[drbh]

this pr adds a rust-cuda backend where the kernel is a single crate compiled with cuda-oxide. its an alternative to #907, opened separately so the two approaches can be compared

similar to nvcc -gencode the crate is built once per capability in cuda-capabilities (same CUDA_KERNEL_ARCHS default as cuda)

• each pass sets cfg(cuda_arch = "XX") for its capability and every lower one, so #[cfg(cuda_arch = "90")] acts like __CUDA_ARCH__ >= 900
• each pass's ptx is compiled to sass with ptxas and packed into a fatbin in .nv_fatbin so cuobjdump can inspect it
• the embedded ptx (.oxart) is stripped so nothing is jit compiled at load. +PTX ships that capability's ptx like cuda does
• the new kernels-cuda crate loads the fatbin with cuModuleLoadData and launches on torch's current stream
• cuda bindings come from https://github.com/drbh/cuda-bindings-prebuilt so bindgen is skipped

note: there can only be one rust-cuda component per extension and kernels should be loaded with kernels-cuda (the generated load doesnt link). build time also scales with the number of capabilities since its one cargo pass each (flash-attn2-rs is 787s for 5 caps vs 351s single pass)

this pr is dependent on NVIDIA/cuda-rust#1601 (CudaStream::borrow_raw) and pins a fork until that is merged

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[github-actions]

Coverage report — kernels/

Measured on: Python 3.10 / Torch 2.13.0.
Other CI configurations are not included in this number.
Hardware-gated code paths (ROCm/XPU/NPU/Darwin/Windows) are excluded or unreachable on the Linux+CUDA runner.

Total coverage: 88.5% — threshold: 80% — ✅

Per-file breakdown

Name	Stmts	Miss	Cover	Missing
src/kernels/__init__.py	14	0	100%
src/kernels/_system.py	6	1	83%	10
src/kernels/_versions.py	130	14	89%	53, 59-60, 63-64, 102, 165-170, 199, 219
src/kernels/archs.py	56	1	98%	94
src/kernels/backends.py	215	63	71%	42, 46, 50-53, 70, 92, 110, 119, 123, 127-129, 150, 159, 163, 167-169, 190, 201, 203, 210-213, 222, 226, 230-255, 263, 286-306
src/kernels/compat.py	9	1	89%	5
src/kernels/deps.py	70	1	99%	56
src/kernels/digest.py	58	5	91%	71-73, 136-137
src/kernels/hf_hub.py	62	2	97%	21, 23
src/kernels/importer.py	74	5	93%	158, 162, 165, 179-180
src/kernels/install.py	21	7	67%	76-100
src/kernels/layer/__init__.py	7	0	100%
src/kernels/layer/_interval_tree.py	103	4	96%	23, 52, 147, 150
src/kernels/layer/device.py	48	12	75%	42, 47-49, 91, 96-98, 101, 149, 152, 156
src/kernels/layer/func.py	93	5	95%	94, 119, 195, 315, 390
src/kernels/layer/globals.py	5	0	100%
src/kernels/layer/inspect.py	27	0	100%
src/kernels/layer/kernelize.py	83	7	92%	306, 314-315, 321, 325, 341-343
src/kernels/layer/layer.py	320	14	96%	212, 259, 447, 606-607, 628, 636, 647, 688, 692, 705, 760, 801, 876
src/kernels/layer/mode.py	14	0	100%
src/kernels/layer/repos.py	136	36	74%	32, 35-42, 73, 86, 90, 99-100, 103, 106-109, 116-117, 120, 123-126, 133-134, 137, 140-143, 150-151, 154, 157-160, 235
src/kernels/load.py	71	2	97%	338, 378
src/kernels/locking.py	89	64	28%	35-83, 91-98, 102-125, 137, 152-159, 165-175, 179-186
src/kernels/python_deps.py	58	6	90%	59-60, 64-65, 101, 104
src/kernels/resolver.py	162	1	99%	248
src/kernels/status.py	50	2	96%	25, 79
src/kernels/validate.py	104	7	93%	9, 101, 172, 195-196, 232-233
src/kernels/variants.py	278	17	94%	65, 96, 117, 147, 256-257, 299-302, 304, 388-394, 400-406, 455-461
src/kernels/verify.py	102	6	94%	44, 170-172, 260-261
TOTAL	2465	283	89%

Updated by the Test kernels workflow on commit 90c61854d8fbc9e563f34859bc469d6fc66fc552.