mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 22:33:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

model: conversion and hyperparameter fixes for ministral and devstral (#13424)

Browse Source
This commit is contained in:
Jeffrey Morgan
2025-12-11 13:04:00 -08:00
committed by GitHub
parent 1c4e85b4df
commit a838421ea3
4 changed files with 250 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
convert/convert_mistral.go
View File

@@ -30,13 +30,15 @@ type mistral3Model struct {
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"`
BetaFast float32 `json:"beta_fast"`
BetaSlow float32 `json:"beta_slow"`
Factor float32 `json:"factor"`
Llama4ScalingBeta *float32 `json:"llama_4_scaling_beta"`
OrigMaxPositionEmbeddings uint32 `json:"original_max_position_embeddings"`
RopeType string `json:"rope_type"`
RopeTheta float32 `json:"rope_theta"`
Mscale *float32 `json:"mscale"`
MscaleAllDim *float32 `json:"mscale_all_dim"`
} `json:"rope_parameters"`
} `json:"text_config"`
VisionModel struct {
@@ -50,6 +52,9 @@ type mistral3Model struct {
HeadDim uint32 `json:"head_dim"`
HiddenAct string `json:"hidden_act"`
RopeTheta float32 `json:"rope_theta"`
RopeParameters struct {
RopeTheta float32 `json:"rope_theta"`
} `json:"rope_parameters"`
} `json:"vision_config"`
MultiModalProjectorBias bool `json:"multimodal_projector_bias"`
ProjectorHiddenAct string `json:"projector_hidden_act"`
@@ -72,10 +77,22 @@ func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
kv["mistral3.attention.value_length"] = p.TextModel.HeadDim
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)
kv["mistral3.rope.scaling.factor"] = p.TextModel.RopeParameters.Factor
kv["mistral3.rope.scaling.type"] = p.TextModel.RopeParameters.RopeType
kv["mistral3.rope.scaling.beta_fast"] = p.TextModel.RopeParameters.BetaFast
kv["mistral3.rope.scaling.beta_slow"] = p.TextModel.RopeParameters.BetaSlow
if p.TextModel.RopeParameters.Mscale != nil {
kv["mistral3.rope.scaling.mscale"] = *p.TextModel.RopeParameters.Mscale
}
if p.TextModel.RopeParameters.MscaleAllDim != nil {
kv["mistral3.rope.scaling.mscale_all_dim"] = *p.TextModel.RopeParameters.MscaleAllDim
}
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
}
if p.TextModel.RopeParameters.Llama4ScalingBeta != nil {
kv["mistral3.rope.scaling_beta"] = *p.TextModel.RopeParameters.Llama4ScalingBeta
}
// Vision configuration
@@ -88,7 +105,7 @@ func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
kv["mistral3.vision.patch_size"] = p.VisionModel.PatchSize
kv["mistral3.vision.num_channels"] = p.VisionModel.NumChannels
// kv["mistral3.vision.attention.layer_norm_epsilon"] = 1e-05 // Default value
kv["mistral3.vision.rope.freq_base"] = p.VisionModel.RopeTheta
kv["mistral3.vision.rope.freq_base"] = cmp.Or(p.VisionModel.RopeTheta, p.VisionModel.RopeParameters.RopeTheta)
// Multimodal configuration
kv["mistral3.image_token_index"] = p.ImageTokenIndex