TensorRT-LLM Requests

[ncomly-nvidia]

Hi all, this issue will track the feature requests you've made to TensorRT-LLM & provide a place to see what TRT-LLM is currently working on.

Last update: Jan 14th, 2024
🚀 = in development

Models

Decoder Only

• 🚀 Zephyr-7B - Support for Zephyr 7B model #157
• DeciLM-7B - Support for other llm like Decilm? #853
• ChatGLM 3 - Support for ChatGLM3 plz #180, Support ChatGLM3 #270
• Mistral-7B - Mistral 7B support #49
• Mixtral-7B - Mixtral 8x 7B - MoE by Mistral AI #616

Encoder / Encoder-Decoder

• DeBERTa - Support for DeBerta  #174
• RoBERTa - [FEA] Support Roberta model #124
• 🚀 BART, mBART - Cross-attention returns wrong results #285, [Feature request] Support MBartForCausalLM Request!!! #360
• FLAN-T5 - How can I use flan-t5-base? #251, Cross-attention returns wrong results #285, [Feature request] Support soft_prompt or inputs_embeds? #310

Multi-Modal

• BLIP2 + T5 - [Feature request] Support soft_prompt or inputs_embeds? #310, [feature request]Blip2 T5 support request #531
• LLaVa - Is it even possible to have multiple input layers #641,
• Qwen-VL - Does the repo suport qwen-vl? #728
• Generic Vision Encoder + LLM Support - Is it even possible to have multiple input layers #641, [Feature request] Support soft_prompt or inputs_embeds? #310
• BLIP2
• Whisper - Support non LLM transformer networks #323

Other

• YaRN - [Feature Request] support YaRN request #792
• Expert Caching - [Feature Request] Mixtral Offloading #849
• LoRA - Llama 2 with LoRA #68
• Mixtral - Mixtral 8x 7B - MoE by Mistral AI #616

Features & Optimizations

• Context Chunking - [Feature request] Dynamic splitfuse from Deepspeed (2x throughput) #317
• Speculative Decoding - Feature: Speculative sampling / Assisted Generation #169, Smaller available space for paged KV cache compared with vLLM #224, Falcon-40b build causing memory leaks and failure #226
  implementation done - documentation in progress

KV Cache

• Reuse KV Cache - [Feature reuqest] support interactive-generation #292, Add automatic reuse of common key value cache blocks between requests #620
• Attention Sinks (StreamingLLM, H2O) - Attention sink #104

Quantization

• StarCoder INT8 SQ - Feature request: Support SmoothQuant variant of StarCoder #324
• Qwen INT4 - [Feature request] AutoAWQ support #345
• INT8 Weight only - Support weight only quantization from bfloat16 to int8? #110

Sampling

• 🚀 support frequnecy_penalty - Support for frequency_penalty #275
• Logit Manipulators - Add Transformers logits manipulators #241
• Combine repetition & presence penalties - Support for combining repetition_penalty, presence_penalty #274

Workflow

Front-ends

• OpenAI compatible API - Provide an interface similar to OpenAI API #334
• Flag for end-of-stream - Flag indicate end of stream #240
• Load from Buffer - GptManager add support for loading from buffer #144
• Paged KV Cache Utilization Metric - How to know the utility of paged kv cache ? #512
• Log Probabilities - Return log probabilities for tokens #238
• Return only new tokens - How to get the newly generated tokens only? #227

Integrations

• 🚀 LlamaIndex
• 🚀 LangChain
• Mojo - Question about a Mojo Integration #556

Usage / Installation

• pip install - waiting for pre-built wheel package #790,

Platform Support

• Jetson - Nvidia Jetson device Support #62, How can I running successful on jetson orin NX? #488, TensorRT installation in TRT-LLM #619
• V100, T4 MHA - Why FMHA is not supported in V100 and T4 #320

[teis-e]

Please add CohereAI!!

CohereForAI/c4ai-command-r-plus

[EwoutH]

Llama 3 would be great (both 8B and 70B): #1470

Maybe quantized to 8 or even 4 bit.

[StephennFernandes]

currently llama 3 throws a bunch of errors converting to TensorRT LLM

any ideal about the support for llama 3

[EwoutH]

Phi-3-mini should be amazing! Such a small 3.8B model could run quantized on many GPUs, with as little as 4GB VRAM.

• Paper: https://arxiv.org/abs/2404.14219
• Model weights: https://huggingface.co/collections/microsoft/phi-3-6626e15e9585a200d2d761e3

[oscarbg]

+1 for Phi-3

[user-0a]

+1 for Command R Plus!

CohereForAI/c4ai-command-r-plus