highperfocused/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-08-24 06:51:13 +02:00
Code Issues Packages Projects Releases Wiki Activity

gpt-oss: disable quantized kv cache (#11929)

Browse Source
This commit is contained in:
Michael Yang
2025-08-15 15:01:05 -07:00
committed by GitHub
parent 026bc29237
commit df335aac09
1 changed files with 5 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
fs/ggml/ggml.go
View File

@@ -752,6 +752,11 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
// SupportsKVCacheType checks if the requested cache type is supported
func (f GGML) SupportsKVCacheType(cacheType string) bool {
if arch := f.KV().Architecture(); slices.Contains([]string{"gptoss", "gpt-oss"}, arch) {
// gpt-oss uses attention with sinks which does not support quantized cache types
slog.Warn("model only supports non-quantized cache types ", "mode", arch)
return cacheType == "f16"
}
return slices.Contains([]string{"f16", "q8_0", "q4_0"}, cacheType)
}