Hugging Face's logo
AbstractPhila's picture
13 2 15

AbstractPhila PRO

AbstractPhil
Geoion99's profile picture mnemic's profile picture manh-linh's profile picture
·
https://civitai.com/user/AbstractPhila

AI & ML interests

datasets, research papers, experimentation, vision, classification, text encoders, tokenization, llms, diffusion, distillation, and more.

Recent Activity

posted an update about 12 hours ago
pytorch-parallel-compiler v0.5.0 upgrades: *Complex benchmarking for wide primitive objects is supported now. This includes multiple presets for quick tests on hardware. * All supported primitive either have validity checks or will have them. * 6 new wide layers supported directly, and will be a key part to the autotuner before v1.0 * WideTracedModel is a preliminary auto-builder so the user doesn't need to build them manually by gathering layers. https://github.com/AbstractEyes/pytorch-parallel-compiler New Layers for 0.5.0: WideGRU, WideLSTM, WideGroupNorm, WideMultiheadedAttention, WideInstancenorm1/2d, WideConv3d, Upcoming for 1.0: * WideTracedModel fully building any supported layer patterns with multiple autotune potentials for autoselection. * Module cherry-picking for use-case only; E.G. WideLinear replace only benefits your case 35% while Attention reduces by 10% no attn. * All (roughly 32 more) commonly used pytorch layer systems supported in one form or another with wide-batched kernels to benefit both eager and compiled, many of which require reworks or completely remaking them. * Autotuning wide formats based on hardware response to the kernels. Kernel chunking for big slow processes such as LSTM, kernel fusion for small process with excess overhead, expanding kernels with masking to fit specific use-case paradigms with hardwares, and a series of smaller and more important optimizations along the way. * Full transformer and rope support with wide-batched optimizations throughout the structures to allow more robust autoregression throughput. * Additional Conv1d, Conv2d, and Conv3d optimizations. >version 1.0 : * Entire diffusion structures specifically kernelized for high-efficiency utilization with eager and compilation. * Video diffusion specific targets meant to heavily reduce computation costs on the gpu and increase computation throughput on the gpu.
replied to their post 3 days ago
The Long: this is a proof of concept; ensemble compilation vmap prototype is functional and can be used to increase throughput for wider batches on FFN, MLP, LLM, or other models than just ensembles. This system traces your model and creates stages of functional activation. Based on the stage it will combine or remove combinations of stages meant to assign your layers to batched functional calls meant to put pressure on your GPU with less loops with directly curated cudagraph compliance where applicable. Identical weights yield identical results at the cost of hardware and vram. TLDR: This is an ensemble optimization adapted to standard models. This will yield high-capacity speed improvements through increased throughput for inference and training alike using carefully traced staged vmap structures. https://github.com/AbstractEyes/pytorch-parallel-compiler The early list of layers isn't fully represented yet, so this is a preliminary look into the potentials of this structure when fully fleshed out. MLP (N=100, batch=32, CUDA): ``` Eager: 2-3x speedup Compiled: 35-40x speedup ``` ResBlock (N=20, batch=8, CUDA): ``` Eager: ~5x speedup Compiled: ~10x speedup ``` This is early testing and so far the yields indicate that WIDENING your model with adjacent shared batched vmaps for uniformly staged models will yield considerably higher output for inference at the cost of additional hardware utilization. This is akin to lining up all your systems and uniformly passing the necessary implications through a shared frozen representation gate. Training for this is not tested nor supported yet, use at your own risk.
posted an update 3 days ago
The Long: this is a proof of concept; ensemble compilation vmap prototype is functional and can be used to increase throughput for wider batches on FFN, MLP, LLM, or other models than just ensembles. This system traces your model and creates stages of functional activation. Based on the stage it will combine or remove combinations of stages meant to assign your layers to batched functional calls meant to put pressure on your GPU with less loops with directly curated cudagraph compliance where applicable. Identical weights yield identical results at the cost of hardware and vram. TLDR: This is an ensemble optimization adapted to standard models. This will yield high-capacity speed improvements through increased throughput for inference and training alike using carefully traced staged vmap structures. https://github.com/AbstractEyes/pytorch-parallel-compiler The early list of layers isn't fully represented yet, so this is a preliminary look into the potentials of this structure when fully fleshed out. MLP (N=100, batch=32, CUDA): ``` Eager: 2-3x speedup Compiled: 35-40x speedup ``` ResBlock (N=20, batch=8, CUDA): ``` Eager: ~5x speedup Compiled: ~10x speedup ``` This is early testing and so far the yields indicate that WIDENING your model with adjacent shared batched vmaps for uniformly staged models will yield considerably higher output for inference at the cost of additional hardware utilization. This is akin to lining up all your systems and uniformly passing the necessary implications through a shared frozen representation gate. Training for this is not tested nor supported yet, use at your own risk.
View all activity

Organizations

DeepGHS's profile picture BangumiBase's profile picture Abstract Powered Research's profile picture