summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorVineet Kumar <git@vineetk.net>2026-05-17 16:56:56 -0400
committerVineet Kumar <git@vineetk.net>2026-05-17 16:56:56 -0400
commit6dc14eccb88e4bcb9b083974a38d5fc1d1b5f9d4 (patch)
tree97b399e697fb25a8a3ecbd740979ee5f5001caeb
parent120ddeccd153782de13e66c4abecafed38e800b5 (diff)
packages/machine-learning: update llama-cpp-rocm to honorary b9203
no tags yet after b9198. two commits after that improves mtp=on pp speed, so might as well use latest commit as of right now.
-rw-r--r--epistemia/packages/machine-learning.scm21
1 files changed, 12 insertions, 9 deletions
diff --git a/epistemia/packages/machine-learning.scm b/epistemia/packages/machine-learning.scm
index 18baf5e..d7dcb0e 100644
--- a/epistemia/packages/machine-learning.scm
+++ b/epistemia/packages/machine-learning.scm
@@ -16,7 +16,7 @@
16 #:use-module (gnu packages rocm-libs)) 16 #:use-module (gnu packages rocm-libs))
17 17
18(define-public llama-cpp-rocm 18(define-public llama-cpp-rocm
19 (let ((tag "b9082")) 19 (let ((tag "87589042cac2c390cec8d68fb2fad64e0a2a252a"))
20 (package 20 (package
21 (inherit llama-cpp) 21 (inherit llama-cpp)
22 (name "llama-cpp-rocm") 22 (name "llama-cpp-rocm")
@@ -25,28 +25,31 @@
25 (origin 25 (origin
26 (method git-fetch) 26 (method git-fetch)
27 (uri (git-reference 27 (uri (git-reference
28 (url "https://github.com/ggml-org/llama.cpp") 28 (url "https://github.com/ggml-org/llama.cpp")
29 (commit tag))) 29 (commit tag)))
30 (file-name (git-file-name name tag)) 30 (file-name (git-file-name name tag))
31 (sha256 31 (sha256
32 (base32 "1y1qf4ig4xziyrx1vp87s7v1c2zs475l59yfx4ld5f2gpmibazgy")))) 32 (base32 "08q6k02rrrzizhlhfyw3bdqq03vrl6yaz8h91i5zach26ck01c3k"))))
33 (arguments 33 (arguments
34 (substitute-keyword-arguments (package-arguments llama-cpp) 34 (substitute-keyword-arguments (package-arguments llama-cpp)
35 ((#:configure-flags flags #~'()) 35 ((#:configure-flags flags #~'())
36 #~(append #$flags 36 #~(append #$flags
37 '(;; CMAKE_HIP_FLAGS is to workaround regression in ROCm 7.2 37 '(;; CMAKE_HIP_FLAGS is to workaround regression in ROCm 7.2
38 "-DCMAKE_HIP_FLAGS=\"-mllvm --amdgpu-unroll-threshold-local=600\"" 38 "-DCMAKE_HIP_FLAGS=\"-mllvm --amdgpu-unroll-threshold-local=600\""
39 "-DGPU_TARGETS=gfx1100"))) 39 "-DGPU_TARGETS=gfx1100"
40 ;; temporarily disable webui building until I figure out how to add second source for hf bucket
41 "-DLLAMA_USE_PREBUILT_UI=OFF" "-DLLAMA_BUILD_UI=OFF")))
40 ((#:phases phases) 42 ((#:phases phases)
41 #~(modify-phases #$phases 43 #~(modify-phases #$phases
42 ;; the test-chat test got two new extra lines that need to be removed 44 ;; mtp added test that also needs internet for downloading its model
43 (add-after 'fix-tests 'fix-tests-new 45 (add-after 'fix-tests 'fix-tests-new
44 (lambda _ 46 (lambda _
45 (substitute* "tests/CMakeLists.txt" 47 (substitute* "tests/CMakeLists.txt"
46 (("target_include_directories\\(test-chat PRIVATE.*") 48 (("llama_build_and_test\\(test-recurrent-state-rollback\\.cpp .*")
47 "") 49 "")
48 (("target_link_libraries\\(test-chat PRIVATE.*") 50 (("set_tests_properties\\(test-recurrent-state-rollback PROPERTIES.*")
49 "")))))))) 51 ""))))
52 (delete 'fix-python-shebang)))))
50 (inputs 53 (inputs
51 (modify-inputs (package-inputs llama-cpp) 54 (modify-inputs (package-inputs llama-cpp)
52 (replace "ggml" ggml-rocm)))))) 55 (replace "ggml" ggml-rocm))))))