baalajimaestro/ollama
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

handle asymmetric embedding KVs

Browse source
This commit is contained in:
Michael Yang 2024-06-20 09:40:17 -07:00
parent fedf71635e
commit 8e0641a9bf
2 changed files with 35 additions and 9 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

40
llm/ggml.go
View file

@ -69,6 +69,30 @@ func (kv KV) HeadCountKV() uint64 {
return 1
}
func (kv KV) EmbeddingHeadCount() uint64 {
if heads := kv.HeadCount(); heads > 0 {
return kv.EmbeddingLength() / kv.HeadCount()
}
return 0
}
func (kv KV) EmbeddingHeadCountK() uint64 {
if k := kv.u64(fmt.Sprintf("%s.attention.key_length", kv.Architecture())); k > 0 {
return k
}
return kv.EmbeddingHeadCount()
}
func (kv KV) EmbeddingHeadCountV() uint64 {
if v := kv.u64(fmt.Sprintf("%s.attention.value_length", kv.Architecture())); v > 0 {
return v
}
return kv.EmbeddingHeadCount()
}
func (kv KV) GQA() uint64 {
return kv.HeadCount() / kv.HeadCountKV()
}
@ -299,6 +323,9 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
headsKV := llm.KV().HeadCountKV()
vocab := uint64(len(llm.KV()["tokenizer.ggml.tokens"].([]any)))
embeddingHeads := llm.KV().EmbeddingHeadCount()
embeddingHeadsK := llm.KV().EmbeddingHeadCountK()
layers := llm.Tensors().Layers()
switch llm.KV().Architecture() {
@ -308,7 +335,7 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
partialOffload = 4 * batch * embedding
partialOffload += max(
// 4*batch*(4+6*embedding+context*(2*heads)+llm.KV().GQA()),
4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embedding/heads*headsKV),
4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embeddingHeads*headsKV),
4*batch*(embedding+vocab)+embedding*vocab*105/128,
)
@ -316,15 +343,15 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
// mixtral 8x22b
ff := uint64(llm.KV()["llama.feed_forward_length"].(uint32))
partialOffload = max(
3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embedding/heads*headsKV),
4*(context*batch*heads+context*embedding/heads*headsKV+batch*1024+embedding/heads*headsKV*batch),
3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embeddingHeads*headsKV),
4*(context*batch*heads+context*embeddingHeads*headsKV+batch*1024+embeddingHeads*headsKV*batch),
)
} else if ffnGateWeight, ok := layers["blk.0"]["ffn_gate.0.weight"]; ok {
// mixtral 8x7b
ffnGateWeight1 := ffnGateWeight.Shape[1]
fullOffload = 4 * batch * (2 + 3*embedding + context*(1+heads) + 2*headsKV + ffnGateWeight1)
partialOffload = max(
4*batch*(3+embedding/heads*headsKV+embedding+context*(1+heads)+ffnGateWeight1)+(embedding*embedding+3*embedding*headsKV*ffnGateWeight1)*9/16,
4*batch*(3+embeddingHeads*headsKV+embedding+context*(1+heads)+ffnGateWeight1)+(embedding*embedding+3*embedding*headsKV*ffnGateWeight1)*9/16,
4*batch*(1+2*embedding+context*(1+heads))+embedding*(6*context*headsKV/heads+embedding*9/16),
)
}
@ -368,15 +395,14 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
fullOffload,
)
case "deepseek2":
keys := uint64(llm.KV()["deepseek2.attention.key_length"].(uint32))
fullOffload = max(
4*batch*(3*embedding+vocab),
4*batch*(3*embedding+2+context*(1+headsKV)+2*keys*headsKV),
4*batch*(3*embedding+2+context*(1+headsKV)+2*embeddingHeadsK*headsKV),
)
partialOffload = max(
4*batch*(3*embedding+vocab)+embedding*vocab*105/128,
4*batch*(2*embedding+1+2*keys*headsKV+context+context*headsKV)+4*keys*context*headsKV+embedding*keys*headsKV*9/16,
4*batch*(2*embedding+1+2*embeddingHeadsK*headsKV+context+context*headsKV)+4*embeddingHeadsK*context*headsKV+embedding*embeddingHeadsK*headsKV*9/16,
)
}

4
llm/memory.go
View file

@ -115,8 +115,8 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
slog.Warn("model missing blk.0 layer size")
}
// fp16 k,v = (1 (k) + 1 (v)) * sizeof(float16) * n_ctx * n_layer * n_embd / n_head * n_head_kv
var kv uint64 = 2 * 2 * uint64(opts.NumCtx) * ggml.KV().BlockCount() * ggml.KV().EmbeddingLength() / ggml.KV().HeadCount() * ggml.KV().HeadCountKV()
// fp16 k,v = sizeof(float16) * n_ctx * n_layer * (n_embd_head_k + n_embd_head_v) * n_head_kv
var kv uint64 = 2 * uint64(opts.NumCtx) * ggml.KV().BlockCount() * (ggml.KV().EmbeddingHeadCountK() + ggml.KV().EmbeddingHeadCountV()) * ggml.KV().HeadCountKV()
// KV is proportional to the number of layers
layerSize += kv / ggml.KV().BlockCount()