matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-10 07:46:59 +00:00
Code Issues Packages Projects Releases Wiki Activity

model: Update encoder cache to use multimodal input processing handler

Browse Source 
The encoder cache needs to know the position of images in the input
stream so that it knows when to delete them. Previously images didn't
have a position, so we implied one by breaking batches before an
image and then assuming the image was in the first position. However,
multimodal objects are now given explicit positions in the input
stream, so we can use that instead.

Breaking batches was also a way to simulate a cross attention mask
for mllama. However, given that it only supports a single sequence
and a single image, this mask doesn't serve any real purpose.
Removing the batch break does not appear to affect the quality of
the output.

Most of this is simply moving the input data structures to a new
package to avoid import cycles.
This commit is contained in:
Jesse Gross
2025-03-08 15:45:31 -08:00
committed by Jesse Gross
parent 4614fafae0
commit a1cda80bcb
13 changed files with 157 additions and 160 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
runner/ollamarunner/cache.go
View File

@@ -10,6 +10,7 @@ import (
"github.com/ollama/ollama/kvcache"
"github.com/ollama/ollama/ml"
"github.com/ollama/ollama/model"
"github.com/ollama/ollama/model/input"
)
type InputCache struct {
@@ -79,7 +80,7 @@ type InputCacheSlot struct {
Id int
// Inputs that are stored in the KV cache
Inputs []model.Input
Inputs []input.Input
// is this cache actively being processed as part of a sequence?
InUse bool
@@ -88,7 +89,7 @@ type InputCacheSlot struct {
lastUsed time.Time
}
func (c *InputCache) LoadCacheSlot(prompt []model.Input, cachePrompt bool) (*InputCacheSlot, []model.Input, error) {
func (c *InputCache) LoadCacheSlot(prompt []input.Input, cachePrompt bool) (*InputCacheSlot, []input.Input, error) {
var slot *InputCacheSlot
var numPast int32
var err error
@@ -139,7 +140,7 @@ func (c *InputCache) LoadCacheSlot(prompt []model.Input, cachePrompt bool) (*Inp
return slot, prompt, nil
}
func (c *InputCache) findLongestCacheSlot(prompt []model.Input) (*InputCacheSlot, int32, error) {
func (c *InputCache) findLongestCacheSlot(prompt []input.Input) (*InputCacheSlot, int32, error) {
longest := int32(-1)
var longestSlot *InputCacheSlot
@@ -162,7 +163,7 @@ func (c *InputCache) findLongestCacheSlot(prompt []model.Input) (*InputCacheSlot
return longestSlot, longest, nil
}
func (c *InputCache) findBestCacheSlot(prompt []model.Input) (*InputCacheSlot, int32, error) {
func (c *InputCache) findBestCacheSlot(prompt []input.Input) (*InputCacheSlot, int32, error) {
oldest := time.Now()
var oldestSlot *InputCacheSlot
@@ -198,7 +199,7 @@ func (c *InputCache) findBestCacheSlot(prompt []model.Input) (*InputCacheSlot, i
if longest > 0 && longestSlot != oldestSlot {
slog.Debug("forking cache slot", "src", longestSlot.Id, "dst", oldestSlot.Id, "inputs", longest, "total",
len(longestSlot.Inputs))
oldestSlot.Inputs = make([]model.Input, longest)
oldestSlot.Inputs = make([]input.Input, longest)
copy(oldestSlot.Inputs, longestSlot.Inputs[:longest])
if c.cache != nil {
c.cache.CopyPrefix(longestSlot.Id, oldestSlot.Id, longest)
@@ -208,7 +209,7 @@ func (c *InputCache) findBestCacheSlot(prompt []model.Input) (*InputCacheSlot, i
return oldestSlot, longest, nil
}
func countCommonPrefix(a []model.Input, b []model.Input) int32 {
func countCommonPrefix(a []input.Input, b []input.Input) int32 {
var count int32
for i := range a {