ollama/model/models/olmo/model.go

package olmo

import (
	"fmt"
	"math"

	"github.com/ollama/ollama/fs"
	"github.com/ollama/ollama/kvcache"
	"github.com/ollama/ollama/ml"
	"github.com/ollama/ollama/ml/nn"
	"github.com/ollama/ollama/ml/nn/rope"
	"github.com/ollama/ollama/model"
	"github.com/ollama/ollama/model/input"
)

const (
	cacheTypeSWA = iota
	cacheTypeCausal
)

type Options struct {
	hiddenSize, numHeads, numKVHeads int
	// headDim, ropeDim                 int
	eps, ropeBase, ropeScale float32

	originalContextLength int
	attnFactor            float32

	ropeType          string
	ropeExtrapolation float32

	ropeBetaFast float32
	ropeBetaSlow float32

	slidingWindowPattern []bool
}

type Model struct {
	model.Base
	model.TextProcessor

	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
	Layers         []Layer       `gguf:"blk"`
	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
	Output         *nn.Linear    `gguf:"output,alt:token_embd"`

	Options
}

func New(c fs.Config) (model.Model, error) {
	vocabulary := model.Vocabulary{
		Values: c.Strings("tokenizer.ggml.tokens"),
		Scores: c.Floats("tokenizer.ggml.scores"),
		Types:  c.Ints("tokenizer.ggml.token_type"),
		Merges: c.Strings("tokenizer.ggml.merges"),
		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(
			[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
			c.Ints("tokenizer.ggml.eos_token_ids")...,
		),
	}

	if c.String("tokenizer.ggml.model") != "gpt2" {
		return nil, model.ErrUnsupportedTokenizer
	}

	var pretokenizers []string
	if c.String("tokenizer.ggml.pre") != "default" {
		pretokenizers = []string{
			"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
		}
	}
	processor := model.NewBytePairEncoding(&vocabulary, pretokenizers...)

	hiddenSize := int(c.Uint("embedding_length"))
	numHeads := int(c.Uint("attention.head_count"))
	numKVHeads := int(c.Uint("attention.head_count_kv"))
	eps := c.Float("attention.layer_norm_rms_epsilon")
	ropeBase := c.Float("rope.freq_base", 1e4)
	ropeScale := c.Float("rope.scaling.factor", 1)
	originalContextLength := int(c.Uint("rope.scaling.original_context_length"))
	attnFactor := c.Float("rope.scaling.attn_factor", 1)
	ropeType := c.String("rope.scaling.type")
	ropeExtrapolation := c.Float("rope.scaling.extrapolation_factor", 1.0)

	fmt.Printf("hiddenSize: %d\n", hiddenSize)
	fmt.Printf("numHeads: %d\n", numHeads)
	fmt.Printf("numKVHeads: %d\n", numKVHeads)
	fmt.Printf("eps: %f\n", eps)
	fmt.Printf("ropeBase: %f\n", ropeBase)
	fmt.Printf("ropeScale: %f\n", ropeScale)
	fmt.Printf("originalContextLength: %d\n", originalContextLength)
	fmt.Printf("attnFactor: %f\n", attnFactor)
	fmt.Printf("ropeType: %s\n", ropeType)
	fmt.Printf("ropeExtrapolation: %f\n", ropeExtrapolation)
	fmt.Printf("sliding_window_pattern: %v\n", c.Bools("attention.sliding_window_pattern"))

	m := Model{
		TextProcessor: processor,
		Layers:        make([]Layer, c.Uint("block_count")),
		Options: Options{
			hiddenSize:            hiddenSize,
			numHeads:              numHeads,
			numKVHeads:            numKVHeads,
			eps:                   eps,
			ropeBase:              ropeBase,
			ropeScale:             ropeScale,
			originalContextLength: originalContextLength,
			attnFactor:            attnFactor,
			ropeType:              ropeType,
			ropeExtrapolation:     ropeExtrapolation,
			ropeBetaFast:          32.0,
			ropeBetaSlow:          1.0,
			slidingWindowPattern:  c.Bools("attention.sliding_window_pattern"),
		},
	}

	m.Cache = kvcache.NewWrapperCache(
		kvcache.NewSWACache(int32(c.Uint("attention.sliding_window")), m.Shift),
		kvcache.NewCausalCache(m.Shift),
	)

	return &m, nil
}

type SelfAttention struct {
	Query       *nn.Linear  `gguf:"attn_q"`
	Key         *nn.Linear  `gguf:"attn_k"`
	Value       *nn.Linear  `gguf:"attn_v"`
	Output      *nn.Linear  `gguf:"attn_output"`
	QNorm       *nn.RMSNorm `gguf:"attn_q_norm"`
	KNorm       *nn.RMSNorm `gguf:"attn_k_norm"`
	RopeFactors ml.Tensor   `gguf:"rope_freqs.weight"`
}

func (m *Model) applyRoPE(ctx ml.Context, states, positions ml.Tensor, ropeDim int, isSWA bool) ml.Tensor {

	var ropeOpts []func(*rope.Options)

	ropeOpts = append(ropeOpts, rope.WithTypeNeoX())

	// Both SWA and non-SWA use beta_fast and beta_slow
	// defaults
	ropeOpts = append(ropeOpts,
		rope.WithBetaFast(m.ropeBetaFast),
		rope.WithBetaSlow(m.ropeBetaSlow),
	)

	// SWA uses freq_scale=1.0, ext_factor=0.0, attn_factor=1.0
	// Non-SWA uses full yarn parameters
	if m.originalContextLength > 0 {
		ropeOpts = append(ropeOpts,
			rope.WithOriginalContextLength(m.originalContextLength),
		)

		// no yarn for swa
		if isSWA {
			ropeOpts = append(ropeOpts,
				rope.WithExtrapolationFactor(0),
				rope.WithAttentionFactor(1.),
			)
		} else {
			ropeOpts = append(ropeOpts,
				rope.WithExtrapolationFactor(m.ropeExtrapolation),
				rope.WithAttentionFactor(m.attnFactor),
			)
		}
	}

	freqScale := float32(1.0)
	if !isSWA {
		freqScale = 1. / m.ropeScale
	}

	return nn.RoPE(ctx, states, positions, ropeDim, m.ropeBase, freqScale, ropeOpts...)
}

func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.Tensor, cache kvcache.Cache, m *Model, isSWA bool) ml.Tensor {
	batchSize := hiddenState.Dim(1)
	headDim := m.hiddenSize / m.numHeads
	ropeDim := headDim

	query := sa.Query.Forward(ctx, hiddenState)
	// double check type
	query = sa.QNorm.Forward(ctx, query, m.eps)
	query = query.Reshape(ctx, headDim, m.numHeads, batchSize)
	//check here

	query = m.applyRoPE(ctx, query, positions, ropeDim, isSWA)

	key := sa.Key.Forward(ctx, hiddenState)
	key = sa.KNorm.Forward(ctx, key, m.eps)
	key = key.Reshape(ctx, headDim, m.numKVHeads, batchSize)
	key = m.applyRoPE(ctx, key, positions, ropeDim, isSWA)

	value := sa.Value.Forward(ctx, hiddenState)
	value = value.Reshape(ctx, headDim, m.numKVHeads, batchSize)

	// check attention scaling as well
	attention := nn.Attention(ctx, query, key, value, 1.0/math.Sqrt(float64(headDim)), cache)
	attention = attention.Reshape(ctx, m.hiddenSize, batchSize)

	return sa.Output.Forward(ctx, attention)
}

func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
	ropeDim := m.hiddenSize / m.numHeads
	isSWA := m.isSWALayer(layer)
	return m.applyRoPE(ctx, key, shift, ropeDim, isSWA), nil
}

type MLP struct {
	Up   *nn.Linear `gguf:"ffn_up"`
	Down *nn.Linear `gguf:"ffn_down"`
	Gate *nn.Linear `gguf:"ffn_gate"`
}

func (mlp *MLP) Forward(ctx ml.Context, hiddenState ml.Tensor, m *Model) ml.Tensor {
	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx, mlp.Up.Forward(ctx, hiddenState))
	return mlp.Down.Forward(ctx, hiddenState)
}

type Layer struct {
	SelfAttention     *SelfAttention
	PostAttentionNorm *nn.RMSNorm `gguf:"post_attention_norm"`
	MLP               *MLP
	PostFFWNorm       *nn.RMSNorm `gguf:"post_ffw_norm"`
}

func (l *Layer) Forward(ctx ml.Context, hiddenState, positions, outputs ml.Tensor, cache kvcache.Cache, m *Model, isSWA bool) ml.Tensor {
	residual := hiddenState

	hiddenState = l.SelfAttention.Forward(ctx, hiddenState, positions, cache, m, isSWA)

	// return hiddenState

	if outputs != nil {
		hiddenState = hiddenState.Rows(ctx, outputs)
		residual = residual.Rows(ctx, outputs)
	}
	// i think this should be after getting the rows?
	hiddenState = l.PostAttentionNorm.Forward(ctx, hiddenState, m.eps)

	hiddenState = hiddenState.Add(ctx, residual)
	residual = hiddenState

	hiddenState = l.MLP.Forward(ctx, hiddenState, m)
	hiddenState = l.PostFFWNorm.Forward(ctx, hiddenState, m.eps)

	return hiddenState.Add(ctx, residual)
}

// Olmo3 has Sliding Window Attention (SWA) 3 out of 4 layers.
func (m *Model) isSWALayer(layerIdx int) bool {
	return m.Options.slidingWindowPattern[layerIdx]
}

func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
	positions := ctx.Input().FromInts(batch.Positions, len(batch.Positions))

	hiddenState := m.TokenEmbedding.Forward(ctx, batch.Inputs)

	for i, layer := range m.Layers {
		m.Cache.SetLayer(i)
		cacheType := cacheTypeSWA

		isSWA := m.isSWALayer(i)
		if !isSWA {
			cacheType = cacheTypeCausal
		}

		wc := m.Cache.(*kvcache.WrapperCache)
		wc.SetLayerType(cacheType)
		// would need to check the cache at the layer instead
		if causal, ok := wc.UnderlyingCache().(*kvcache.Causal); ok {
			// TODO: not sure about the index here
			causal.SetCausal(ctx, kvcache.CausalOptions{Except: []int{}})
		}

		var outputs ml.Tensor
		if i == len(m.Layers)-1 {
			outputs = batch.Outputs
		}

		hiddenState = layer.Forward(ctx, hiddenState, positions, outputs, m.Cache, m, isSWA)

		// return hiddenState, nil
	}

	hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)
	return m.Output.Forward(ctx, hiddenState), nil
}

func init() {
	model.Register("olmo2", New)
}