From 22399861677270bfb770f805fc559c9d767065ce Mon Sep 17 00:00:00 2001 From: Georg Kolling Date: Tue, 31 Oct 2023 16:12:36 +0000 Subject: [PATCH] Disable Cutlass LSTM codepath for 128 wide layers, as this kernel is broken --- dorado/nn/CRFModel.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/dorado/nn/CRFModel.cpp b/dorado/nn/CRFModel.cpp index 57202322..04f3f087 100644 --- a/dorado/nn/CRFModel.cpp +++ b/dorado/nn/CRFModel.cpp @@ -106,7 +106,7 @@ static LstmMode get_cuda_lstm_mode(int layer_idx, int layer_size) { bool is_TX2 = (prop->major == 6 && prop->minor == 2); bool is_A100_H100 = ((prop->major == 8 || prop->major == 9) && prop->minor == 0); - if (is_A100_H100 && layer_size <= 1024 && (layer_size % 128) == 0) { + if (is_A100_H100 && layer_size <= 1024 && layer_size > 128 && (layer_size % 128) == 0) { return (layer_idx == 0) ? LstmMode::CUTLASS_TNC_F16 : LstmMode::CUTLASS_TNC_I8; } else if (!is_TX2 && (layer_size == 96 || layer_size == 128)) { return LstmMode::QUANTISED_NTC;