Skip to content

Commit

Permalink
remove force DMMV
Browse files Browse the repository at this point in the history
  • Loading branch information
YellowRoseCx committed Aug 15, 2023
1 parent cd61aa0 commit 7d11961
Showing 1 changed file with 1 addition and 5 deletions.
6 changes: 1 addition & 5 deletions Makefile
Original file line number Diff line number Diff line change
Expand Up @@ -199,28 +199,24 @@ ifdef LLAMA_HIPBLAS
CXX := $(ROCM_PATH)/llvm/bin/clang++
GPU_TARGETS = gfx803 gfx900 gfx906 gfx908 gfx90a gfx1030 gfx1100
LLAMA_CUDA_DMMV_X ?= 128
LLAMA_CUDA_MMV_Y ?= 1
LLAMA_CUDA_MMV_Y ?= 2
LLAMA_CUDA_KQUANTS_ITER ?= 1
LLAMA_CUDA_FORCE_DMMV ?= true
HIPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS $(shell $(ROCM_PATH)/bin/hipconfig -C)
HIPLDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib -lhipblas -lamdhip64 -lrocblas
HIP_OBJS += ggml-cuda.o ggml_v2-cuda.o ggml_v2-cuda-legacy.o
ggml-cuda.o: HIPFLAGS += $(addprefix --offload-arch=,$(GPU_TARGETS)) \
-DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X) \
-DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y) \
-DGGML_CUDA_FORCE_DMMV \
-DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER) \
-DCC_TURING=1000000000
ggml_v2-cuda.o: HIPFLAGS += $(addprefix --offload-arch=,$(GPU_TARGETS)) \
-DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X) \
-DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y) \
-DGGML_CUDA_FORCE_DMMV \
-DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER) \
-DCC_TURING=1000000000
ggml_v2-cuda-legacy.o: HIPFLAGS += $(addprefix --offload-arch=,$(GPU_TARGETS)) \
-DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X) \
-DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y) \
-DGGML_CUDA_FORCE_DMMV \
-DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER) \
-DCC_TURING=1000000000 # DGGML_CUDA_DMMV_F16 does not currently work with AMD.
ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
Expand Down

0 comments on commit 7d11961

Please sign in to comment.