Bump llama-cpp-python, +tensor_split by @shouyiwang, +mul_mat_q (#3610)

README.md

@@ -261,7 +261,9 @@ 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. |
+| `--mul_mat_q` | Activate new mulmat kernels. |
 | `--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,
@@ -110,9 +116,11 @@ class LlamacppHF(PreTrainedModel):
             'n_batch': shared.args.n_batch,
             'use_mmap': not shared.args.no_mmap,
             'use_mlock': shared.args.mlock,
+            'mul_mat_q': shared.args.mul_mat_q,
             '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,
@@ -63,9 +69,11 @@ class LlamaCppModel:
             'n_batch': shared.args.n_batch,
             'use_mmap': not shared.args.no_mmap,
             'use_mlock': shared.args.mlock,
+            'mul_mat_q': shared.args.mul_mat_q,
             '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

@@ -67,11 +67,13 @@ loaders_and_params = OrderedDict({
         'n_gqa',
         'rms_norm_eps',
         'n_gpu_layers',
+        'tensor_split',
         'n_batch',
         'threads',
         'no_mmap',
         'low_vram',
         'mlock',
+        'mul_mat_q',
         'llama_cpp_seed',
         'alpha_value',
         'compress_pos_emb',
@@ -82,11 +84,13 @@ loaders_and_params = OrderedDict({
         'n_gqa',
         'rms_norm_eps',
         'n_gpu_layers',
+        'tensor_split',
         'n_batch',
         'threads',
         'no_mmap',
         'low_vram',
         'mlock',
+        'mul_mat_q',
         'alpha_value',
         'compress_pos_emb',
         'cpu',

modules/shared.py

@@ -119,8 +119,10 @@ parser.add_argument('--n_batch', type=int, default=512, help='Maximum number of
 parser.add_argument('--no-mmap', action='store_true', help='Prevent mmap from being used.')
 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('--mul_mat_q', action='store_true', help='Activate new mulmat kernels.')
 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

@@ -68,7 +68,9 @@ def list_model_elements():
         'no_mmap',
         'low_vram',
         'mlock',
+        'mul_mat_q',
         'n_gpu_layers',
+        'tensor_split',
         'n_ctx',
         'n_gqa',
         'rms_norm_eps',

modules/ui_model_menu.py

@@ -33,6 +33,7 @@ def create_ui():
                 default_gpu_mem.append(int(re.sub('[a-zA-Z ]', '', i)))
             else:
                 default_gpu_mem.append(int(re.sub('[a-zA-Z ]', '', i)) * 1000)
+
     while len(default_gpu_mem) < len(total_mem):
         default_gpu_mem.append(0)
 
@@ -109,6 +110,8 @@ def create_ui():
                             shared.gradio['no_mmap'] = gr.Checkbox(label="no-mmap", value=shared.args.no_mmap)
                             shared.gradio['low_vram'] = gr.Checkbox(label="low-vram", value=shared.args.low_vram)
                             shared.gradio['mlock'] = gr.Checkbox(label="mlock", value=shared.args.mlock)
+                            shared.gradio['mul_mat_q'] = gr.Checkbox(label="mul_mat_q", value=shared.args.mul_mat_q)
+                            shared.gradio['tensor_split'] = gr.Textbox(label='tensor_split', info='Split the model across multiple GPUs, comma-separated list of proportions, e.g. 18,17')
                             shared.gradio['llama_cpp_seed'] = gr.Number(label='Seed (0 for random)', value=shared.args.llama_cpp_seed)
                             shared.gradio['trust_remote_code'] = gr.Checkbox(label="trust-remote-code", value=shared.args.trust_remote_code, info='Make sure to inspect the .py files inside the model folder before loading it with this option enabled.')
                             shared.gradio['gptq_for_llama_info'] = gr.Markdown('GPTQ-for-LLaMa support is currently only kept for compatibility with older GPUs. AutoGPTQ or ExLlama is preferred when compatible. GPTQ-for-LLaMa is installed by default with the webui on supported systems. Otherwise, it has to be installed manually following the instructions here: [instructions](https://github.com/oobabooga/text-generation-webui/blob/main/docs/GPTQ-models-(4-bit-mode).md#installation-1).')

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"
+llama-cpp-python==0.1.78; 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"
+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"