mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-23 23:18:26 +00:00
Code Issues Packages Projects Releases Wiki Activity

use split activations when possible (#12293)

Browse Source 
* use ggml_*_split activations when possible

* forward qkv
This commit is contained in:
Michael Yang
2025-09-16 09:51:19 -07:00
committed by GitHub
parent c253433d68
commit ad95d5b30b
16 changed files with 59 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
model/models/llama4/model_text.go
View File

@@ -58,14 +58,14 @@ type TextMLP struct {
}
func (mlp *TextMLP) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOptions) ml.Tensor {
hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenStates))
hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
return mlp.Down.Forward(ctx, hiddenStates)
}
type TextExperts struct {
Gate *nn.Linear `gguf:"ffn_gate_exps"`
Up *nn.Linear `gguf:"ffn_up_exps"`
Down *nn.Linear `gguf:"ffn_down_exps"`
Gate *nn.LinearBatch `gguf:"ffn_gate_exps"`
Up *nn.LinearBatch `gguf:"ffn_up_exps"`
Down *nn.LinearBatch `gguf:"ffn_down_exps"`
}
func (e *TextExperts) Forward(ctx ml.Context, hiddenStates, routerLogits ml.Tensor, opts *TextOptions) ml.Tensor {
@@ -76,9 +76,9 @@ func (e *TextExperts) Forward(ctx ml.Context, hiddenStates, routerLogits ml.Tens
hiddenStates = hiddenStates.Repeat(ctx, 1, opts.numExpertsUsed)
hiddenStates = hiddenStates.Mul(ctx, scores)
upStates := e.Up.Weight.MulmatID(ctx, hiddenStates, experts)
gateStates := e.Gate.Weight.MulmatID(ctx, hiddenStates, experts)
downStates := e.Down.Weight.MulmatID(ctx, upStates.Mul(ctx, gateStates.SILU(ctx)), experts)
upStates := e.Up.Forward(ctx, hiddenStates, experts)
gateStates := e.Gate.Forward(ctx, hiddenStates, experts)
downStates := e.Down.Forward(ctx, upStates.Mul(ctx, gateStates.SILU(ctx)), experts)
nextStates := downStates.View(ctx, 0, hiddenStates.Dim(0), downStates.Stride(2), hiddenStates.Dim(2))
for i := 1; i < opts.numExpertsUsed; i++ {
@@ -96,7 +96,7 @@ type TextSharedExpert struct {
}
func (mlp *TextSharedExpert) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOptions) ml.Tensor {
hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenStates))
hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
return mlp.Down.Forward(ctx, hiddenStates)
}