mirror of
https://github.com/huggingface/text-generation-inference.git
synced 2025-04-20 14:22:08 +00:00
cfd4fbb479
* Add llamacpp backend Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Get rid of llama_batch_get_one() Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Use max_batch_total_tokens Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Handle max_batch_size Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Add some input validation checks Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Handle ctx args & fix sampling Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Add GPU args Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Add --defrag-threshold Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Add a stupid batch mechanism Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Cleanup Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Add --numa Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Fix args Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Enable flash attention by default Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Add --offload-kqv Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Fix batch_pos Signed-off-by: Adrien Gallouët <angt@huggingface.co> * backend(llama): add CUDA Dockerfile_llamacpp for now * Only export the latest logits Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Output real logprobs Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Fix batching Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Fix seq iterations Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Auto-detect n_threads when not provided Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Clear request cache after completion Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Remove warmup Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Cleanup Signed-off-by: Adrien Gallouët <angt@huggingface.co> * backend(llama): add CUDA architectures build argument for Dockerfile * Add specific args for batch Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Add --type-v & --type-k Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Bump llamacpp to b4623 Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Disable graceful shutdown in debug mode Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Update Dockerfile_llamacpp Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Cleanup Dockerfile Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Update Cargo.lock Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Update args Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Simplify batching logic Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Set TGI_LLAMA_PKG_CUDA from CUDA_VERSION Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Rename bindings Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Remove n_ctx Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Make max_batch_total_tokens optional Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Ensure all samplers are freed on error Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Initialize penalty_last_n with llamacpp default value Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Cleanup Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Improve default settings Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Add doc Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Update docs Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Thanks clippy Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Thanks cargo fmt Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Update docs Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Do not use HOSTNAME env Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Bump llama.cpp & cuda Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Fix requirements.txt Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Fix fmt Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Enable KQV offload by default Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Remove Ngrok tunneling Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Remove .cargo/config.toml Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Fix Dockerfile Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Add missing cuda prefix Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Handle custom llama.cpp dir Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Cleanup Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Add README.md Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Add HF transfer Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Fix bool args Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Update doc Signed-off-by: Adrien Gallouët <angt@huggingface.co> * Update doc Signed-off-by: Adrien Gallouët <angt@huggingface.co> --------- Signed-off-by: Adrien Gallouët <angt@huggingface.co> Co-authored-by: Morgan Funtowicz <funtowiczmo@gmail.com>
1.2 KiB
1.2 KiB
Multi-backend support
TGI (Text Generation Inference) offers flexibility by supporting multiple backends for serving large language models (LLMs). With multi-backend support, you can choose the backend that best suits your needs, whether you prioritize performance, ease of use, or compatibility with specific hardware. API interaction with TGI remains consistent across backends, allowing you to switch between them seamlessly.
Supported backends:
- TGI CUDA backend: This high-performance backend is optimized for NVIDIA GPUs and serves as the default option within TGI. Developed in-house, it boasts numerous optimizations and is used in production by various projects, including those by Hugging Face.
- TGI TRTLLM backend: This backend leverages NVIDIA's TensorRT library to accelerate LLM inference. It utilizes specialized optimizations and custom kernels for enhanced performance. However, it requires a model-specific compilation step for each GPU architecture.
- TGI Llamacpp backend: This backend facilitates the deployment of large language models (LLMs) by integrating [llama.cpp][llama.cpp], an advanced inference engine optimized for both CPU and GPU computation.