cuda: add bit-exact warp decode path for low-row INT4

GPU_BACKENDS.md

@@ -41,6 +41,10 @@ distributing or on machines with an unsupported iGPU visible to the runtime
   header) because gfx GPUs report `compute_major >= 7` and a runtime check
   alone would select empty kernel bodies under HIP. On AMD, all compute uses
   the portable kernels; rocWMMA matrix-core support is a possible follow-up.
+- `COLI_CUDA_WARP_DECODE=1` — opt-in low-row INT4 decode kernel on NVIDIA.
+  Maps one output neuron to one 32-lane warp and eight adjacent outputs to
+  each block, while preserving the original 256-thread accumulation order.
+  This removes block-wide barriers without changing results bit-for-bit.
 
 ## Validation
 

c/Makefile

@@ -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 ($(IS_WIN),)
 # the CUDA installer sets CUDA_PATH system-wide (e.g. C:\Program Files\NVIDIA
 # GPU Computing Toolkit\CUDA\v13.2); fall back to it before the POSIX default.
@@ -228,6 +230,16 @@ CFLAGS   += -DCOLI_CUDA
 LDFLAGS  += -L$(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)
@@ -324,7 +336,7 @@ gemm-test: tests/test_gemm_largebatch.mm backend_metal.mm backend_metal.h
 
 cuda-test: backend_cuda.cu backend_cuda.h backend_gpu_compat.h tests/test_backend_cuda.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)
 
 # convenience alias: kernel correctness on AMD (same test, hipcc toolchain)
@@ -334,11 +346,11 @@ 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)
 
 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