| # One backend per `[kernel.*]` section, all implementing the same two entry points declared in | |
| # torch-ext/torch_binding.h. Adding Metal means adding a section and a `gguf_metal/` directory — | |
| # no change to the bindings' schema or to the Python API. | |
| # | |
| # `vendor/` is a pinned subset of llama.cpp (revision in vendor/UPSTREAM); refresh it with | |
| # `python vendor.py --rev <sha>`. It holds upstream's Metal kernels as they ship -- no local copy is | |
| # generated from them. | |
| [general] | |
| name = "ggml-quantization" | |
| version = 1 | |
| license = "MIT" | |
| # Metal only for now. The cuda kernel and its build section were dropped rather than kept building | |
| # untested -- `git log -- gguf_cuda` has them, and `vendor/` still carries ggml's cuda sources. | |
| backends = ["metal"] | |
| [general.hub] | |
| repo-id = "marcsun13/ggml-quantization" | |
| [torch] | |
| src = [ | |
| "torch-ext/torch_binding.cpp", | |
| "torch-ext/torch_binding.h", | |
| ] | |
| [kernel.gguf_metal] | |
| backend = "metal" | |
| depends = ["torch"] | |
| # ggml-metal.metal includes "ggml-common.h" from vendor/src, so the shader compile needs it | |
| # on its include path. | |
| include = ["gguf_metal", "torch-ext", "vendor/src", "vendor/src/ggml-metal", "vendor/src/ggml-metal/kernels"] | |
| src = [ | |
| "gguf_metal/gguf_metal.cpp", | |
| "gguf_metal/ggml_dispatch.mm", | |
| "gguf_metal/common.h", | |
| # Upstream's Metal kernels, compiled as they ship: one file per operation, so a package | |
| # lists the ones it dispatches. | |
| "vendor/src/ggml-metal/kernels/mul_mv.metal", | |
| "vendor/src/ggml-metal/kernels/mul_mm.metal", | |
| "vendor/src/ggml-metal/kernels/quantize.metal", | |
| "vendor/src/ggml-metal/kernels/common.h", | |
| "vendor/src/ggml-metal/kernels/dequantize.h", | |
| "vendor/src/ggml-metal/kernels/quantize.h", | |
| "vendor/src/ggml-metal/ggml-metal-impl.h", | |
| "vendor/src/ggml-common.h", | |
| ] | |