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

ggml: Support heterogeneous KV cache layer sizes in memory estimation

Browse Source 
Gemma3 uses sliding windows for its context on 5/6 layers, significantly
reducing memory usage but leading to uneven usage across layers,
which makes allocation to the correct GPU difficult. We currently
estimate very conservatively by assuming all layers are consistent
at the max size.

Llama3.2-vision is also inconsistent between self attention and cross
attention layers - at moment, we calculate the correct total size
and then average this across layers. In some cases, this may lead
to crashes if a large layer is placed on a GPU sized by the average.

This allows memory estimation to calculate per-layer KV cache size
and take this account when placing layers onto GPUs. We already do
this for weights that vary per-tensor, so this is a logical extension.

Fixes #9730
Fixes #9890
This commit is contained in:
Jesse Gross
2025-03-24 13:39:07 -07:00
committed by Jesse Gross
parent f4f0992b6e
commit f66216e399
5 changed files with 49 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
fs/ggml/ggml.go
View File

@@ -413,7 +413,7 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
}, offset, nil
}
func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialOffload, fullOffload uint64) {
func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string) (kv []uint64, partialOffload, fullOffload uint64) {
embedding := f.KV().EmbeddingLength()
heads := f.KV().HeadCount()
headsKV := f.KV().HeadCountKV()
@@ -426,7 +426,10 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
layers := f.Tensors().GroupLayers()
bytesPerElement := kvCacheBytesPerElement(kvCacheType)
kv = uint64(float64(context*f.KV().BlockCount()*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
kv = make([]uint64, f.KV().BlockCount())
for i := range kv {
kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
}
switch f.KV().Architecture() {
case "llama":
@@ -460,16 +463,14 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
case "mllama":
var visionTokens, tiles uint64 = 1601, 4
if crossAttentionLayers, ok := f.KV()["mllama.attention.cross_attention_layers"].(*array); ok {
kv = headsKV *
(embeddingHeadsK + embeddingHeadsV) * // one for K, one for V
(2* // sizeof(float16)
(f.KV().BlockCount()-uint64(crossAttentionLayers.size))* // num non-cross attention layers
context +
4* // sizeof(float32)
uint64(crossAttentionLayers.size)* // num cross attention layers
visionTokens*
tiles)
crossAttentionLayers := f.KV().Uints("attention.cross_attention_layers")
for i := range kv {
if slices.Contains(crossAttentionLayers, uint32(i)) {
kv[i] = headsKV * (embeddingHeadsK + embeddingHeadsV) *
4 * // sizeof(float32)
visionTokens *
tiles
}
}
fullOffload = max(
@@ -505,6 +506,20 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
4*embeddingHeadsK*context*8+
embedding*embeddingHeadsK*heads*9/16,
)
// Gemma2 also has sliding window attention but we only have an optimized implementation in the Ollama
// engine. Gemma3 always uses the Ollama engine.
if f.KV().Architecture() == "gemma3" {
const gemma3GlobalCacheCount = 6
slidingWindow := (uint64(numParallel) * uint64(f.KV().Uint("attention.sliding_window"))) + batch
for i := range kv {
// Every 6th layer is a global layer, which is the full context size that has already been set. The other
// layers are the smaller local (sliding) layers.
if (i+1)%gemma3GlobalCacheCount != 0 {
kv[i] = uint64(float64(slidingWindow*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
}
}
}
case "command-r":
fullOffload = max(
4*batch*(embedding+vocab),