mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 22:33:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

GGML update to ec98e2002 (#13451)

Browse Source 
* Revert "add support for NVIDIA Nemotron 3 Nano"

This reverts commit e7d2ae9d69421012e9a8765c06a3fdf0e45b12f3.

* GGML update to 380b4c984

Remove MaskBatchPadding as GGML_KQ_MASK_PAD is no longer present (no
padding required)

* update to c45f89d55

* ec98e2002

solar pro needed more adjusting - needs verification

* review comments
This commit is contained in:
Daniel Hiltgen
2025-12-17 13:13:55 -08:00
committed by GitHub
parent 1c094038bc
commit 49a9c9ba6a
127 changed files with 8128 additions and 6710 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
kvcache/causal.go
View File

@@ -140,10 +140,6 @@ func (c *Causal) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity
c.config.CachePadding = 1
}
if c.config.MaskBatchPadding == 0 {
c.config.MaskBatchPadding = 1
}
if c.config.MaskDType == ml.DTypeOther {
c.config.MaskDType = ml.DTypeF32
}
@@ -364,15 +360,12 @@ func roundUp(length, pad int) int {
// token in the history should apply. This is based on both the sequence and causality (the
// position of the history is not ahead of the token in the batch).
func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
// Align and pad the two dimensions as required by the backend
batchSize := roundUp(c.curBatchSize, c.config.MaskBatchPadding)
c.curCellRange.min = roundDown(c.curCellRange.min, c.config.CachePadding)
c.curCellRange.max = roundUp(c.curCellRange.max+1, c.config.CachePadding) - 1
length := c.curCellRange.max - c.curCellRange.min + 1
mask := make([]float32, batchSize*length)
mask := make([]float32, c.curBatchSize*length)
for i := range c.curBatchSize {
enabled := !slices.Contains(c.opts.Except, i)
@@ -386,13 +379,7 @@ func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
}
}
// Mask out any padding tokens we added. For padding that we added to the cache history, this
// has already been masked out because the sequence doesn't match.
for i := c.curBatchSize * length; i < len(mask); i++ {
mask[i] = float32(math.Inf(-1))
}
maskTensor := ctx.Input().FromFloats(mask, length, batchSize)
maskTensor := ctx.Input().FromFloats(mask, length, c.curBatchSize)
if c.config.MaskDType != ml.DTypeF32 {
maskTensor = maskTensor.Cast(ctx, c.config.MaskDType)