llm: Enable flash attention for mistral3 by default

Although the vision component of multimodal models typically already
call the optimized nn.Attention, it is converted into non-fused
operations. That is because the backend-specific fused kernels may
have requirements, such as padding, and they is performed by the
cache, which vision encoders don't use.

This implements a fallback path in the backend, softening the
requirements into optimizations. In turn, this allows flash attention
to be used for vision encoders, saving a significant amount of VRAM
and improving performance.

cmd/bench/bench.go

@@ -48,8 +48,8 @@ func OutputMetrics(w io.Writer, format string, metrics []Metrics, verbose bool)
 	case "benchstat":
 		if verbose {
 			printHeader := func() {
-				fmt.Printf("sysname: %s\n", runtime.GOOS)
-				fmt.Printf("machine: %s\n", runtime.GOARCH)
+				fmt.Fprintf(w, "sysname: %s\n", runtime.GOOS)
+				fmt.Fprintf(w, "machine: %s\n", runtime.GOARCH)
 			}
 			once.Do(printHeader)
 		}
@@ -147,6 +147,17 @@ func BenchmarkChat(fOpt flagOptions) error {
 		return err
 	}
 
+	var out io.Writer = os.Stdout
+	if fOpt.outputFile != nil && *fOpt.outputFile != "" {
+		f, err := os.OpenFile(*fOpt.outputFile, os.O_CREATE|os.O_WRONLY, 0o644)
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "ERROR: cannot open output file %s: %v\n", *fOpt.outputFile, err)
+			return err
+		}
+		defer f.Close()
+		out = f
+	}
+
 	for _, model := range models {
 		for range *fOpt.epochs {
 			options := make(map[string]interface{})
@@ -241,13 +252,14 @@ func BenchmarkChat(fOpt flagOptions) error {
 				},
 			}
 
-			OutputMetrics(os.Stdout, *fOpt.format, metrics, *fOpt.verbose)
+			OutputMetrics(out, *fOpt.format, metrics, *fOpt.verbose)
 
 			if *fOpt.keepAlive > 0 {
 				time.Sleep(time.Duration(*fOpt.keepAlive*float64(time.Second)) + 200*time.Millisecond)
 			}
 		}
 	}
+
 	return nil
 }
 

convert/convert.go

@@ -208,6 +208,8 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &gptossModel{}
 	case "DeepseekOCRForCausalLM":
 		conv = &deepseekocr{}
+	case "DeepseekV3ForCausalLM":
+		conv = &deepseek2Model{}
 	default:
 		return fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}

convert/convert_deepseek2.go

@@ -0,0 +1,173 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"log/slog"
+	"regexp"
+	"strconv"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type deepseek2Model struct {
+	ModelParameters               // architectures, vocab_size
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	HiddenLayers          uint32  `json:"num_hidden_layers"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	RMSNormEPS            float32 `json:"rms_norm_eps"`
+
+	RopeTheta     float32 `json:"rope_theta"`
+	QKNopeHeadDim uint32  `json:"qk_nope_head_dim"`
+	QKRopeHeadDim uint32  `json:"qk_rope_head_dim"`
+	KVLoraRank    uint32  `json:"kv_lora_rank"`
+	QLoraRank     uint32  `json:"q_lora_rank"`
+	VHeadDim      uint32  `json:"v_head_dim"`
+
+	ExpertCount            uint32  `json:"n_routed_experts"`
+	ExpertSharedCount      uint32  `json:"n_shared_experts"`
+	ExpertIntermediateSize uint32  `json:"moe_intermediate_size"`
+	ExpertUsedCount        uint32  `json:"num_experts_per_tok"`
+	ExpertWeightsNorm      bool    `json:"norm_topk_prob"`
+	ExpertWeightsScale     float32 `json:"routed_scaling_factor"`
+
+	ScoringFunc            string `json:"scoring_func"`
+	LeadingDenseBlockCount uint32 `json:"first_k_dense_replace"`
+
+	RopeScaling struct {
+		Factor                        float32 `json:"factor"`
+		OriginalMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
+		Type                          string  `json:"type"`
+		MScaleAllDim                  float32 `json:"mscale_all_dim"`
+	} `json:"rope_scaling"`
+
+	Architecture string
+}
+
+func (p *deepseek2Model) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "deepseek2"
+	kv["general.type"] = "model"
+	kv["deepseek2.block_count"] = p.HiddenLayers
+
+	numHeads := p.NumAttentionHeads
+	numKVHeads := p.NumKeyValueHeads
+
+	kv["deepseek2.attention.head_count"] = numHeads
+	kv["deepseek2.attention.head_count_kv"] = numKVHeads
+	kv["deepseek2.attention.key_length"] = p.QKNopeHeadDim + p.QKRopeHeadDim
+	kv["deepseek2.attention.kv_lora_rank"] = p.KVLoraRank
+	kv["deepseek2.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["deepseek2.attention.q_lora_rank"] = p.QLoraRank
+	kv["deepseek2.attention.value_length"] = p.VHeadDim
+	kv["deepseek2.context_length"] = p.MaxPositionEmbeddings
+	kv["deepseek2.embedding_length"] = p.HiddenSize
+	kv["deepseek2.expert_count"] = p.ExpertCount
+	kv["deepseek2.expert_feed_forward_length"] = p.ExpertIntermediateSize
+	kv["deepseek2.expert_shared_count"] = p.ExpertSharedCount
+
+	var scoringFunc uint32
+	switch p.ScoringFunc {
+	case "softmax":
+		// not currently supported in the model, but needed for Deepseek-OCR
+		scoringFunc = 1
+	case "sigmoid":
+		scoringFunc = 2
+	}
+	kv["deepseek2.expert_gating_func"] = scoringFunc
+	kv["deepseek2.expert_used_count"] = p.ExpertUsedCount
+	kv["deepseek2.expert_weights_norm"] = p.ExpertWeightsNorm
+	kv["deepseek2.expert_weights_scale"] = p.ExpertWeightsScale
+	kv["deepseek2.feed_forward_length"] = p.IntermediateSize
+	kv["deepseek2.leading_dense_block_count"] = p.LeadingDenseBlockCount
+
+	kv["deepseek2.rope.dimension_count"] = p.QKRopeHeadDim
+	kv["deepseek2.rope.freq_base"] = cmp.Or(p.RopeTheta, 10000.0)
+	kv["deepseek2.rope.scaling.factor"] = p.RopeScaling.Factor
+	kv["deepseek2.rope.scaling.original_context_length"] = p.RopeScaling.OriginalMaxPositionEmbeddings
+	kv["deepseek2.rope.scaling.type"] = p.RopeScaling.Type
+	kv["deepseek2.rope.scaling.yarn_log_multiplier"] = 0.1 * p.RopeScaling.MScaleAllDim
+
+	kv["tokenizer.ggml.pre"] = "deepseek-v3"
+
+	return kv
+}
+
+func (p *deepseek2Model) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"language_model.", "",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.kv_a_proj_with_mqa", "attn_kv_a_mqa",
+		"self_attn.kv_a_layernorm", "attn_kv_a_norm",
+		"self_attn.kv_b_proj", "attn_kv_b",
+		"self_attn.q_a_proj", "attn_q_a",
+		"self_attn.q_a_layernorm", "attn_q_a_norm",
+		"self_attn.q_b_proj", "attn_q_b",
+		"self_attn.o_proj", "attn_output",
+		"post_attention_layernorm", "ffn_norm",
+		"mlp.shared_experts.down_proj", "ffn_down_shexp",
+		"mlp.shared_experts.gate_proj", "ffn_gate_shexp",
+		"mlp.shared_experts.up_proj", "ffn_up_shexp",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"mlp.gate.e_score_correction_bias", "exp_probs_b.bias",
+		"mlp.gate", "ffn_gate_inp",
+	}
+}
+
+func (p *deepseek2Model) Tensors(s []Tensor) (out []*ggml.Tensor) {
+	merges := make([]merge, p.HiddenLayers*3)
+	for i := range p.HiddenLayers {
+		merges[i*3+0] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.gate_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
+		}
+		merges[i*3+1] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.up_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
+		}
+		merges[i*3+2] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.down_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
+		}
+	}
+
+	skipLayer := func(n string, minValue uint32) bool {
+		re := regexp.MustCompile(`^blk\.(\d+)`)
+		matches := re.FindStringSubmatch(n)
+		if matches == nil {
+			return false
+		}
+
+		blkNum, err := strconv.Atoi(matches[1])
+		if err != nil {
+			return false
+		}
+
+		return uint32(blkNum) >= minValue
+	}
+
+	out, s = mergeTensors(s, merges...)
+	for _, t := range s {
+		// skip any additional layers (such as the Multi-Token Prediction layer)
+		if skipLayer(t.Name(), p.HiddenLayers) {
+			slog.Debug("skipping layer", "name", t.Name())
+			continue
+		}
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+	return out
+}

fs/ggml/ggml.go

@@ -831,6 +831,7 @@ func (f GGML) FlashAttention() bool {
 	return slices.Contains([]string{
 		"gemma3",
 		"gptoss", "gpt-oss",
+		"mistral3",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
 	}, f.KV().String("general.architecture"))

ml/backend.go

@@ -233,8 +233,10 @@ type Tensor interface {
 //
 // kqv := value.Mulmat(ctx, kq)
 // return kqv.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
+//
+// cacheConfigApplied indicates whether the optimizations requested through CacheConfig have been performed
 type ScaledDotProductAttention interface {
-	ScaledDotProductAttention(ctx Context, key, value, mask, sinks Tensor, vmla Tensor, scale float64) Tensor
+	ScaledDotProductAttention(ctx Context, key, value, mask, sinks Tensor, vmla Tensor, scale float64, cacheConfigApplied bool) Tensor
 }
 
 type number interface {

ml/backend/ggml/ggml.go

@@ -687,7 +687,7 @@ func (b *Backend) CacheConfig() ml.CacheConfig {
 	if b.flashAttention {
 		return ml.CacheConfig{CachePadding: 256, MaskDType: ml.DTypeF16, MaskBatchPadding: C.GGML_KQ_MASK_PAD}
 	} else {
-		return ml.CacheConfig{CachePadding: 32, PermutedV: true}
+		return ml.CacheConfig{CachePadding: 256, PermutedV: true}
 	}
 }
 
@@ -1645,7 +1645,29 @@ func (t *Tensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
 	}
 }
 
-func (t *Tensor) ScaledDotProductAttention(ctx ml.Context, key, value, mask, sinks ml.Tensor, vmla ml.Tensor, scale float64) ml.Tensor {
+func (t *Tensor) ScaledDotProductAttention(ctx ml.Context, key, value, mask, sinks ml.Tensor, vmla ml.Tensor, scale float64, cacheConfigApplied bool) ml.Tensor {
+	// If the cache didn't help us with required transformations, do them here
+	if !cacheConfigApplied {
+		cacheConfig := t.b.CacheConfig()
+
+		// Padding key and value to CachePadding is a performance optimization, not a requirement, so we don't do it if it wasn't done by the caller
+
+		if cacheConfig.PermutedV {
+			value = value.Permute(ctx, 1, 2, 0, 3).Contiguous(ctx)
+		}
+
+		if mask != nil {
+			padSize := int(pad(C.size_t(mask.Dim(1)), C.size_t(cacheConfig.MaskBatchPadding))) - mask.Dim(1)
+			if padSize > 0 {
+				mask = mask.Pad(ctx, 0, padSize, 0, 0)
+			}
+
+			if mask.DType() != cacheConfig.MaskDType {
+				mask = mask.Cast(ctx, cacheConfig.MaskDType)
+			}
+		}
+	}
+
 	var kqMask *C.struct_ggml_tensor
 	if mask != nil {
 		kqMask = mask.(*Tensor).t

ml/nn/attention.go

@@ -57,10 +57,9 @@ func AttentionWithVMLA(ctx ml.Context, query, key, value, sinks ml.Tensor, vmla
 		key, value, mask = cache.Get(ctx)
 	}
 
-	// Only use the fast SDPA implementation if we have a cache, since that's what
-	// will do any expected backend-specific transformations for us
-	if sdpa, ok := query.(ml.ScaledDotProductAttention); ok && cache != nil {
-		return sdpa.ScaledDotProductAttention(ctx, key, value, mask, sinks, vmla, scale)
+	if sdpa, ok := query.(ml.ScaledDotProductAttention); ok {
+		cacheConfigApplied := cache != nil
+		return sdpa.ScaledDotProductAttention(ctx, key, value, mask, sinks, vmla, scale, cacheConfigApplied)
 	} else {
 		query = query.Permute(ctx, 0, 2, 1, 3)
 		key = key.Permute(ctx, 0, 2, 1, 3)