[bug#72027] [PATCH] gnu: Add whisper-cpp.

[bug#72027] [PATCH] gnu: Add whisper-cpp.

[Andy Tai]

* gnu/packages/machine-learning.scm (whisper-cpp): New variable.

Change-Id: If191f434a3f66b16afdc702069eaed1ca22e2427
---
 gnu/packages/machine-learning.scm | 23 +++++++++++++++++++++++
 1 file changed, 23 insertions(+)

diff --git a/gnu/packages/machine-learning.scm b/gnu/packages/machine-learning.scm
index 8ca9a6376e..61c25b49da 100644
--- a/gnu/packages/machine-learning.scm
+++ b/gnu/packages/machine-learning.scm
@@ -5982,6 +5982,29 @@ (define-public oneapi-dnnl-for-r-torch
        (sha256
         (base32 "1zyw5rd8x346bb7gac9a7x3saviw3zvp6aqz2z1l9sv163vmjfz6"))))))
 
+(define-public whisper-cpp
+    (package
+      (name "whisper-cpp")
+      (version "1.6.2")
+      (source
+       (origin
+         (method git-fetch)
+         (uri (git-reference
+               (url "https://github.com/ggerganov/whisper.cpp")
+               (commit (string-append "v" version))))
+         (file-name (git-file-name name version))
+         (sha256
+          (base32 "01q4j602wkvsf9vw0nsazzgvjppf4fhpy90vqnm9affynyxhi0c4"))))
+      (build-system cmake-build-system)
+      (properties '((tunable? . #true))) ;use AVX512, FMA, etc. when available
+      (home-page "https://github.com/ggerganov/whisper.cpp")
+      (synopsis "Port of OpenAI's Whisper model in C/C++ ")
+      (description "This package is a high-performance inference of OpenAI's
+Whisper automatic speech recognition (ASR) model, implemented in plain C/C++
+without dependencies")
+      (license license:expat)))
+
+
 (define-public python-gguf
   (package
     (name "python-gguf")

base-commit: e13f7d48e5b989f5dbd27c19ac81989ec6b3ec6e
-- 
2.34.1

[bug#72027] [PATCH v2] gnu: Add whisper-cpp.

[Andy Tai]

* gnu/packages/machine-learning.scm (whisper-cpp): New variable.

Change-Id: If191f434a3f66b16afdc702069eaed1ca22e2427
---
 gnu/packages/machine-learning.scm | 33 +++++++++++++++++++++++++++++++
 1 file changed, 33 insertions(+)

diff --git a/gnu/packages/machine-learning.scm b/gnu/packages/machine-learning.scm
index 8ca9a6376e..2af3c188c2 100644
--- a/gnu/packages/machine-learning.scm
+++ b/gnu/packages/machine-learning.scm
@@ -28,6 +28,7 @@
 ;;; Copyright © 2024 Timothee Mathieu <timothee.mathieu@inria.fr>
 ;;; Copyright © 2024 Spencer King <spencer.king@geneoscopy.com>
 ;;; Copyright © 2024 David Elsing <david.elsing@posteo.net>
+;;; Copyright © 2024 Andy Tai <atai@atai.org>
 ;;;
 ;;; This file is part of GNU Guix.
 ;;;
@@ -117,6 +118,7 @@ (define-module (gnu packages machine-learning)
   #:use-module (gnu packages rdf)
   #:use-module (gnu packages regex)
   #:use-module (gnu packages rpc)
+  #:use-module (gnu packages sdl)
   #:use-module (gnu packages serialization)
   #:use-module (gnu packages sphinx)
   #:use-module (gnu packages statistics)
@@ -5982,6 +5984,37 @@ (define-public oneapi-dnnl-for-r-torch
        (sha256
         (base32 "1zyw5rd8x346bb7gac9a7x3saviw3zvp6aqz2z1l9sv163vmjfz6"))))))
 
+(define-public whisper-cpp
+  (package
+    (name "whisper-cpp")
+    (version "1.6.2")
+    (source
+     (origin
+       (method git-fetch)
+       (uri (git-reference
+             (url "https://github.com/ggerganov/whisper.cpp")
+             (commit (string-append "v" version))))
+       (file-name (git-file-name name version))
+       (sha256
+        (base32 "01q4j602wkvsf9vw0nsazzgvjppf4fhpy90vqnm9affynyxhi0c4"))))
+    (build-system cmake-build-system)
+    (arguments
+     (list
+      #:configure-flags #~'("-DWHISPER_STANDALONE=TRUE")))
+    ;; "-DWHISPER_FFMPEG=TRUE"  ; TO DO
+
+    ;; (inputs (list ffmpeg openblas sdl2)) ;TO DO:
+    (native-inputs (list pkg-config))
+    (properties '((tunable? . #t))) ;use AVX512, FMA, etc. when available
+    (home-page "https://github.com/ggerganov/whisper.cpp")
+    (synopsis "Port of OpenAI's Whisper model in C/C++")
+    (description
+     "This package is a high-performance inference of OpenAI's
+Whisper automatic speech recognition (ASR) model, implemented in plain C/C++
+without dependencies")
+    (license license:expat)))
+
+
 (define-public python-gguf
   (package
     (name "python-gguf")

base-commit: 71694363f6cc79e44955acbd9f7ca108e658caca
-- 
2.34.1

[bug#72027] [PATCH v3] gnu: Add whisper-cpp.

[Andy Tai]

* gnu/packages/machine-learning.scm (whisper-cpp): New variable.

Change-Id: If191f434a3f66b16afdc702069eaed1ca22e2427
---
 gnu/packages/machine-learning.scm | 34 +++++++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)

diff --git a/gnu/packages/machine-learning.scm b/gnu/packages/machine-learning.scm
index 8ca9a6376e..0650c9620a 100644
--- a/gnu/packages/machine-learning.scm
+++ b/gnu/packages/machine-learning.scm
@@ -28,6 +28,7 @@
 ;;; Copyright © 2024 Timothee Mathieu <timothee.mathieu@inria.fr>
 ;;; Copyright © 2024 Spencer King <spencer.king@geneoscopy.com>
 ;;; Copyright © 2024 David Elsing <david.elsing@posteo.net>
+;;; Copyright © 2024 Andy Tai <atai@atai.org>
 ;;;
 ;;; This file is part of GNU Guix.
 ;;;
@@ -117,6 +118,7 @@ (define-module (gnu packages machine-learning)
   #:use-module (gnu packages rdf)
   #:use-module (gnu packages regex)
   #:use-module (gnu packages rpc)
+  #:use-module (gnu packages sdl)
   #:use-module (gnu packages serialization)
   #:use-module (gnu packages sphinx)
   #:use-module (gnu packages statistics)
@@ -5982,6 +5984,38 @@ (define-public oneapi-dnnl-for-r-torch
        (sha256
         (base32 "1zyw5rd8x346bb7gac9a7x3saviw3zvp6aqz2z1l9sv163vmjfz6"))))))
 
+(define-public whisper-cpp
+  (package
+    (name "whisper-cpp")
+    (version "1.6.2")
+    (source
+     (origin
+       (method git-fetch)
+       (uri (git-reference
+             (url "https://github.com/ggerganov/whisper.cpp")
+             (commit (string-append "v" version))))
+       (file-name (git-file-name name version))
+       (sha256
+        (base32 "01q4j602wkvsf9vw0nsazzgvjppf4fhpy90vqnm9affynyxhi0c4"))))
+    (build-system cmake-build-system)
+    (arguments
+     (list
+      #:configure-flags #~'("-DWHISPER_STANDALONE=TRUE"
+                            "-DWHISPER_SDL2=TRUE")))
+    ;; "-DWHISPER_FFMPEG=TRUE"  ; TO DO
+
+    (inputs (list sdl2)) ; ffmpeg openblas  ;TO DO:
+    (native-inputs (list pkg-config))
+    (properties '((tunable? . #t))) ;use AVX512, FMA, etc. when available
+    (home-page "https://github.com/ggerganov/whisper.cpp")
+    (synopsis "Port of OpenAI's Whisper model in C/C++")
+    (description
+     "This package is a high-performance inference of OpenAI's
+Whisper automatic speech recognition (ASR) model, implemented in plain C/C++
+without dependencies")
+    (license license:expat)))
+
+
 (define-public python-gguf
   (package
     (name "python-gguf")

base-commit: 71694363f6cc79e44955acbd9f7ca108e658caca
-- 
2.34.1

[bug#72027] [PATCH v4] gnu: Add whisper-cpp.

[Andy Tai]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain; charset=a, Size: 3025 bytes --]

* gnu/packages/machine-learning.scm (whisper-cpp): New variable.

Change-Id: If191f434a3f66b16afdc702069eaed1ca22e2427
---
 gnu/packages/machine-learning.scm | 42 +++++++++++++++++++++++++++++++
 1 file changed, 42 insertions(+)

diff --git a/gnu/packages/machine-learning.scm b/gnu/packages/machine-learning.scm
index 303e123a15..b6ceca6cb0 100644
--- a/gnu/packages/machine-learning.scm
+++ b/gnu/packages/machine-learning.scm
@@ -28,6 +28,7 @@
 ;;; Copyright © 2024 Timothee Mathieu <timothee.mathieu@inria.fr>
 ;;; Copyright © 2024 Spencer King <spencer.king@geneoscopy.com>
 ;;; Copyright © 2024 David Elsing <david.elsing@posteo.net>
+;;; Copyright © 2024 Andy Tai <atai@atai.org>
 ;;;
 ;;; This file is part of GNU Guix.
 ;;;
@@ -117,6 +118,7 @@ (define-module (gnu packages machine-learning)
   #:use-module (gnu packages rdf)
   #:use-module (gnu packages regex)
   #:use-module (gnu packages rpc)
+  #:use-module (gnu packages sdl)
   #:use-module (gnu packages serialization)
   #:use-module (gnu packages sphinx)
   #:use-module (gnu packages statistics)
@@ -5990,6 +5992,46 @@ (define-public oneapi-dnnl-for-r-torch
        (sha256
         (base32 "1zyw5rd8x346bb7gac9a7x3saviw3zvp6aqz2z1l9sv163vmjfz6"))))))
 
+(define-public whisper-cpp
+  (package
+    (name "whisper-cpp")
+    (version "1.6.2")
+    (source
+     (origin
+       (method git-fetch)
+       (uri (git-reference
+             (url "https://github.com/ggerganov/whisper.cpp")
+             (commit (string-append "v" version))))
+       (file-name (git-file-name name version))
+       (sha256
+        (base32 "01q4j602wkvsf9vw0nsazzgvjppf4fhpy90vqnm9affynyxhi0c4"))))
+    (build-system cmake-build-system)
+    (arguments
+     (list
+      #:configure-flags #~'("-DWHISPER_STANDALONE=TRUE" "-DWHISPER_SDL2=TRUE")
+      ;; "-DWHISPER_FFMPEG=TRUE"  ; TO DO
+      #:phases #~(modify-phases %standard-phases
+             #$@(if (not (target-64bit?))
+                  #~(add-after 'unpack 'skip-failing-tests
+                     (lambda _
+                              ;; large model does not fit in RAM in 32-bit system,
+                              ;; disable large model test
+                              `((substitute* "tests/CMakeLists.txt"
+                                  (("LABELS \"large\"")
+                                   "DISABLED true")))))
+                  #~()))))
+    (inputs (list sdl2)) ;ffmpeg openblas  ;TO DO:
+    (native-inputs (list pkg-config))
+    (properties '((tunable? . #t))) ;use AVX512, FMA, etc. when available
+    (home-page "https://github.com/ggerganov/whisper.cpp")
+    (synopsis "Port of OpenAI's Whisper model in C/C++")
+    (description
+     "This package is a high-performance inference of OpenAI's
+Whisper automatic speech recognition (ASR) model, implemented in plain C/C++
+without dependencies")
+    (license license:expat)))
+
+
 (define-public python-gguf
   (package
     (name "python-gguf")

base-commit: 8cbc5e5bf0ca35f12cf5bbf123efbc3b6a57ad44
-- 
2.34.1

[bug#72027] [PATCH v5] gnu: Add whisper-cpp.

[Andy Tai]

* gnu/packages/machine-learning.scm (whisper-cpp): New variable.

Change-Id: If191f434a3f66b16afdc702069eaed1ca22e2427
---
 gnu/packages/machine-learning.scm | 43 +++++++++++++++++++++++++++++++
 1 file changed, 43 insertions(+)

diff --git a/gnu/packages/machine-learning.scm b/gnu/packages/machine-learning.scm
index 3680cab044..0471db8d46 100644
--- a/gnu/packages/machine-learning.scm
+++ b/gnu/packages/machine-learning.scm
@@ -28,6 +28,7 @@
 ;;; Copyright © 2024 Timothee Mathieu <timothee.mathieu@inria.fr>
 ;;; Copyright © 2024 Spencer King <spencer.king@geneoscopy.com>
 ;;; Copyright © 2024 David Elsing <david.elsing@posteo.net>
+;;; Copyright © 2024 Andy Tai <atai@atai.org>
 ;;;
 ;;; This file is part of GNU Guix.
 ;;;
@@ -117,6 +118,7 @@ (define-module (gnu packages machine-learning)
   #:use-module (gnu packages rdf)
   #:use-module (gnu packages regex)
   #:use-module (gnu packages rpc)
+  #:use-module (gnu packages sdl)
   #:use-module (gnu packages serialization)
   #:use-module (gnu packages sphinx)
   #:use-module (gnu packages statistics)
@@ -5990,6 +5992,47 @@ (define-public oneapi-dnnl-for-r-torch
        (sha256
         (base32 "1zyw5rd8x346bb7gac9a7x3saviw3zvp6aqz2z1l9sv163vmjfz6"))))))
 
+(define-public whisper-cpp
+  (package
+    (name "whisper-cpp")
+    (version "1.6.2")
+    (source
+     (origin
+       (method git-fetch)
+       (uri (git-reference
+             (url "https://github.com/ggerganov/whisper.cpp")
+             (commit (string-append "v" version))))
+       (file-name (git-file-name name version))
+       (sha256
+        (base32 "01q4j602wkvsf9vw0nsazzgvjppf4fhpy90vqnm9affynyxhi0c4"))))
+    (build-system cmake-build-system)
+    (arguments
+     (list
+      #:configure-flags #~'("-DWHISPER_STANDALONE=TRUE" "-DWHISPER_SDL2=TRUE")
+      ;; "-DWHISPER_FFMPEG=TRUE"  ; TO DO
+      #:phases #~(modify-phases %standard-phases
+           #$@(if (not (target-64bit?))
+                  '((add-after 'unpack 'skip-failing-tests
+                     (lambda _
+                              ;; 32-bit system
+                              ;; large model does not fit in RAM in 32-bit system,
+                              ;; disable large model test
+                              (substitute* "tests/CMakeLists.txt"
+                                  (("LABELS \"large\"")
+                                   "DISABLED true")))))
+                  '()))))
+    (inputs (list sdl2)) ;ffmpeg openblas  ;TO DO:
+    (native-inputs (list pkg-config))
+    (properties '((tunable? . #t))) ;use AVX512, FMA, etc. when available
+    (home-page "https://github.com/ggerganov/whisper.cpp")
+    (synopsis "Port of OpenAI's Whisper model in C/C++")
+    (description
+     "This package is a high-performance inference of OpenAI's
+Whisper automatic speech recognition (ASR) model, implemented in plain C/C++
+without dependencies")
+    (license license:expat)))
+
+
 (define-public python-gguf
   (package
     (name "python-gguf")

base-commit: 2d6a3799fcda5c017f653c6e96b91964b07a7ee0
-- 
2.34.1

[bug#72027] [PATCH v6] gnu: Add whisper-cpp.

[Andy Tai]

* gnu/packages/machine-learning.scm (whisper-cpp): New variable.

Change-Id: If191f434a3f66b16afdc702069eaed1ca22e2427
---
 gnu/packages/machine-learning.scm | 43 +++++++++++++++++++++++++++++++
 1 file changed, 43 insertions(+)

diff --git a/gnu/packages/machine-learning.scm b/gnu/packages/machine-learning.scm
index 4c5b713cbf..09e83a7636 100644
--- a/gnu/packages/machine-learning.scm
+++ b/gnu/packages/machine-learning.scm
@@ -28,6 +28,7 @@
 ;;; Copyright © 2024 Timothee Mathieu <timothee.mathieu@inria.fr>
 ;;; Copyright © 2024 Spencer King <spencer.king@geneoscopy.com>
 ;;; Copyright © 2024 David Elsing <david.elsing@posteo.net>
+;;; Copyright © 2024 Andy Tai <atai@atai.org>
 ;;;
 ;;; This file is part of GNU Guix.
 ;;;
@@ -117,6 +118,7 @@ (define-module (gnu packages machine-learning)
   #:use-module (gnu packages rdf)
   #:use-module (gnu packages regex)
   #:use-module (gnu packages rpc)
+  #:use-module (gnu packages sdl)
   #:use-module (gnu packages serialization)
   #:use-module (gnu packages sphinx)
   #:use-module (gnu packages statistics)
@@ -5937,6 +5939,47 @@ (define-public oneapi-dnnl-for-r-torch
        (sha256
         (base32 "1zyw5rd8x346bb7gac9a7x3saviw3zvp6aqz2z1l9sv163vmjfz6"))))))
 
+(define-public whisper-cpp
+  (package
+    (name "whisper-cpp")
+    (version "1.6.2")
+    (source
+     (origin
+       (method git-fetch)
+       (uri (git-reference
+             (url "https://github.com/ggerganov/whisper.cpp")
+             (commit (string-append "v" version))))
+       (file-name (git-file-name name version))
+       (sha256
+        (base32 "01q4j602wkvsf9vw0nsazzgvjppf4fhpy90vqnm9affynyxhi0c4"))))
+    (build-system cmake-build-system)
+    (arguments
+     (list
+      #:configure-flags #~'("-DWHISPER_STANDALONE=TRUE" "-DWHISPER_SDL2=TRUE")
+      ;; "-DWHISPER_FFMPEG=TRUE"  ; TO DO
+      #:phases #~(modify-phases %standard-phases
+           #$@(if (not (target-64bit?))
+                  '((add-after 'unpack 'skip-failing-tests
+                     (lambda _
+                              ;; 32-bit system
+                              ;; large model does not fit in RAM in 32-bit system,
+                              ;; disable large model test
+                              (substitute* "tests/CMakeLists.txt"
+                                  (("LABELS \"large\"")
+                                   "DISABLED true")))))
+                  '()))))
+    (inputs (list sdl2)) ;ffmpeg openblas  ;TO DO:
+    (native-inputs (list pkg-config))
+    (properties '((tunable? . #t))) ;use AVX512, FMA, etc. when available
+    (home-page "https://github.com/ggerganov/whisper.cpp")
+    (synopsis "Port of OpenAI's Whisper model in C/C++")
+    (description
+     "This package is a high-performance inference of OpenAI's
+Whisper automatic speech recognition (ASR) model, implemented in plain C/C++
+without dependencies")
+    (license license:expat)))
+
+
 (define-public python-gguf
   (package
     (name "python-gguf")

base-commit: 58a839273d1e4fac2b3a0ec456aabdf82deaa124
-- 
2.34.1

[bug#72027] [PATCH v6] gnu: Add whisper-cpp.

[Juliana Sims via Guix-patches via]

Hi Andy,

Thanks for submitting this patch!  I'm suffering from RSI pain at the 
moment and was tasked, as I ease back into working, with packaging this 
and the associated Emacs mode to help avoid RSI issues in the future.  
Seeing there is already a patch up, I am quite excited and thankful I 
don't have to write one myself!

I had planned to review this patch, but I notice there are several 
TODOs.  So instead, I will ask: what is the status of this patch?  Is 
it in a state where it is ready for review, or is there more work to be 
done first?  If it is the latter, what (if anything) can I do to help 
out?

Thanks,
Juli

[bug#72027] [PATCH v6] gnu: Add whisper-cpp.

[Juliana Sims via Guix-patches via]

Well, I know I just asked if this was ready for review, but I figured 
I'd go ahead and do some "low-hanging fruit review" while I wait for 
you to get a chance to respond ;)

Fortunately, most of what I've noticed so far is minor and stylistic.  
There are some bigger, discussion-opening comments near the end.

Firstly, TODO comments are conventionally spelled as one word.  This is 
not grammatically correct, nor is it a requirement, but it is the 
convention and folks searching files for such comments will likely 
search for "TODO" rather than "TO DO".

Secondly, the synopsis should probably either drop the "Port of" or 
change "in" to "to".  I don't think this is a major issue; it just 
flows better like that imo.  I wouldn't oppose merging over this so if 
you prefer it the way it is, that's fine.

Next, the description wants expansion.  It should be three to five 
sentences.  The sentence there now is a fine start, though it needs 
ending punctuation.  Perhaps a sentence explaining how this package can 
be used ("whisper-cpp can be integrated with other programs to provide 
speech-to-text support" or something) and one explaining what makes it 
unique ("because whisper-cpp uses a leading speech recognition model, 
it is able to perform speech recognition rapidly with relatively few 
resources").  You could perhaps mention that it can run on CPU as well 
as GPU, that it offers some integrations with certain hardware, or so 
on.  Whatever you think is important or interesting; these are just 
some ideas :)

Now we enter the discussion-launching comments.

I notice that ggml is vendored.  This one is tricky.  Firstly, there is 
no standalone ggml package (yet; I saw your patch 65284).  Usually, 
Guix only asks that dependencies be unvendored if there is already a 
standalone package for the dependency, so unless and until that is 
merged there is no real issue.  Secondly, though, and this touches on 
the ggml patches, it seems that there are no formal releases of ggml 
yet, and development is happening in the repositories for whisper-cpp 
and llama-cpp as well as its own repository.  It seems these three 
versions of ggml are all slightly out of sync with each other.  It 
would be nice if upstream used git submodules to ensure their work was 
synchronized.  Alas, we as Guix can't do anything about that, and we 
must ensure the packages we offer work correctly.  The inconsistencies 
between these versions of ggml make me think packaging it separately 
would risk breakage.  With that in mind, might it be best to drop the 
standalone ggml patchset and just let llama-cpp and whisper-cpp vendor 
their versions?  While suboptimal because it results in building "the 
same package" multiple times, I would argue that the divergence in the 
code means they are not, in fact, the same package.

Finally, is this package complete?  Looking at the store directory for 
the package, I see headers and the like but no actual models.  Is this 
sufficient for using the inference?  Are client libraries or programs 
supposed to install models themselves?  Or can this package be used to 
generate models as described in the project's README?  If not, should 
it be able to?  I am admittedly fairly ignorant about the machine 
learning ecosystem, so feel free to explain as much as you think may be 
necessary.  My goal in these questions is ensuring users get what they 
expect from this package.

Relatedly, if this package is complete but requires further setup, I 
would strongly support explaining that in the package description.  As 
a user, I've encountered a few packages that require more setup and 
don't mention that they do, and I'm then frustrated and confused when I 
learn this from trying to use the package and then have no support from 
Guix in trying to make things work properly.  (Another step beyond this 
may be offering a system service which performs configuration, but that 
can be a future, separate patch.)

To circle back to what I mentioned in my first email, I would like to 
package the whisper.el Emacs mode[1].  Currently, whisper.el plans to 
install and compile whisper.cpp and its models itself; I think we as 
Guix should make this unnecessary for an imagined future emacs-whisper 
package.

Looking forward to hearing from you,
Juli

[1] https://github.com/natrys/whisper.el

[bug#72027] [PATCH v6] gnu: Add whisper-cpp.

[Andy Tai]

I will take a look of your review. Thanks


On Fri, Aug 23, 2024 at 7:16 AM Juliana Sims <juli@incana.org> wrote:
>
> Hi Andy,
>
> Thanks for submitting this patch!  I'm suffering from RSI pain at the
> moment and was tasked, as I ease back into working, with packaging this
> and the associated Emacs mode to help avoid RSI issues in the future.
> Seeing there is already a patch up, I am quite excited and thankful I
> don't have to write one myself!
>
> I had planned to review this patch, but I notice there are several
> TODOs.  So instead, I will ask: what is the status of this patch?  Is
> it in a state where it is ready for review, or is there more work to be
> done first?  If it is the latter, what (if anything) can I do to help
> out?
>
> Thanks,
> Juli
>
>

[bug#72027] [PATCH v6] gnu: Add whisper-cpp.

[Andy Tai]

Juliana, I have made some minor adjustments per your review, in a new
patch.   I think the patch is usable but not complete; integration
with ffmpeg is missing. So feel free to take over the patch and to
update it with anything you see missing such as information on ML
model.   Generally I do not package ML models together with the
software as a ML model can be tricky being free or not, so I avoid
that.

>
> On Fri, Aug 23, 2024 at 7:16 AM Juliana Sims <juli@incana.org> wrote:
> >
> > Hi Andy,
> >
> > Thanks for submitting this patch!  I'm suffering from RSI pain at the
> > moment and was tasked, as I ease back into working, with packaging this
> > and the associated Emacs mode to help avoid RSI issues in the future.
> > Seeing there is already a patch up, I am quite excited and thankful I
> > don't have to write one myself!
> >
> > I had planned to review this patch, but I notice there are several
> > TODOs.  So instead, I will ask: what is the status of this patch?  Is
> > it in a state where it is ready for review, or is there more work to be
> > done first?  If it is the latter, what (if anything) can I do to help
> > out?
> >
> > Thanks,
> > Juli
> >
> >

[bug#72027] [PATCH v7] gnu: Add whisper-cpp.

[Andy Tai]

* gnu/packages/machine-learning.scm (whisper-cpp): New variable.

Change-Id: If191f434a3f66b16afdc702069eaed1ca22e2427
---
 gnu/packages/machine-learning.scm | 54 +++++++++++++++++++++++++++++++
 1 file changed, 54 insertions(+)

diff --git a/gnu/packages/machine-learning.scm b/gnu/packages/machine-learning.scm
index 4c5b713cbf..bad7033894 100644
--- a/gnu/packages/machine-learning.scm
+++ b/gnu/packages/machine-learning.scm
@@ -28,6 +28,7 @@
 ;;; Copyright © 2024 Timothee Mathieu <timothee.mathieu@inria.fr>
 ;;; Copyright © 2024 Spencer King <spencer.king@geneoscopy.com>
 ;;; Copyright © 2024 David Elsing <david.elsing@posteo.net>
+;;; Copyright © 2024 Andy Tai <atai@atai.org>
 ;;;
 ;;; This file is part of GNU Guix.
 ;;;
@@ -117,6 +118,7 @@ (define-module (gnu packages machine-learning)
   #:use-module (gnu packages rdf)
   #:use-module (gnu packages regex)
   #:use-module (gnu packages rpc)
+  #:use-module (gnu packages sdl)
   #:use-module (gnu packages serialization)
   #:use-module (gnu packages sphinx)
   #:use-module (gnu packages statistics)
@@ -5937,6 +5939,58 @@ (define-public oneapi-dnnl-for-r-torch
        (sha256
         (base32 "1zyw5rd8x346bb7gac9a7x3saviw3zvp6aqz2z1l9sv163vmjfz6"))))))
 
+(define-public whisper-cpp
+  (package
+    (name "whisper-cpp")
+    (version "1.6.2")
+    (source
+     (origin
+       (method git-fetch)
+       (uri (git-reference
+             (url "https://github.com/ggerganov/whisper.cpp")
+             (commit (string-append "v" version))))
+       (file-name (git-file-name name version))
+       (sha256
+        (base32 "01q4j602wkvsf9vw0nsazzgvjppf4fhpy90vqnm9affynyxhi0c4"))))
+    (build-system cmake-build-system)
+    (arguments
+     (list
+      #:configure-flags #~'("-DWHISPER_STANDALONE=TRUE" "-DWHISPER_SDL2=TRUE")
+      ;; "-DWHISPER_FFMPEG=TRUE"  ; TODO
+      #:phases #~(modify-phases %standard-phases
+           #$@(if (not (target-64bit?))
+                  '((add-after 'unpack 'skip-failing-tests
+                     (lambda _
+                              ;; 32-bit system
+                              ;; large model does not fit in RAM in 32-bit system,
+                              ;; disable large model test
+                              (substitute* "tests/CMakeLists.txt"
+                                  (("LABELS \"large\"")
+                                   "DISABLED true")))))
+                  '()))))
+    (inputs (list sdl2)) ;ffmpeg openblas  ;TODO:
+    (native-inputs (list pkg-config))
+    (properties '((tunable? . #t))) ;use AVX512, FMA, etc. when available
+    (home-page "https://github.com/ggerganov/whisper.cpp")
+    (synopsis "OpenAI's Whisper model in C/C++")
+    (description
+     "This package is a high-performance inference of OpenAI's
+Whisper automatic speech recognition (ASR) model, implemented in plain C/C++
+without dependencies, with
+@itemize
+@item AVX intrinsics support for x86 architectures
+@item VSX intrinsics support for POWER architectures
+@item Mixed F16 / F32 precision
+@item 4-bit and 5-bit integer quantization support
+@item Zero memory allocations at runtime
+@item Support for CPU-only inference
+@item Efficient GPU support for NVIDIA
+@item OpenVINO Support
+@item C-style API
+@end itemize")
+    (license license:expat)))
+
+
 (define-public python-gguf
   (package
     (name "python-gguf")

base-commit: de25a21433bf8ac213302c52789af06a37f4f972
-- 
2.34.1