1 files changed, 83 insertions, 0 deletions

diff --git a/epistemia/packages/machine-learning.scm b/epistemia/packages/machine-learning.scm
new file mode 100644
index 0000000..b3e121b
--- /dev/null
+++ b/epistemia/packages/machine-learning.scm
@@ -0,0 +1,83 @@
+(define-module (epistemia packages machine-learning)
+  #:use-module (guix packages)
+  #:use-module (guix download)
+  #:use-module (guix utils)
+  #:use-module (guix git-download)
+  #:use-module (guix gexp)
+  #:use-module (gnu packages llvm)
+  #:use-module (gnu packages machine-learning)
+  #:use-module (gnu packages rocm)
+  #:use-module (gnu packages rocm-libs))
+
+(define-public llama-cpp-rocm
+  (let ((tag "b8931"))
+    (package
+      (inherit llama-cpp)
+      (name "llama-cpp-rocm")
+      (version (string-append "0.0.0-" tag))
+      (source
+       (origin
+         (method git-fetch)
+         (uri (git-reference
+               (url "https://github.com/ggml-org/llama.cpp")
+               (commit tag)))
+         (file-name (git-file-name name tag))
+         (sha256
+          (base32 "1bjhkmgj83z5br70j4pgz1sa1jvbc4rmxm7kca2yr4f6wxk47sgx"))))
+      (arguments
+       (substitute-keyword-arguments (package-arguments llama-cpp)
+         ((#:configure-flags flags #~'())
+          #~(append #$flags
+                    '(;; CMAKE_HIP_FLAGS is to workaround regression in ROCm 7.2
+                      "-DCMAKE_HIP_FLAGS=\"-mllvm --amdgpu-unroll-threshold-local=600\""
+                      "-DGPU_TARGETS=gfx1100")))
+         ((#:phases phases)
+          #~(modify-phases #$phases
+              ;; the test-chat test got two new extra lines that need to be removed
+              (add-after 'fix-tests 'fix-tests-new
+                (lambda _
+                  (substitute* "tests/CMakeLists.txt"
+                    (("target_include_directories\\(test-chat PRIVATE.*")
+                     "")
+                    (("target_link_libraries\\(test-chat PRIVATE.*")
+                     ""))))))))
+      (inputs
+       (modify-inputs (package-inputs llama-cpp)
+         (replace "ggml" ggml-rocm))))))
+
+(define-public ggml-rocm
+  (package
+    (inherit (@@ (gnu packages machine-learning) ggml))
+    (name "ggml-rocm")
+    (version "0.10.0")
+    (source
+     (origin
+       (method git-fetch)
+       (uri (git-reference
+              (url "https://github.com/ggml-org/ggml")
+              (commit (string-append "v" version))))
+       (file-name (git-file-name name version))
+       (sha256
+        (base32 "0pp3zsv5rckb4rr6ddd3d1qmpmg7ifzx46v932i5dndfaa8bp2yr"))))
+    (arguments
+     (substitute-keyword-arguments (package-arguments (@@ (gnu packages machine-learning) ggml))
+       ((#:configure-flags flags #~'())
+        #~(append #$flags (list (string-append "-DCMAKE_HIP_COMPILER_ROCM_ROOT=" #$rocm-hip-runtime)
+                                (string-append "-DCMAKE_HIP_FLAGS=--rocm-path=" #$rocm-hip-runtime
+                                               " --rocm-device-lib-path=" #$rocm-device-libs "/amdgcn/bitcode")
+                                "-DGPU_TARGETS=gfx1100"
+                                "-DGGML_HIP=ON"
+                                "-DGGML_HIP_ROCWMMA_FATTN=OFF"
+                                "-DGGML_CUDA_FA_ALL_QUANTS=ON")))))
+    (inputs
+     (modify-inputs (package-inputs (@@ (gnu packages machine-learning) ggml))
+       (append clang-rocm
+               lld-rocm
+               rocm-cmake
+               rocm-device-libs
+               rocm-hipcc
+               rocm-hip-runtime
+               rocr-runtime
+               hipblas
+               hipblas-common
+               rocblas)))))