Document StreamingLLM

README.md

@@ -269,6 +269,9 @@ List of command-line flags
 | `--logits_all`| Needs to be set for perplexity evaluation to work. Otherwise, ignore it, as it makes prompt processing slower. |
 | `--no_offload_kqv` | Do not offload the K, Q, V to the GPU. This saves VRAM but reduces the performance. |
 | `--cache-capacity CACHE_CAPACITY`   | Maximum cache capacity (llama-cpp-python). Examples: 2000MiB, 2GiB. When provided without units, bytes will be assumed. |
+| `--row_split`                               | Split the model by rows across GPUs. This may improve multi-gpu performance. |
+| `--streaming-llm`                           | Activate StreamingLLM to avoid re-evaluating the entire prompt when old messages are removed. |
+| `--attention-sink-size ATTENTION_SINK_SIZE` | StreamingLLM: number of sink tokens. Only used if the trimmed prompt doesn't share a prefix with the old prompt. |
 
 #### ExLlamav2
 

docs/04 - Model Tab.md

@@ -80,16 +80,17 @@ Example: https://huggingface.co/TheBloke/Llama-2-7b-Chat-GGUF
 
 * **n-gpu-layers**: The number of layers to allocate to the GPU. If set to 0, only the CPU will be used. If you want to offload all layers, you can simply set this to the maximum value.
 * **n_ctx**: Context length of the model. In llama.cpp, the cache is preallocated, so the higher this value, the higher the VRAM. It is automatically set to the maximum sequence length for the model based on the metadata inside the GGUF file, but you may need to lower this value be able to fit the model into your GPU. After loading the model, the "Truncate the prompt up to this length" parameter under "Parameters" > "Generation" is automatically set to your chosen "n_ctx" so that you don't have to set the same thing twice.
+* **tensor_split**: For multi-gpu only. Sets the amount of memory to allocate per GPU as proportions. Not to be confused with other loaders where this is set in GB; here you can set something like `30,70` for 30%/70%.
+* **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).
-* **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.
+* **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).
 * **numa**: May improve performance on certain multi-cpu systems.
-* **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. 
-* **tensor_split**: For multi-gpu only. Sets the amount of memory to allocate per GPU.
-* **Seed**: The seed for the llama.cpp random number generator. Not very useful as it can only be set once (that I'm aware).
 
 ### llamacpp_HF
 

modules/shared.py

@@ -130,8 +130,8 @@ group.add_argument('--logits_all', action='store_true', help='Needs to be set fo
 group.add_argument('--no_offload_kqv', action='store_true', help='Do not offload the  K, Q, V to the GPU. This saves VRAM but reduces the performance.')
 group.add_argument('--cache-capacity', type=str, help='Maximum cache capacity (llama-cpp-python). Examples: 2000MiB, 2GiB. When provided without units, bytes will be assumed.')
 group.add_argument('--row_split', action='store_true', help='Split the model by rows across GPUs. This may improve multi-gpu performance.')
-group.add_argument('--streaming-llm', action='store_true', help='Activates StreamingLLM, which prevents the prompt from ever being reevaluated when old chat messages are removed due to the context length for the model being reached.')
+group.add_argument('--streaming-llm', action='store_true', help='Activate StreamingLLM to avoid re-evaluating the entire prompt when old messages are removed.')
-group.add_argument('--attention-sink-size', type=int, default=5, help='Minimum attention sink length from StreamingLLM.')
+group.add_argument('--attention-sink-size', type=int, default=5, help='StreamingLLM: number of sink tokens. Only used if the trimmed prompt does not share a prefix with the old prompt.')
 
 # ExLlamaV2
 group = parser.add_argument_group('ExLlamaV2')

modules/ui_model_menu.py

@@ -118,7 +118,7 @@ def create_ui():
                             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)
+                            shared.gradio['attention_sink_size'] = gr.Number(label="attention_sink_size", value=shared.args.attention_sink_size, 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.')
                             shared.gradio['row_split'] = gr.Checkbox(label="row_split", value=shared.args.row_split, info='Split the model by rows across GPUs. This may improve multi-gpu performance.')
                             shared.gradio['no_offload_kqv'] = gr.Checkbox(label="no_offload_kqv", value=shared.args.no_offload_kqv, info='Do not offload the  K, Q, V to the GPU. This saves VRAM but reduces the performance.')