gsrlabs/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-01-29 07:12:03 +03:00
Code Issues Packages Projects Releases Wiki Activity

glm4moelite: quantize more tensors to q8_0 and avoid double BOS token (#13891)

Browse Source
This commit is contained in:
Jeffrey Morgan
2026-01-24 16:33:54 -08:00
committed by GitHub
parent f3b476c592
commit 16750865d1
2 changed files with 8 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
model/models/glm4moelite/model.go
View File

@@ -246,7 +246,7 @@ func New(c fs.Config) (model.Model, error) {
Values: c.Strings("tokenizer.ggml.tokens"),
Types: c.Ints("tokenizer.ggml.token_type"),
Merges: c.Strings("tokenizer.ggml.merges"),
AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
AddBOS: c.Bool("tokenizer.ggml.add_bos_token", false),
BOS: []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
EOS: append(

7
server/quantization.go
View File

@@ -95,6 +95,13 @@ func getTensorNewType(kv fsggml.KV, qs *quantizeState, newType fsggml.TensorType
// for the 8-expert model, bumping this to Q8_0 trades just ~128MB
newType = fsggml.TensorTypeQ8_0
}
} else if strings.Contains(name, "attn_k_b.weight") ||
strings.Contains(name, "attn_v_b.weight") ||
strings.Contains(name, "attn_kv_a_mqa.weight") ||
strings.Contains(name, "attn_q_a.weight") ||
strings.Contains(name, "attn_q_b.weight") {
// MLA tensors need higher precision to avoid quality degradation
newType = fsggml.TensorTypeQ8_0
} else if strings.Contains(name, "ffn_down") {
iLayer := qs.iFfnDown
n_layer := qs.nFfnDown