Skip to content

cuda: add experimental lossless compressed expert tier #651

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
JustVugg merged 6 commits into from
Jul 31, 2026
Merged

cuda: add experimental lossless compressed expert tier #651

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
ca05b77
cuda: reuse decode activations across attention projections
ZacharyZcR Jul 27, 2026
66211fa
Reclaim detached expert arena pages
ZacharyZcR Jul 27, 2026
8cc534c
cuda: add lossless compressed expert tier
ZacharyZcR Jul 27, 2026
3d586d7
docs: document compressed CUDA expert tier
ZacharyZcR Jul 27, 2026
2e4cd63
cuda: pipeline compressed sidecar loading
ZacharyZcR Jul 27, 2026
a6def3b
docs: clarify compressed tier output semantics
ZacharyZcR Jul 28, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 15 additions & 3 deletions c/Makefile
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -130,6 +130,8 @@ INSTALL ?= install
# clean. See backend_loader.c and README "cuda-dll" below.
CUDA ?= 0
CUDA_DLL ?= 0
COLI_ANS ?= 0
DIETGPU_ROOT ?=
ifneq ($(IS_WIN),)
# the CUDA installer sets CUDA_PATH system-wide (e.g. C:\Program Files\NVIDIA
# GPU Computing Toolkit\CUDA\v13.2); fall back to it before the POSIX default.
Expand Down Expand Up @@ -243,6 +245,16 @@ CFLAGS += -DCOLI_CUDA
LDFLAGS += -L$(CUDA_HOME)/lib64 -Wl,-rpath,$(CUDA_HOME)/lib64 -lcudart -lstdc++
CUDA_OBJ = backend_cuda.o
INK_CUDA_OBJ = backend_cuda_ink.o
ifeq ($(COLI_ANS),1)
ifeq ($(strip $(DIETGPU_ROOT)),)
$(error DIETGPU_ROOT is required when COLI_ANS=1)
endif
CFLAGS += -DCOLI_ANS
GPUFLAGS += -DCOLI_ANS -I$(DIETGPU_ROOT) -I$(DIETGPU_ROOT)/build/third_party/glog -I$(DIETGPU_ROOT)/third_party/glog/src
ANS_LIBS = -L$(DIETGPU_ROOT)/build/lib -Wl,-rpath,$(DIETGPU_ROOT)/build/lib -lgpu_ans -ldietgpu_utils -lglog
ANS_NVCC_LIBS = -L$(DIETGPU_ROOT)/build/lib -Xlinker -rpath -Xlinker $(DIETGPU_ROOT)/build/lib -lgpu_ans -ldietgpu_utils -lglog
LDFLAGS += $(ANS_LIBS)
endif
endif

ifeq ($(HIP),1)
Expand Down Expand Up @@ -413,7 +425,7 @@ $(RANSLIB): tools/rans_ctypes.c rans.h

cuda-test: backend_cuda.cu backend_cuda.h backend_gpu_compat.h tests/test_backend_cuda.cu tests/test_ragged_attention.cu
@command -v "$(GPUCC)" >/dev/null 2>&1 || { echo "$(GPUCC_NAME) not found" >&2; exit 1; }
"$(GPUCC)" $(GPUFLAGS) backend_cuda.cu tests/test_backend_cuda.cu -o backend_cuda_test$(EXE)
"$(GPUCC)" $(GPUFLAGS) backend_cuda.cu tests/test_backend_cuda.cu -o backend_cuda_test$(EXE) $(ANS_NVCC_LIBS)
./backend_cuda_test$(EXE)
"$(GPUCC)" $(GPUFLAGS) backend_cuda.cu tests/test_ragged_attention.cu -o ragged_attention_test$(EXE)
./ragged_attention_test$(EXE)
Expand All @@ -425,12 +437,12 @@ hip-test:
# CI: compile the backend and its test binary WITHOUT executing them (for
# runners with the toolchain but no GPU). Pass CUDA_ARCH/HIP=1+HIP_ARCH.
gpu-compile: backend_cuda.o
"$(GPUCC)" $(GPUFLAGS) backend_cuda.cu tests/test_backend_cuda.cu -o backend_cuda_test$(EXE)
"$(GPUCC)" $(GPUFLAGS) backend_cuda.cu tests/test_backend_cuda.cu -o backend_cuda_test$(EXE) $(ANS_NVCC_LIBS)
"$(GPUCC)" $(GPUFLAGS) backend_cuda.cu tests/test_ragged_attention.cu -o ragged_attention_test$(EXE)

cuda-bench: backend_cuda.cu backend_cuda.h backend_gpu_compat.h tests/bench_tensor_core.cu
@command -v "$(GPUCC)" >/dev/null 2>&1 || { echo "$(GPUCC_NAME) not found" >&2; exit 1; }
"$(GPUCC)" $(GPUFLAGS) backend_cuda.cu tests/bench_tensor_core.cu -o backend_cuda_bench$(EXE)
"$(GPUCC)" $(GPUFLAGS) backend_cuda.cu tests/bench_tensor_core.cu -o backend_cuda_bench$(EXE) $(ANS_NVCC_LIBS)
./backend_cuda_bench$(EXE)

olmoe$(EXE): olmoe.c st.h json.h compat.h sample.h tok.h tok_unicode.h tok_unicode_o200k.h
Expand Down
Loading
Loading