GPU backends: CUDA and HIP/ROCm

colibrì's GPU expert backend is one source file (c/backend_cuda.cu) compiled for either vendor through c/backend_gpu_compat.h — the same one-shim-header pattern compat.h uses for the Windows port. Compiled by nvcc the shim is a pass-through to cuda_runtime.h (the NVIDIA path is byte-identical to the pre-HIP tree); compiled by hipcc it maps the 14-symbol CUDA runtime surface the backend uses onto HIP 1:1. The kernels use only shared syntax (__global__, __shared__, __syncthreads__, <<<>>>), no vendor intrinsics.

Rule for contributors: vendor differences go in backend_gpu_compat.h only — never #ifdef __HIP__ (or CUDA-specific code) in backend_cuda.cu.

Supported environments

backend	platform	toolchain	build
CUDA (CUDA=1)	Linux x86-64	CUDA toolkit (nvcc), CUDA_HOME=/usr/local/cuda default	make -C c glm CUDA=1 [CUDA_ARCH=native|sm_XX]
HIP (HIP=1)	Linux x86-64	ROCm (hipcc), ROCM_HOME=/opt/rocm default; tested on ROCm 7.2	make -C c glm HIP=1 [HIP_ARCH=native|gfxXXXX]
HIP DLL (HIP_DLL=1) — build and runtime; validated on one configuration	Windows x86-64	HIP SDK (hipcc) + a compatible MSVC x64 host toolchain; HIP_SDK_ROOT from HIP_PATH	make -C c hip-dll HIP_DLL=1 HIP_SDK_ROOT=<sdk-root> HIP_ARCH=gfxNNNN → c/coli_hip.dll

CUDA=1 and HIP=1 are mutually exclusive and both opt-in: the default build remains pure, dependency-free CPU. Both are directly linked paths and remain Linux-only, refused elsewhere with an early $(error). The Windows HIP path is selected separately as HIP_DLL=1 and does not go through HIP=1; on Windows CUDA_DLL=1 and HIP_DLL=1 are likewise mutually exclusive. *_ARCH=native targets the local GPU; pass an explicit arch when distributing or on machines with an unsupported iGPU visible to the runtime (and mask iGPUs at runtime with HIP_VISIBLE_DEVICES=<ordinal> on ROCm). On Windows HIP_ARCH must be an explicit gfxNNNN — see below.

Windows HIP DLL

Mirrors the Windows CUDA split: MinGW gcc cannot compile .cu, and Windows hipcc targets the MSVC ABI, so the backend is built into a standalone coli_hip.dll instead of being linked into the host. The same c/backend_cuda.cu and the same coli_cuda_* ABI the Linux HIP path already reuses are used unchanged.

The host loads coli_hip.dll at runtime through the same loader seam the Windows CUDA split uses, and binds the HIP runtime explicitly: the directory holding amdhip64_7.dll is named by COLI_HIP_RUNTIME_DIR, and the loader fails closed if the mapped module is not that exact file or if a second module of the same basename is already present. It never silently accepts a copy from System32, from PATH, or from an unrelated ROCm install.

Read what this was tested on before relying on it; docs/windows.md has the full setup walkthrough.

The two halves are built separately.

Host build mode — prepares the host for the DLL split. Needs no HIP SDK and no HIP_ARCH, because it only compiles c/backend_loader.c and links colibri.exe; amdhip64 is never linked into the host:

make -C c colibri.exe HIP_DLL=1

Backend DLL — requires a selected SDK and an explicit architecture:

make -C c hip-dll \
    HIP_DLL=1 \
    HIP_SDK_ROOT=<sdk-root> \
    HIP_ARCH=gfxNNNN

It produces c/coli_hip.dll; the linker may also emit c/coli_hip.lib (and .exp/.pdb on toolchains that generate them). All are ignored by git and removed by make -C c clean.

SDK selection variables

No install location is hardcoded. HIP_SDK_ROOT defaults from the HIP_PATH environment variable the Windows HIP SDK installer sets; pass it explicitly for a relocated or source-built SDK (for example from rocm-sdk path --root). Every component can be overridden independently, because packaged layouts do not all keep runtime, development and device files under one root:

variable	default	selects
HIP_SDK_ROOT	$(HIP_PATH)	SDK root (--hip-path)
HIP_BIN_DIR	$(HIP_SDK_ROOT)/bin	directory holding hipcc
HIP_INCLUDE_DIR	$(HIP_SDK_ROOT)/include	headers (-I)
HIP_LIB_DIR	$(HIP_SDK_ROOT)/lib	amdhip64 import library (-L)
HIP_DEVICE_LIB_PATH	$(HIP_SDK_ROOT)/lib/llvm/amdgcn/bitcode	device bitcode (--rocm-device-lib-path)
HIPCC	$(HIP_BIN_DIR)/hipcc.exe	compiler driver
HIP_ARCH	(none — must be explicit)	--offload-arch=

--hip-path pins the SDK deliberately: a machine can carry both a driver-installed HIP SDK and a source-built one, and clang otherwise injects the former on its own, which silently mixes headers from one tree with the import library and device bitcode from the other.

HIP_ARCH=native is rejected on Windows: rocm_agent_enumerator is not available there, so there is nothing to resolve native against.

What this was tested on, and what it does not claim

Validated on one configuration: AMD Radeon(TM) 8060S Graphics reporting gfx1151, TheRock HIP 7.14.60850, VS2022 MSVC 14.44.35207, Windows SDK 10.0.26100.0. Nothing here is a statement about other GPUs, drivers or SDKs.

• Hybrid placement. In the validated workload CUDA_DENSE=1 put the eligible dense tensors on HIP while routed experts stayed on CPU. Routed experts reach the GPU only through the existing expert-placement controls (CUDA_EXPERT_GB plus a pin/usage source), which is untested here. This is not full-GPU MoE inference.
• Device discovery is not proof of GPU work. A run can print [CUDA] device 0: ... and still execute every tensor on CPU. The number that settles it is [CUDA] resident set: N tensors — N = 0 means no model tensor was resident on the GPU. Check that, not the device line.
• CPU fallback keeps the command successful. A per-tensor upload failure falls back to CPU and the process still exits 0. Read the first fallback diagnostics and the final fallback count together with the resident set.
• Lifecycle. Normal one-shot model exits rely on Windows process teardown rather than an explicit coli_cuda_shutdown call. This matches the existing host path and is not presented as explicit backend shutdown.
• There is no hosted CI coverage for the Windows HIP build or runtime — no hosted runner provides a Windows HIP toolchain or an AMD GPU. engine-hip-syntax covers the Linux HIP compile only; the Windows job added alongside this work runs the loader contract tests with synthetic DLL fixtures and no GPU.
• Physical validation therefore comes from a local Windows AMD/HIP environment and is reported as external evidence, not as CI coverage.

Runtime configuration (identical for both vendors)

• COLI_CUDA=1 + COLI_GPU=N (or COLI_GPUS=0,1,...) — enable, select devices
• CUDA_EXPERT_GB=G — VRAM budget for the expert tier (clamped to free VRAM minus projected dense set and 2 GB headroom per device)
• CUDA_RELEASE_HOST=1 — GPU-tier experts drop their host backing after upload (default on multi-GPU); combined with PIN=auto/PIN_FILL, VRAM becomes additional pinned capacity at zero RAM cost. The engine rematerializes an expert from disk (expert_host_ensure) whenever the CPU path needs one whose host copy was released — validated under total GPU failure below.
• CUDA_DENSE=1 — experimental resident-dense path (unchanged)
• COLI_CUDA_TC_W4A16=1 — opt-in W4A16 tensor-core path. NVIDIA-only: the WMMA kernels are compile-gated (COLI_GPU_HAS_WMMA in the compat 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.

Validation

Unit tests (run on GPU hardware)

make -C c cuda-test [CUDA_ARCH=...]    # NVIDIA
make -C c hip-test  [HIP_ARCH=...]     # AMD (same test source)

Covers q8/q4/q2/f32 matmul correctness, multi-device placement/stats, and tensor_update — the standard upstream suite, unchanged, compiled by hipcc. (A companion PR adds failure-path tests for the backend; they are vendor-neutral and run under hip-test identically.)

CI (no GPU required)

The engine-hip-syntax job in .github/workflows/ci.yml compiles the backend and its test binary with hipcc (rocm/dev container pinned to 6.2, gfx1100) on every PR, mirroring engine-cuda-syntax. Kernel execution is not possible on hosted runners; that is what hip-test on real hardware is for (matrix below).

Hardware test matrix (documented results)

environment	result
AMD RX 9070 XT (gfx1201), ROCm 7.2.4, Linux 7.0	hip-test pass (all cases above); GLM-5.2 end-to-end runs (0.32 tok/s @ 61% expert hit with CUDA_RELEASE_HOST=1); benchmark series in PR #112
NVIDIA	compile-verified in CI (sm_80); nvcc path is a pass-through include — runtime run of make cuda-test on NVIDIA hardware welcomed, the test source is vendor-neutral

Known behavior notes

• GPU float matmuls round differently than the CPU int8-dot (IDOT) kernels: greedy output is not token-identical across backends (consistent with the shape-dependence documented in #100), and MTP draft acceptance measures lower on GPU-heavy configs (~40% → ~31% on the PR #112 machine). A numerics-matched integer GPU kernel is the planned follow-up.
• An earlier revision of this branch carried CUDA_EXTEND=1 (VRAM tier holding experts beyond the RAM pin). It was superseded by upstream's PIN=auto + PIN_FILL + CUDA_RELEASE_HOST, which achieve the same capacity extension with deeper engine integration; this branch's safety and validation work now targets that mechanism.