mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 14:26:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

chore(gptoss): cleanup dead code (#12932)

Browse Source
This commit is contained in:
Michael Yang
2025-11-03 11:27:15 -08:00
committed by GitHub
parent 60829f7ec6
commit ce3eb0a315
1 changed files with 4 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
model/models/gptoss/model.go
View File

@@ -32,7 +32,6 @@ func (m *Transformer) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, err
hiddenStates := m.TokenEmbedding.Forward(ctx, batch.Inputs)
positions := ctx.Input().FromInts(batch.Positions, len(batch.Positions))
one := ctx.Input().FromFloats([]float32{1}, 1)
for i, block := range m.TransformerBlocks {
m.Cache.SetLayer(i)
if c, ok := m.Cache.(*kvcache.WrapperCache); ok {
@@ -45,7 +44,7 @@ func (m *Transformer) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, err
outputs = batch.Outputs
}
hiddenStates = block.Forward(ctx, hiddenStates, positions, outputs, one, m.Cache, &m.Options)
hiddenStates = block.Forward(ctx, hiddenStates, positions, outputs, m.Cache, &m.Options)
}
hiddenStates = m.OutputNorm.Forward(ctx, hiddenStates, m.eps)
@@ -90,13 +89,13 @@ type TransformerBlock struct {
MLP *MLPBlock
}
func (d *TransformerBlock) Forward(ctx ml.Context, hiddenStates, positions, outputs, one ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
func (d *TransformerBlock) Forward(ctx ml.Context, hiddenStates, positions, outputs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
hiddenStates = d.Attention.Forward(ctx, hiddenStates, positions, cache, opts)
if outputs != nil {
hiddenStates = hiddenStates.Rows(ctx, outputs)
}
hiddenStates = d.MLP.Forward(ctx, hiddenStates, one, opts)
hiddenStates = d.MLP.Forward(ctx, hiddenStates, opts)
return hiddenStates
}
@@ -177,7 +176,7 @@ type MLPBlock struct {
Down *nn.LinearBatch `gguf:"ffn_down_exps"`
}
func (mlp *MLPBlock) Forward(ctx ml.Context, hiddenStates, one ml.Tensor, opts *Options) ml.Tensor {
func (mlp *MLPBlock) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options) ml.Tensor {
hiddenDim, sequenceLength, batchSize := hiddenStates.Dim(0), hiddenStates.Dim(1), hiddenStates.Dim(2)
residual := hiddenStates