Bump llama-cpp-python

README.md

@@ -262,6 +262,7 @@ Optionally, you can use the following command-line flags:
 | `--no-mmap` | Prevent mmap from being used. |
 | `--mlock`   | Force the system to keep the model in RAM. |
 | `--cache-capacity CACHE_CAPACITY`   | Maximum cache capacity. Examples: 2000MiB, 2GiB. When provided without units, bytes will be assumed. |
+| `--tensor_split TENSOR_SPLIT` | Split the model across multiple GPUs, comma-separated list of proportions, e.g. 18,17 |
 | `--llama_cpp_seed SEED` | Seed for llama-cpp models. Default 0 (random). |
 | `--n_gqa N_GQA`         | grouped-query attention. Must be 8 for llama-2 70b. |
 | `--rms_norm_eps RMS_NORM_EPS`  | 5e-6 is a good value for llama-2 models. |

modules/llamacpp_hf.py

@@ -102,6 +102,12 @@ class LlamacppHF(PreTrainedModel):
             model_file = list(path.glob('*ggml*.bin'))[0]
 
         logger.info(f"llama.cpp weights detected: {model_file}\n")
+
+        if shared.args.tensor_split is None or shared.args.tensor_split.strip() == '':
+            tensor_split_list = None
+        else:
+            tensor_split_list = [float(x) for x in shared.args.tensor_split.strip().split(",")]
+
         params = {
             'model_path': str(model_file),
             'n_ctx': shared.args.n_ctx,
@@ -113,6 +119,7 @@ class LlamacppHF(PreTrainedModel):
             'low_vram': shared.args.low_vram,
             'n_gpu_layers': shared.args.n_gpu_layers,
             'rope_freq_base': 10000 * shared.args.alpha_value ** (64 / 63.),
+            'tensor_split': tensor_split_list,
             '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,

modules/llamacpp_model.py

@@ -55,6 +55,12 @@ class LlamaCppModel:
                 cache_capacity = int(shared.args.cache_capacity)
 
         logger.info("Cache capacity is " + str(cache_capacity) + " bytes")
+
+        if shared.args.tensor_split is None or shared.args.tensor_split.strip() == '':
+            tensor_split_list = None
+        else:
+            tensor_split_list = [float(x) for x in shared.args.tensor_split.strip().split(",")]
+
         params = {
             'model_path': str(path),
             'n_ctx': shared.args.n_ctx,
@@ -66,6 +72,7 @@ class LlamaCppModel:
             'low_vram': shared.args.low_vram,
             'n_gpu_layers': shared.args.n_gpu_layers,
             'rope_freq_base': 10000 * shared.args.alpha_value ** (64 / 63.),
+            'tensor_split': tensor_split_list,
             '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,

modules/loaders.py

@@ -67,6 +67,7 @@ loaders_and_params = OrderedDict({
         'n_gqa',
         'rms_norm_eps',
         'n_gpu_layers',
+        'tensor_split',
         'n_batch',
         'threads',
         'no_mmap',
@@ -82,6 +83,7 @@ loaders_and_params = OrderedDict({
         'n_gqa',
         'rms_norm_eps',
         'n_gpu_layers',
+        'tensor_split',
         'n_batch',
         'threads',
         'no_mmap',

modules/shared.py

@@ -121,6 +121,7 @@ parser.add_argument('--low-vram', action='store_true', help='Low VRAM Mode')
 parser.add_argument('--mlock', action='store_true', help='Force the system to keep the model in RAM.')
 parser.add_argument('--cache-capacity', type=str, help='Maximum cache capacity. Examples: 2000MiB, 2GiB. When provided without units, bytes will be assumed.')
 parser.add_argument('--n-gpu-layers', type=int, default=0, help='Number of layers to offload to the GPU.')
+parser.add_argument('--tensor_split', type=str, default=None, help="Split the model across multiple GPUs, comma-separated list of proportions, e.g. 18,17")
 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 llama-2 70b.')

modules/ui.py

@@ -69,6 +69,7 @@ def list_model_elements():
         'low_vram',
         'mlock',
         'n_gpu_layers',
+        'tensor_split',
         'n_ctx',
         'n_gqa',
         'rms_norm_eps',

requirements.txt

@@ -31,8 +31,8 @@ https://github.com/jllllll/exllama/releases/download/0.0.10/exllama-0.0.10+cu117
 https://github.com/jllllll/exllama/releases/download/0.0.10/exllama-0.0.10+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
 
 # llama-cpp-python without GPU support
-llama-cpp-python==0.1.77; 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==0.1.78; platform_system != "Windows"
+https://github.com/abetlen/llama-cpp-python/releases/download/v0.1.78/llama_cpp_python-0.1.78-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-0.1.77+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"