vendor/ollama
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

model: add lfm2 architecture and LFM2.5-1.2B-Thinking support (#13792)

Browse Source 
Co-Authored-By: TommyBoiss <165361500+TommyBoiss@users.noreply.github.com>
This commit is contained in:
Jeffrey Morgan
2026-01-20 12:20:53 -08:00
committed by GitHub
parent 31085d5e53
commit 01cf7445f3
18 changed files with 3441 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
server/quantization.go
View File

@@ -198,8 +198,8 @@ func newType(t *fsggml.Tensor, kv fsggml.KV, qs *quantizeState, ftype fsggml.Fil
name := t.Name
quantize := strings.HasSuffix(name, "weight")
// don't quantize vision stuff
quantize = quantize && (!strings.Contains(name, "v.") || strings.Contains(name, "_v."))
// don't quantize vision encoder tensors (named with "v." prefix)
quantize = quantize && !strings.HasPrefix(name, "v.")
quantize = quantize && !strings.Contains(name, "mm.")
// quantize only 2D and 3D tensors (experts)
@@ -219,6 +219,9 @@ func newType(t *fsggml.Tensor, kv fsggml.KV, qs *quantizeState, ftype fsggml.Fil
// NOTE: can't use LLM_TN here because the layer number is not known
quantize = quantize && !strings.Contains(name, "ssm_conv1d.weight")
// do not quantize LFM2's shortconv kernel weights
quantize = quantize && !strings.Contains(name, "shortconv.conv.weight")
// do not quantize RWKV's time_mix_first tensors
quantize = quantize && !strings.Contains(name, "time_mix_first.weight")
quantize = quantize && !strings.Contains(name, "time_mix_w1.weight")