feat(llama.cpp): upgrade and use libmtmd (#5379)

* WIP

* wip

* wip

* Make it compile

* Update json.hpp

* this shouldn't be private for now

* Add logs

* Reset auto detected template

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Re-enable grammars

* This seems to be broken - 360a9c98e1 (diff-a18a8e64e12a01167d8e98fc)[…]cccf0d4eed09d76d879L2998-L3207

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Placeholder

* Simplify image loading

* use completion type

* disable streaming

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* correctly return timings

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Remove some debug logging

* Adapt tests

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Keep header

* embedding: do not use oai type

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Sync from server.cpp

* Use utils and json directly from llama.cpp

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Sync with upstream

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* fix: copy json.hpp from the correct location

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* fix: add httplib

* sync llama.cpp

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* Embeddiongs: set OAICOMPAT_TYPE_EMBEDDING

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat: sync with server.cpp by including it

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* make it darwin-compatible

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

---------

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
This commit is contained in:
Ettore Di Giacinto
2025-05-17 16:02:53 +02:00
committed by GitHub
parent 6ef383033b
commit 6d5bde860b
8 changed files with 648 additions and 27490 deletions

View File

@@ -21,7 +21,8 @@ type MultimodalContent struct {
ID int
}
const DefaultMultiModalTemplate = "{{ range .Audio }}[audio-{{.ID}}]{{end}}{{ range .Images }}[img-{{.ID}}]{{end}}{{ range .Video }}[vid-{{.ID}}]{{end}}{{.Text}}"
// https://github.com/ggml-org/llama.cpp/blob/be1d4a13db26750fac702ceb3af88ae4f39dc9f4/tools/mtmd/mtmd.h#L42
const DefaultMultiModalTemplate = "{{ range .Audio }}[audio-{{.ID}}]{{end}}{{ range .Images }}<__image__>{{end}}{{ range .Video }}[vid-{{.ID}}]{{end}}{{.Text}}"
func TemplateMultiModal(templateString string, opts MultiModalOptions, text string) (string, error) {
if templateString == "" {

View File

@@ -20,7 +20,7 @@ var _ = Describe("EvaluateTemplate", func() {
VideosInMessage: 0,
}, "bar")
Expect(err).NotTo(HaveOccurred())
Expect(result).To(Equal("[img-0]bar"))
Expect(result).To(Equal("<__image__>bar"))
})
It("should handle messages with more images correctly", func() {
@@ -33,7 +33,7 @@ var _ = Describe("EvaluateTemplate", func() {
VideosInMessage: 0,
}, "bar")
Expect(err).NotTo(HaveOccurred())
Expect(result).To(Equal("[img-0][img-1]bar"))
Expect(result).To(Equal("<__image__><__image__>bar"))
})
It("should handle messages with more images correctly", func() {
result, err := TemplateMultiModal("", MultiModalOptions{
@@ -45,7 +45,7 @@ var _ = Describe("EvaluateTemplate", func() {
VideosInMessage: 0,
}, "bar")
Expect(err).NotTo(HaveOccurred())
Expect(result).To(Equal("[audio-0][img-2][img-3]bar"))
Expect(result).To(Equal("[audio-0]<__image__><__image__>bar"))
})
It("should handle messages with more images correctly", func() {
result, err := TemplateMultiModal("", MultiModalOptions{
@@ -57,7 +57,7 @@ var _ = Describe("EvaluateTemplate", func() {
VideosInMessage: 0,
}, "bar")
Expect(err).NotTo(HaveOccurred())
Expect(result).To(Equal("[audio-0][img-2]bar"))
Expect(result).To(Equal("[audio-0]<__image__>bar"))
})
It("should handle messages with more images correctly", func() {
result, err := TemplateMultiModal("", MultiModalOptions{