chore(krz01): Put all ollama/whisper stuff in separate modules #157
10 changed files with 109 additions and 43 deletions
|
@ -1,11 +1,4 @@
|
|||
{
|
||||
config,
|
||||
lib,
|
||||
pkgs,
|
||||
meta,
|
||||
name,
|
||||
...
|
||||
}:
|
||||
{ config, lib, ... }:
|
||||
|
||||
lib.extra.mkConfig {
|
||||
enabledModules = [
|
||||
|
@ -18,6 +11,8 @@ lib.extra.mkConfig {
|
|||
"microvm-ml01"
|
||||
"microvm-router01"
|
||||
"nvidia-tesla-k80"
|
||||
"ollama"
|
||||
"whisper"
|
||||
"proxmox"
|
||||
];
|
||||
|
||||
|
@ -33,45 +28,11 @@ lib.extra.mkConfig {
|
|||
];
|
||||
};
|
||||
|
||||
services.netbird.enable = true;
|
||||
|
||||
# We are going to use CUDA here.
|
||||
nixpkgs.config.cudaSupport = true;
|
||||
hardware.graphics.enable = true;
|
||||
environment.systemPackages = [
|
||||
((pkgs.openai-whisper-cpp.override { cudaPackages = pkgs.cudaPackages_11; }).overrideAttrs (old: {
|
||||
src = pkgs.fetchFromGitHub {
|
||||
owner = "ggerganov";
|
||||
repo = "whisper.cpp";
|
||||
rev = "v1.7.1";
|
||||
hash = "sha256-EDFUVjud79ZRCzGbOh9L9NcXfN3ikvsqkVSOME9F9oo=";
|
||||
};
|
||||
env = {
|
||||
WHISPER_CUBLAS = "";
|
||||
GGML_CUDA = "1";
|
||||
};
|
||||
# We only need Compute Capability 3.7.
|
||||
CUDA_ARCH_FLAGS = [ "sm_37" ];
|
||||
# We are GPU-only anyway.
|
||||
patches = (old.patches or [ ]) ++ [
|
||||
./no-weird-microarch.patch
|
||||
./all-nvcc-arch.patch
|
||||
];
|
||||
}))
|
||||
];
|
||||
|
||||
services = {
|
||||
ollama = {
|
||||
enable = true;
|
||||
host = meta.network.${name}.netbirdIp;
|
||||
package = pkgs.callPackage ./ollama.nix {
|
||||
cudaPackages = pkgs.cudaPackages_11;
|
||||
# We need to thread our nvidia x11 driver for CUDA.
|
||||
extraLibraries = [ config.hardware.nvidia.package ];
|
||||
};
|
||||
};
|
||||
};
|
||||
|
||||
services.netbird.enable = true;
|
||||
networking.firewall.trustedInterfaces = [ "wt0" ];
|
||||
};
|
||||
|
||||
|
|
20
machines/krz01/ollama/default.nix
Normal file
20
machines/krz01/ollama/default.nix
Normal file
|
@ -0,0 +1,20 @@
|
|||
{
|
||||
config,
|
||||
pkgs,
|
||||
meta,
|
||||
name,
|
||||
...
|
||||
}:
|
||||
{
|
||||
services = {
|
||||
ollama = {
|
||||
enable = true;
|
||||
host = meta.network.${name}.netbirdIp;
|
||||
package = pkgs.callPackage ./package.nix {
|
||||
cudaPackages = pkgs.cudaPackages_11;
|
||||
# We need to thread our nvidia x11 driver for CUDA.
|
||||
extraLibraries = [ config.hardware.nvidia.package ];
|
||||
};
|
||||
};
|
||||
};
|
||||
}
|
26
machines/krz01/whisper/all-nvcc-arch.patch
Normal file
26
machines/krz01/whisper/all-nvcc-arch.patch
Normal file
|
@ -0,0 +1,26 @@
|
|||
From 2278389ef9ac9231349440aa68f9544ddc69cdc7 Mon Sep 17 00:00:00 2001
|
||||
From: Raito Bezarius <masterancpp@gmail.com>
|
||||
Date: Wed, 9 Oct 2024 13:37:08 +0200
|
||||
Subject: [PATCH] fix: sm_37 for nvcc
|
||||
|
||||
Signed-off-by: Raito Bezarius <masterancpp@gmail.com>
|
||||
---
|
||||
Makefile | 2 +-
|
||||
1 file changed, 1 insertion(+), 1 deletion(-)
|
||||
|
||||
diff --git a/Makefile b/Makefile
|
||||
index 2ccb750..70dfd9b 100644
|
||||
--- a/Makefile
|
||||
+++ b/Makefile
|
||||
@@ -537,7 +537,7 @@ endif #GGML_CUDA_NVCC
|
||||
ifdef CUDA_DOCKER_ARCH
|
||||
MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
|
||||
else ifndef CUDA_POWER_ARCH
|
||||
- MK_NVCCFLAGS += -arch=native
|
||||
+ MK_NVCCFLAGS += -arch=sm_37
|
||||
endif # CUDA_DOCKER_ARCH
|
||||
|
||||
ifdef GGML_CUDA_FORCE_DMMV
|
||||
--
|
||||
2.46.0
|
||||
|
25
machines/krz01/whisper/default.nix
Normal file
25
machines/krz01/whisper/default.nix
Normal file
|
@ -0,0 +1,25 @@
|
|||
{ pkgs, ... }:
|
||||
{
|
||||
environment.systemPackages = [
|
||||
((pkgs.openai-whisper-cpp.override { cudaPackages = pkgs.cudaPackages_11; }).overrideAttrs (old: {
|
||||
src = pkgs.fetchFromGitHub {
|
||||
owner = "ggerganov";
|
||||
repo = "whisper.cpp";
|
||||
rev = "v1.7.1";
|
||||
hash = "sha256-EDFUVjud79ZRCzGbOh9L9NcXfN3ikvsqkVSOME9F9oo=";
|
||||
};
|
||||
env = {
|
||||
WHISPER_CUBLAS = "";
|
||||
GGML_CUDA = "1";
|
||||
};
|
||||
# We only need Compute Capability 3.7.
|
||||
CUDA_ARCH_FLAGS = [ "sm_37" ];
|
||||
# We are GPU-only anyway.
|
||||
patches = (old.patches or [ ]) ++ [
|
||||
./no-weird-microarch.patch
|
||||
./all-nvcc-arch.patch
|
||||
];
|
||||
}))
|
||||
];
|
||||
|
||||
}
|
34
machines/krz01/whisper/no-weird-microarch.patch
Normal file
34
machines/krz01/whisper/no-weird-microarch.patch
Normal file
|
@ -0,0 +1,34 @@
|
|||
From 51568b61ef63ecd97867562571411082c32751d3 Mon Sep 17 00:00:00 2001
|
||||
From: Raito Bezarius <masterancpp@gmail.com>
|
||||
Date: Wed, 9 Oct 2024 13:36:51 +0200
|
||||
Subject: [PATCH] fix: avx & f16c in Makefile
|
||||
|
||||
Signed-off-by: Raito Bezarius <masterancpp@gmail.com>
|
||||
---
|
||||
Makefile | 8 ++++----
|
||||
1 file changed, 4 insertions(+), 4 deletions(-)
|
||||
|
||||
diff --git a/Makefile b/Makefile
|
||||
index 32b7cbb..2ccb750 100644
|
||||
--- a/Makefile
|
||||
+++ b/Makefile
|
||||
@@ -361,12 +361,12 @@ ifndef RISCV
|
||||
|
||||
ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
|
||||
# Use all CPU extensions that are available:
|
||||
- MK_CFLAGS += -march=native -mtune=native
|
||||
- HOST_CXXFLAGS += -march=native -mtune=native
|
||||
+ # MK_CFLAGS += -march=native -mtune=native
|
||||
+ # HOST_CXXFLAGS += -march=native -mtune=native
|
||||
|
||||
# Usage AVX-only
|
||||
- #MK_CFLAGS += -mfma -mf16c -mavx
|
||||
- #MK_CXXFLAGS += -mfma -mf16c -mavx
|
||||
+ MK_CFLAGS += -mf16c -mavx
|
||||
+ MK_CXXFLAGS += -mf16c -mavx
|
||||
|
||||
# Usage SSSE3-only (Not is SSE3!)
|
||||
#MK_CFLAGS += -mssse3
|
||||
--
|
||||
2.46.0
|
||||
|
Loading…
Reference in a new issue