mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-24 07:28:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

perf: build graph for next batch async to keep GPU busy (#11863)

Browse Source 
* perf: build graph for next batch in parallel to keep GPU busy

This refactors the main run loop of the ollama runner to perform the main GPU
intensive tasks (Compute+Floats) in a go routine so we can prepare the next
batch in parallel to reduce the amount of time the GPU stalls waiting for the
next batch of work.

* tests: tune integration tests for ollama engine

This tunes the integration tests to focus more on models supported
by the new engine.
This commit is contained in:
Daniel Hiltgen
2025-08-29 14:20:28 -07:00
committed by GitHub
parent ead4a9a1d0
commit 517807cdf2
20 changed files with 591 additions and 235 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
model/models/llama4/model.go
View File

@@ -134,16 +134,16 @@ type separator struct {
y bool
}
func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
var result []input.Input
func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
var result []*input.Input
for _, inp := range inputs {
if len(inp.Multimodal) == 0 {
result = append(result, inp)
continue
}
var imageInputs []input.Input
imageInputs = append(imageInputs, input.Input{Token: 200080}) // <|image_start|>
var imageInputs []*input.Input
imageInputs = append(imageInputs, &input.Input{Token: 200080}) // <|image_start|>
for i, mm := range inp.Multimodal {
patchesPerChunk := mm.Tensor.Dim(1)
@@ -151,20 +151,20 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
if i < len(inp.Multimodal)-1 {
separator := mm.Data.(*separator)
imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
imageInputs = append(imageInputs, slices.Repeat([]input.Input{{Token: 200092}}, patchesPerChunk-1)...)
imageInputs = append(imageInputs, &input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
imageInputs = append(imageInputs, slices.Repeat([]*input.Input{{Token: 200092}}, patchesPerChunk-1)...)
if separator.x {
imageInputs = append(imageInputs, input.Input{Token: 200084}) // <|tile_x_separator|>
imageInputs = append(imageInputs, &input.Input{Token: 200084}) // <|tile_x_separator|>
}
if separator.y {
imageInputs = append(imageInputs, input.Input{Token: 200085}) // <|tile_y_separator|>
imageInputs = append(imageInputs, &input.Input{Token: 200085}) // <|tile_y_separator|>
}
} else {
imageInputs = append(imageInputs, input.Input{Token: 200090}) // <|image|>
imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
imageInputs = append(imageInputs, slices.Repeat([]input.Input{{Token: 200092}}, patchesPerChunk-1)...)
imageInputs = append(imageInputs, input.Input{Token: 200080}) // <|image_end|>
imageInputs = append(imageInputs, &input.Input{Token: 200090}) // <|image|>
imageInputs = append(imageInputs, &input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
imageInputs = append(imageInputs, slices.Repeat([]*input.Input{{Token: 200092}}, patchesPerChunk-1)...)
imageInputs = append(imageInputs, &input.Input{Token: 200080}) // <|image_end|>
}
}