pali112/ollama
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

llm: Remove unneeded warning with flash attention enabled

Browse Source 
If flash attention is enabled without KV cache quanitization, we will
currently always get this warning:
level=WARN source=server.go:226 msg="kv cache type not supported by model" type=""
This commit is contained in:
Jesse Gross
2025-09-09 10:37:28 -07:00
committed by Jesse Gross
parent 5198956372
commit 71cb86af3e
3 changed files with 9 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
fs/ggml/ggml.go
View File

@@ -864,12 +864,16 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
// SupportsKVCacheType checks if the requested cache type is supported
func (f GGML) SupportsKVCacheType(cacheType string) bool {
if cacheType == "" || cacheType == "f16" {
return true
}
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"
slog.Warn("model only supports non-quantized cache types", "model", arch)
return false
}
return slices.Contains([]string{"f16", "q8_0", "q4_0"}, cacheType)
return slices.Contains([]string{"q8_0", "q4_0"}, cacheType)
}
// SupportsFlashAttention checks if the model supports flash attention