From 703475f6c2771600acc27eba76f6a750f54aae50 Mon Sep 17 00:00:00 2001 From: Tyler Michael Smith Date: Fri, 14 Jun 2024 12:30:15 -0400 Subject: [PATCH] [Kernel] Fix CUTLASS 3.x custom broadcast load epilogue (#5516) --- .../quantization/cutlass_w8a8/broadcast_load_epilogue_c3x.hpp | 2 +- vllm/model_executor/layers/quantization/fp8.py | 4 +--- 2 files changed, 2 insertions(+), 4 deletions(-) diff --git a/csrc/quantization/cutlass_w8a8/broadcast_load_epilogue_c3x.hpp b/csrc/quantization/cutlass_w8a8/broadcast_load_epilogue_c3x.hpp index 8f38bbf507901..877a9f5b9e5de 100644 --- a/csrc/quantization/cutlass_w8a8/broadcast_load_epilogue_c3x.hpp +++ b/csrc/quantization/cutlass_w8a8/broadcast_load_epilogue_c3x.hpp @@ -153,7 +153,7 @@ struct Sm90RowOrScalarBroadcast { CUTLASS_DEVICE void begin(uint64_t* full_mbarrier_ptr, int load_iteration, bool issue_tma_load) { - if (params.ptr_row == nullptr) { + if (!params.row_broadcast) { return; } diff --git a/vllm/model_executor/layers/quantization/fp8.py b/vllm/model_executor/layers/quantization/fp8.py index bc08bfcc32b3a..e89fd65813c05 100644 --- a/vllm/model_executor/layers/quantization/fp8.py +++ b/vllm/model_executor/layers/quantization/fp8.py @@ -257,9 +257,7 @@ class Fp8LinearMethod(LinearMethodBase): # If dynamic, layer.input_scale is None and x_scale computed from x. # If static, layer.input_scale is scalar and x_scale is input_scale. - # Temporarily disable CUTLASS kernels due to an illegal memory access - #if bias is None and self.cutlass_fp8_supported: - if False: + if bias is None and self.cutlass_fp8_supported: qinput, x_scale = ops.scaled_fp8_quant(x, layer.input_scale) # Fused GEMM_DQ