From cbd36aa4d16cd253b268b4684fc359f66c5f0660 Mon Sep 17 00:00:00 2001 From: OlivierDehaene Date: Mon, 13 Mar 2023 22:57:08 +0100 Subject: [PATCH] fix(server): revert gpt-neox optims (#123) --- router/src/infer.rs | 6 +++++- server/Makefile | 2 +- 2 files changed, 6 insertions(+), 2 deletions(-) diff --git a/router/src/infer.rs b/router/src/infer.rs index 5955faec..e11f4fe6 100644 --- a/router/src/infer.rs +++ b/router/src/infer.rs @@ -229,7 +229,11 @@ async fn batching_task( shared: Arc, ) { // Minimum batch size after which we try to add more requests - let limit_min_batch_size = (max_batch_size / 2) as u32; + let limit_min_batch_size = if max_batch_size > 1 { + (max_batch_size / 2) as u32 + } else { + 0 + }; // Infinite loop loop { diff --git a/server/Makefile b/server/Makefile index 3de13e8e..44cf7c5c 100644 --- a/server/Makefile +++ b/server/Makefile @@ -1,4 +1,4 @@ -transformers_commit := 2f87dca1ca3e5663d0637da9bb037a6956e57a5e +transformers_commit := 517563354a3226ecfc3dca6e7a38012668d7156a gen-server: # Compile protos