 8 for llama2 70b. | +| `--rms_norm_eps RMS_NORM_EPS` | Must be 1e-5 for llama2 70b. | #### AutoGPTQ diff --git a/modules/llamacpp_hf.py b/modules/llamacpp_hf.py index 069e1007..6a4a0294 100644 --- a/modules/llamacpp_hf.py +++ b/modules/llamacpp_hf.py @@ -106,6 +106,8 @@ class LlamacppHF(PreTrainedModel): 'n_gpu_layers': shared.args.n_gpu_layers, 'rope_freq_base': 10000 * shared.args.alpha_value ** (64/63.), 'rope_freq_scale': 1.0 / shared.args.compress_pos_emb, + 'n_gqa': shared.args.n_gqa or None, + 'rms_norm_eps': shared.args.rms_norm_eps or None, 'logits_all': True, } diff --git a/modules/llamacpp_model.py b/modules/llamacpp_model.py index 048f9eac..0f9c3470 100644 --- a/modules/llamacpp_model.py +++ b/modules/llamacpp_model.py @@ -53,6 +53,8 @@ class LlamaCppModel: 'n_gpu_layers': shared.args.n_gpu_layers, 'rope_freq_base': 10000 * shared.args.alpha_value ** (64/63.), 'rope_freq_scale': 1.0 / shared.args.compress_pos_emb, + 'n_gqa': shared.args.n_gqa or None, + 'rms_norm_eps': shared.args.rms_norm_eps or None, } result.model = Llama(**params) diff --git a/modules/loaders.py b/modules/loaders.py index b760128f..c55cf0ff 100644 --- a/modules/loaders.py +++ b/modules/loaders.py @@ -30,6 +30,8 @@ loaders_and_params = { ], 'llama.cpp': [ 'n_ctx', + 'n_gqa', + 'rms_norm_eps', 'n_gpu_layers', 'n_batch', 'threads', @@ -42,6 +44,8 @@ loaders_and_params = { ], 'llamacpp_HF': [ 'n_ctx', + 'n_gqa', + 'rms_norm_eps', 'n_gpu_layers', 'n_batch', 'threads', diff --git a/modules/shared.py b/modules/shared.py index d805d30f..5249e170 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -127,6 +127,8 @@ parser.add_argument('--cache-capacity', type=str, help='Maximum cache capacity. parser.add_argument('--n-gpu-layers', type=int, default=0, help='Number of layers to offload to the GPU.') parser.add_argument('--n_ctx', type=int, default=2048, help='Size of the prompt context.') parser.add_argument('--llama_cpp_seed', type=int, default=0, help='Seed for llama-cpp models. Default 0 (random)') +parser.add_argument('--n_gqa', type=int, default=0, help='grouped-query attention. Must be 8 for llama2 70b.') +parser.add_argument('--rms_norm_eps', type=float, default=0, help='Must be 1e-5 for llama2 70b.') # GPTQ parser.add_argument('--wbits', type=int, default=0, help='Load a pre-quantized model with specified precision in bits. 2, 3, 4 and 8 are supported.') diff --git a/modules/ui.py b/modules/ui.py index 704be925..0e6cf0b1 100644 --- a/modules/ui.py +++ b/modules/ui.py @@ -61,6 +61,8 @@ def list_model_elements(): 'mlock', 'n_gpu_layers', 'n_ctx', + 'n_gqa', + 'rms_norm_eps', 'llama_cpp_seed', 'gpu_split', 'max_seq_len', diff --git a/requirements.txt b/requirements.txt index 7077f039..f085628a 100644 --- a/requirements.txt +++ b/requirements.txt @@ -27,8 +27,8 @@ https://github.com/PanQiWei/AutoGPTQ/releases/download/v0.3.0/auto_gptq-0.3.0+cu https://github.com/jllllll/exllama/releases/download/0.0.8/exllama-0.0.8+cu117-cp310-cp310-win_amd64.whl; platform_system == "Windows" https://github.com/jllllll/exllama/releases/download/0.0.8/exllama-0.0.8+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" # llama-cpp-python without GPU support -llama-cpp-python==0.1.74; platform_system != "Windows" -https://github.com/abetlen/llama-cpp-python/releases/download/v0.1.74/llama_cpp_python-0.1.74-cp310-cp310-win_amd64.whl; platform_system == "Windows" +llama-cpp-python==0.1.77; platform_system != "Windows" +https://github.com/abetlen/llama-cpp-python/releases/download/v0.1.77/llama_cpp_python-0.1.77-cp310-cp310-win_amd64.whl; platform_system == "Windows" # llama-cpp-python with CUDA support -https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_cuda-0.1.74+cu117-cp310-cp310-win_amd64.whl; platform_system == "Windows" -https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_cuda-0.1.74+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" +https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_cuda-0.1.77+cu117-cp310-cp310-win_amd64.whl; platform_system == "Windows" +https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_cuda-0.1.77+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" diff --git a/server.py b/server.py index babdec2d..97eb5509 100644 --- a/server.py +++ b/server.py @@ -219,6 +219,8 @@ def create_model_menus(): shared.gradio['n_batch'] = gr.Slider(label="n_batch", minimum=1, maximum=2048, value=shared.args.n_batch) shared.gradio['n_gpu_layers'] = gr.Slider(label="n-gpu-layers", minimum=0, maximum=128, value=shared.args.n_gpu_layers) shared.gradio['n_ctx'] = gr.Slider(minimum=0, maximum=16384, step=256, label="n_ctx", value=shared.args.n_ctx) + shared.gradio['n_gqa'] = gr.Slider(minimum=0, maximum=16, step=1, label="n_gqa", value=shared.args.n_gqa, info='grouped-query attention. Must be 8 for llama2 70b.') + shared.gradio['rms_norm_eps'] = gr.Slider(minimum=0, maximum=1e-5, step=1e-6, label="rms_norm_eps", value=shared.args.n_gqa, info='Must be 1e-5 for llama2 70b.') shared.gradio['wbits'] = gr.Dropdown(label="wbits", choices=["None", 1, 2, 3, 4, 8], value=str(shared.args.wbits) if shared.args.wbits > 0 else "None") shared.gradio['groupsize'] = gr.Dropdown(label="groupsize", choices=["None", 32, 64, 128, 1024], value=str(shared.args.groupsize) if shared.args.groupsize > 0 else "None") shared.gradio['model_type'] = gr.Dropdown(label="model_type", choices=["None", "llama", "opt", "gptj"], value=shared.args.model_type or "None")