From e0bb8114f43dbe1d65031dfcde7e30203c3a48d6 Mon Sep 17 00:00:00 2001 From: Hongtao Zhang Date: Wed, 11 Mar 2026 22:53:10 +0000 Subject: [PATCH 1/3] Add SM80 architecture to CUTLASS builds for CUDA 12.8+ and 12.9+ Add sm_80 to the CUTLASS architecture lists to ensure Ampere GPU support is included in builds targeting CUDA 12.8+ and 12.9+. --- third_party/Makefile | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/third_party/Makefile b/third_party/Makefile index 2a09f5990..f9a50212a 100755 --- a/third_party/Makefile +++ b/third_party/Makefile @@ -42,11 +42,11 @@ sb_micro_path: # for cuda 12.9 and later Build from commit v3.9 (3.9 release commit) for blackwell support cuda_cutlass: ifeq ($(shell echo $(CUDA_VER)">=12.9" | bc -l), 1) - $(eval ARCHS := "100;103") + $(eval ARCHS := "80;90;100;103") if [ -d cutlass ]; then rm -rf cutlass; fi git clone --branch v4.1.0 --depth 1 https://github.com/NVIDIA/cutlass.git && cd cutlass else ifeq ($(shell echo $(CUDA_VER)">=12.8" | bc -l), 1) - $(eval ARCHS := "90;100") + $(eval ARCHS := "80;90;100") if [ -d cutlass ]; then rm -rf cutlass; fi git clone --branch v3.9.2 --depth 1 https://github.com/NVIDIA/cutlass.git && cd cutlass else ifeq ($(shell echo $(CUDA_VER)">=11.8" | bc -l), 1) From d56fb5c2f3f1bbf271c53a8b6a1157c21cfb330e Mon Sep 17 00:00:00 2001 From: Hongtao Zhang Date: Tue, 17 Mar 2026 20:45:25 +0000 Subject: [PATCH 2/3] Add arch support for MSCCL. --- third_party/Makefile | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/third_party/Makefile b/third_party/Makefile index f9a50212a..ce1916154 100755 --- a/third_party/Makefile +++ b/third_party/Makefile @@ -270,7 +270,7 @@ apex_rocm: # Build MSCCL for CUDA cuda_msccl: sb_micro_path ifeq ($(shell echo $(CUDA_VER)">=12.9" | bc -l), 1) - $(eval ARCHS := 100 103) + $(eval ARCHS := 80 90 100 103) if [ -d msccl ]; then rm -rf msccl; fi; \ git clone --single-branch --branch main https://github.com/Azure/msccl.git \ && git -C msccl checkout 87048bd && git -C msccl submodule update --recursive --init From b8cfc300be56486e1ab0ec24fb92810541065371 Mon Sep 17 00:00:00 2001 From: Hongtao Zhang Date: Tue, 17 Mar 2026 13:46:44 -0700 Subject: [PATCH 3/3] Apply suggestion from @Copilot Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> --- third_party/Makefile | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/third_party/Makefile b/third_party/Makefile index ce1916154..8c1128463 100755 --- a/third_party/Makefile +++ b/third_party/Makefile @@ -39,7 +39,7 @@ sb_micro_path: mkdir -p $(SB_MICRO_PATH)/lib # Build cutlass. -# for cuda 12.9 and later Build from commit v3.9 (3.9 release commit) for blackwell support +# For CUDA 12.9 and later, build from CUTLASS v4.1.0 (required for Blackwell support). cuda_cutlass: ifeq ($(shell echo $(CUDA_VER)">=12.9" | bc -l), 1) $(eval ARCHS := "80;90;100;103")