Merge pull request #5970 from oobabooga/dev

Merge dev branch

.github/FUNDING.yml

@@ -1 +1,2 @@
+github: oobabooga
 ko_fi: oobabooga

README.md

@@ -107,13 +107,16 @@ pip install -r <requirements file according to table below>
 
 Requirements file to use:
 
-| GPU | requirements file to use |
-|--------|---------|
-| NVIDIA | `requirements.txt` |
-| AMD | `requirements_amd.txt` |
-| CPU only | `requirements_cpu_only.txt` |
-| Apple Intel | `requirements_apple_intel.txt` |
-| Apple Silicon | `requirements_apple_silicon.txt` |
+| GPU | CPU | requirements file to use |
+|--------|---------|---------|
+| NVIDIA | has AVX2 | `requirements.txt` |
+| NVIDIA | no AVX2 | `requirements_noavx2.txt` |
+| AMD | has AVX2 | `requirements_amd.txt` |
+| AMD | no AVX2 | `requirements_amd_noavx2.txt` |
+| CPU only | has AVX2 | `requirements_cpu_only.txt` |
+| CPU only | no AVX2 | `requirements_cpu_only_noavx2.txt` |
+| Apple | Intel | `requirements_apple_intel.txt` |
+| Apple | Apple Silicon | `requirements_apple_silicon.txt` |
 
 ### Start the web UI
 
@@ -129,7 +132,7 @@ Then browse to
 
 ##### AMD GPU on Windows
 
-1) Use `requirements_cpu_only.txt` in the command above.
+1) Use `requirements_cpu_only.txt` or `requirements_cpu_only_noavx2.txt` in the command above.
 
 2) Manually install llama-cpp-python using the appropriate command for your hardware: [Installation from PyPI](https://github.com/abetlen/llama-cpp-python#installation-with-hardware-acceleration).
     * Use the `LLAMA_HIPBLAS=on` toggle.
@@ -252,6 +255,7 @@ List of command-line flags
 
 | Flag        | Description |
 |-------------|-------------|
+| `--tensorcores`  | Use llama-cpp-python compiled with tensor cores support. This increases performance on RTX cards. NVIDIA only. |
 | `--n_ctx N_CTX` | Size of the prompt context. |
 | `--threads` | Number of threads to use. |
 | `--threads-batch THREADS_BATCH` | Number of threads to use for batches/prompt processing. |
@@ -421,3 +425,9 @@ If you would like to contribute to the project, check out the [Contributing guid
 ## Acknowledgment
 
 In August 2023, [Andreessen Horowitz](https://a16z.com/) (a16z) provided a generous grant to encourage and support my independent work on this project. I am **extremely** grateful for their trust and recognition.
+
+## GitHub Sponsors
+
+The following is a list of top-tier sponsors for this project here on GitHub:
+
+* Be the first one! Visit https://github.com/sponsors/oobabooga/.

docs/04 - Model Tab.md

@@ -21,7 +21,7 @@ Options:
 * **alpha_value**: Used to extend the context length of a model with a minor loss in quality. I have measured 1.75 to be optimal for 1.5x context, and 2.5 for 2x context. That is, with alpha = 2.5 you can make a model with 4096 context length go to 8192 context length.
 * **rope_freq_base**: Originally another way to write "alpha_value", it ended up becoming a necessary parameter for some models like CodeLlama, which was fine-tuned with this set to 1000000 and hence needs to be loaded with it set to 1000000 as well.
 * **compress_pos_emb**: The first and original context-length extension method, discovered by [kaiokendev](https://kaiokendev.github.io/til). When set to 2, the context length is doubled, 3 and it's tripled, etc. It should only be used for models that have been fine-tuned with this parameter set to different than 1. For models that have not been tuned to have greater context length, alpha_value will lead to a smaller accuracy loss.
-* **cpu**: Loads the model in CPU mode using Pytorch. The model will be loaded in 32-bit precision, so a lot of RAM will be used. CPU inference with transformers is older than llama.cpp and it works, but it's a lot slower.
+* **cpu**: Loads the model in CPU mode using Pytorch. The model will be loaded in 32-bit precision, so a lot of RAM will be used. CPU inference with transformers is older than llama.cpp and it works, but it's a lot slower. Note: this parameter has a different interpretation in the llama.cpp loader (see below).
 * **load-in-8bit**: Load the model in 8-bit precision using bitsandbytes. The 8-bit kernel in that library has been optimized for training and not inference, so load-in-8bit is slower than load-in-4bit (but more accurate).
 * **bf16**: Use bfloat16 precision instead of float16 (the default). Only applies when quantization is not used.
 * **auto-devices**: When checked, the backend will try to guess a reasonable value for "gpu-memory" to allow you to load a model with CPU offloading. I recommend just setting "gpu-memory" manually instead. This parameter is also needed for loading GPTQ models, in which case it needs to be checked before loading the model.
@@ -84,7 +84,9 @@ Example: https://huggingface.co/TheBloke/Llama-2-7b-Chat-GGUF
 * **n_batch**: Batch size for prompt processing. Higher values are supposed to make generation faster, but I have never obtained any benefit from changing this value.
 * **threads**: Number of threads. Recommended value: your number of physical cores. 
 * **threads_batch**: Number of threads for batch processing. Recommended value: your total number of cores (physical + virtual).
+* **tensorcores**: Use llama.cpp compiled with "tensor cores" support, which improves performance on NVIDIA RTX cards in most cases.
 * **streamingllm**: Experimental feature to avoid re-evaluating the entire prompt when part of it is removed, for instance, when you hit the context length for the model in chat mode and an old message is removed.
+* **cpu**: Force a version of llama.cpp compiled without GPU acceleration to be used. Can usually be ignored. Only set this if you want to use CPU only and llama.cpp doesn't work otherwise. 
 * **no_mul_mat_q**: Disable the mul_mat_q kernel. This kernel usually improves generation speed significantly. This option to disable it is included in case it doesn't work on some system.
 * **no-mmap**: Loads the model into memory at once, possibly preventing I/O operations later on at the cost of a longer load time.
 * **mlock**: Force the system to keep the model in RAM rather than swapping or compressing (no idea what this means, never used it).

modules/llama_cpp_python_hijack.py

@@ -1,11 +1,25 @@
 from typing import Sequence
 
-import llama_cpp
 from tqdm import tqdm
 
 from modules import shared
 from modules.cache_utils import process_llamacpp_cache
 
+try:
+    import llama_cpp
+except:
+    llama_cpp = None
+
+try:
+    import llama_cpp_cuda
+except:
+    llama_cpp_cuda = None
+
+try:
+    import llama_cpp_cuda_tensorcores
+except:
+    llama_cpp_cuda_tensorcores = None
+
 
 def eval_with_progress(self, tokens: Sequence[int]):
     """
@@ -67,7 +81,7 @@ def my_generate(self, *args, **kwargs):
     lib.Llama.generate = my_generate
 
 
-for lib in [llama_cpp]:
+for lib in [llama_cpp, llama_cpp_cuda, llama_cpp_cuda_tensorcores]:
     if lib is not None:
         lib.Llama.eval = eval_with_progress
         monkey_patch_generate(lib)

modules/llamacpp_hf.py

@@ -2,7 +2,6 @@
 from pathlib import Path
 from typing import Any, Dict, Optional, Union
 
-import llama_cpp
 import torch
 from torch.nn import CrossEntropyLoss
 from transformers import GenerationConfig, PretrainedConfig, PreTrainedModel
@@ -11,6 +10,32 @@
 from modules import RoPE, llama_cpp_python_hijack, shared
 from modules.logging_colors import logger
 
+try:
+    import llama_cpp
+except:
+    llama_cpp = None
+
+try:
+    import llama_cpp_cuda
+except:
+    llama_cpp_cuda = None
+
+try:
+    import llama_cpp_cuda_tensorcores
+except:
+    llama_cpp_cuda_tensorcores = None
+
+
+def llama_cpp_lib():
+    if shared.args.cpu and llama_cpp is not None:
+        return llama_cpp
+    elif shared.args.tensorcores and llama_cpp_cuda_tensorcores is not None:
+        return llama_cpp_cuda_tensorcores
+    elif llama_cpp_cuda is not None:
+        return llama_cpp_cuda
+    else:
+        return llama_cpp
+
 
 class LlamacppHF(PreTrainedModel):
     def __init__(self, model, path):
@@ -32,7 +57,7 @@ def __init__(self, model, path):
                 'n_tokens': self.model.n_tokens,
                 'input_ids': self.model.input_ids.copy(),
                 'scores': self.model.scores.copy(),
-                'ctx': llama_cpp.llama_new_context_with_model(model.model, model.context_params)
+                'ctx': llama_cpp_lib().llama_new_context_with_model(model.model, model.context_params)
             }
 
     def _validate_model_class(self):
@@ -195,7 +220,7 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
             'split_mode': 1 if not shared.args.row_split else 2
         }
 
-        Llama = llama_cpp.Llama
+        Llama = llama_cpp_lib().Llama
         model = Llama(**params)
 
         return LlamacppHF(model, model_file)

modules/llamacpp_model.py

@@ -1,7 +1,6 @@
 import re
 from functools import partial
 
-import llama_cpp
 import numpy as np
 import torch
 
@@ -10,6 +9,32 @@
 from modules.logging_colors import logger
 from modules.text_generation import get_max_prompt_length
 
+try:
+    import llama_cpp
+except:
+    llama_cpp = None
+
+try:
+    import llama_cpp_cuda
+except:
+    llama_cpp_cuda = None
+
+try:
+    import llama_cpp_cuda_tensorcores
+except:
+    llama_cpp_cuda_tensorcores = None
+
+
+def llama_cpp_lib():
+    if shared.args.cpu and llama_cpp is not None:
+        return llama_cpp
+    elif shared.args.tensorcores and llama_cpp_cuda_tensorcores is not None:
+        return llama_cpp_cuda_tensorcores
+    elif llama_cpp_cuda is not None:
+        return llama_cpp_cuda
+    else:
+        return llama_cpp
+
 
 def ban_eos_logits_processor(eos_token, input_ids, logits):
     logits[eos_token] = -float('inf')
@@ -35,8 +60,8 @@ def __del__(self):
     @classmethod
     def from_pretrained(self, path):
 
-        Llama = llama_cpp.Llama
-        LlamaCache = llama_cpp.LlamaCache
+        Llama = llama_cpp_lib().Llama
+        LlamaCache = llama_cpp_lib().LlamaCache
 
         result = self()
         cache_capacity = 0
@@ -101,12 +126,12 @@ def load_grammar(self, string):
         if string != self.grammar_string:
             self.grammar_string = string
             if string.strip() != '':
-                self.grammar = llama_cpp.LlamaGrammar.from_string(string)
+                self.grammar = llama_cpp_lib().LlamaGrammar.from_string(string)
             else:
                 self.grammar = None
 
     def generate(self, prompt, state, callback=None):
-        LogitsProcessorList = llama_cpp.LogitsProcessorList
+        LogitsProcessorList = llama_cpp_lib().LogitsProcessorList
         prompt = prompt if type(prompt) is str else prompt.decode()
 
         # Handle truncation

modules/loaders.py

@@ -41,9 +41,11 @@
         'alpha_value',
         'rope_freq_base',
         'compress_pos_emb',
+        'cpu',
         'numa',
         'no_offload_kqv',
         'row_split',
+        'tensorcores',
         'streaming_llm',
         'attention_sink_size',
     ],
@@ -60,13 +62,15 @@
         'alpha_value',
         'rope_freq_base',
         'compress_pos_emb',
+        'cpu',
         'numa',
         'cfg_cache',
         'trust_remote_code',
         'no_use_fast',
         'logits_all',
         'no_offload_kqv',
         'row_split',
+        'tensorcores',
         'streaming_llm',
         'attention_sink_size',
         'llamacpp_HF_info',

modules/shared.py

@@ -114,6 +114,7 @@
 
 # llama.cpp
 group = parser.add_argument_group('llama.cpp')
+group.add_argument('--tensorcores', action='store_true', help='Use llama-cpp-python compiled with tensor cores support. This increases performance on RTX cards. NVIDIA only.')
 group.add_argument('--n_ctx', type=int, default=2048, help='Size of the prompt context.')
 group.add_argument('--threads', type=int, default=0, help='Number of threads to use.')
 group.add_argument('--threads-batch', type=int, default=0, help='Number of threads to use for batches/prompt processing.')
@@ -204,8 +205,7 @@
 group.add_argument('--multimodal-pipeline', type=str, default=None, help='The multimodal pipeline to use. Examples: llava-7b, llava-13b.')
 
 # Deprecated parameters
-group = parser.add_argument_group('Deprecated')
-group.add_argument('--tensorcores', action='store_true', help='DEPRECATED')
+# group = parser.add_argument_group('Deprecated')
 
 args = parser.parse_args()
 args_defaults = parser.parse_args([])
@@ -215,7 +215,7 @@
     if hasattr(args, arg):
         provided_arguments.append(arg)
 
-deprecated_args = ['tensorcores']
+deprecated_args = []
 
 
 def do_cmd_flags_warnings():

modules/ui.py

@@ -103,6 +103,7 @@ def list_model_elements():
         'logits_all',
         'no_offload_kqv',
         'row_split',
+        'tensorcores',
         'streaming_llm',
         'attention_sink_size',
         'hqq_backend',

modules/ui_model_menu.py

@@ -119,6 +119,7 @@ def create_ui():
                             shared.gradio['use_double_quant'] = gr.Checkbox(label="use_double_quant", value=shared.args.use_double_quant)
                             shared.gradio['use_flash_attention_2'] = gr.Checkbox(label="use_flash_attention_2", value=shared.args.use_flash_attention_2, info='Set use_flash_attention_2=True while loading the model.')
                             shared.gradio['auto_devices'] = gr.Checkbox(label="auto-devices", value=shared.args.auto_devices)
+                            shared.gradio['tensorcores'] = gr.Checkbox(label="tensorcores", value=shared.args.tensorcores, info='NVIDIA only: use llama-cpp-python compiled with tensor cores support. This increases performance on RTX cards.')
                             shared.gradio['streaming_llm'] = gr.Checkbox(label="streaming_llm", value=shared.args.streaming_llm, info='(experimental) Activate StreamingLLM to avoid re-evaluating the entire prompt when old messages are removed.')
                             shared.gradio['attention_sink_size'] = gr.Number(label="attention_sink_size", value=shared.args.attention_sink_size, precision=0, info='StreamingLLM: number of sink tokens. Only used if the trimmed prompt doesn\'t share a prefix with the old prompt.')
                             shared.gradio['cpu'] = gr.Checkbox(label="cpu", value=shared.args.cpu, info='llama.cpp: Use llama-cpp-python compiled without GPU acceleration. Transformers: use PyTorch in CPU mode.')