mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 14:26:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

model: ministral w/ llama4 scaling (#13292)

Browse Source 
This change:

* fixes rope scaling in the mistral converter
* updates ministral to include llama4 scaling
* includes a new ministral parser for parsing reasoning and tool calling

---------

Co-authored-by: jmorganca <jmorganca@gmail.com>
This commit is contained in:
Patrick Devine
2025-12-01 23:20:14 -08:00
committed by GitHub
parent 554172759c
commit d3e0a0dee4
9 changed files with 379 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
convert/convert_mistral.go
View File

@@ -29,6 +29,15 @@ type mistral3Model struct {
SlidingWindow *uint32 `json:"sliding_window"`
HiddenAct string `json:"hidden_act"`
VocabSize uint32 `json:"vocab_size"`
RopeParameters struct {
BetaFast float32 `json:"beta_fast"`
BetaSlow float32 `json:"beta_slow"`
Factor float32 `json:"factor"`
ScalingBeta float32 `json:"llama_4_scaling_beta"`
OrigMaxPositionEmbeddings uint32 `json:"original_max_position_embeddings"`
RopeType string `json:"rope_type"`
RopeTheta float32 `json:"rope_theta"`
} `json:"rope_parameters"`
} `json:"text_config"`
VisionModel struct {
NumAttentionHeads uint32 `json:"num_attention_heads"`
@@ -61,8 +70,13 @@ func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
kv["mistral3.attention.layer_norm_rms_epsilon"] = p.TextModel.RMSNormEPS
kv["mistral3.attention.key_length"] = p.TextModel.HeadDim
kv["mistral3.attention.value_length"] = p.TextModel.HeadDim
kv["mistral3.rope.dimension_count"] = p.TextModel.HiddenSize / p.TextModel.NumHiddenLayers
kv["mistral3.rope.freq_base"] = p.TextModel.RopeTheta
kv["mistral3.rope.dimension_count"] = cmp.Or(p.TextModel.HeadDim, p.TextModel.HiddenSize/p.TextModel.NumAttentionHeads)
kv["mistral3.rope.freq_base"] = cmp.Or(p.TextModel.RopeTheta, p.TextModel.RopeParameters.RopeTheta)
if p.TextModel.RopeParameters.OrigMaxPositionEmbeddings > 0 {
kv["mistral3.rope.scaling.original_context_length"] = p.TextModel.RopeParameters.OrigMaxPositionEmbeddings
kv["mistral3.rope.scaling_beta"] = p.TextModel.RopeParameters.ScalingBeta
}
// Vision configuration
kv["mistral3.vision.block_count"] = p.VisionModel.NumHiddenLayers