docs: add molbal/orca-cli to community integrations (#9909)

Gemma3 uses sliding windows for its context on 5/6 layers, significantly
reducing memory usage but leading to uneven usage across layers,
which makes allocation to the correct GPU difficult. We currently
estimate very conservatively by assuming all layers are consistent
at the max size.

Llama3.2-vision is also inconsistent between self attention and cross
attention layers - at moment, we calculate the correct total size
and then average this across layers. In some cases, this may lead
to crashes if a large layer is placed on a GPU sized by the average.

This allows memory estimation to calculate per-layer KV cache size
and take this account when placing layers onto GPUs. We already do
this for weights that vary per-tensor, so this is a logical extension.

Fixes #9730
Fixes #9890

README.md

@@ -394,6 +394,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Reins](https://github.com/ibrahimcetin/reins) (Easily tweak parameters, customize system prompts per chat, and enhance your AI experiments with reasoning model support.)
 - [Ellama](https://github.com/zeozeozeo/ellama) (Friendly native app to chat with an Ollama instance)
 - [screenpipe](https://github.com/mediar-ai/screenpipe) Build agents powered by your screen history
+- [Ollamb](https://github.com/hengkysteen/ollamb) (Simple yet rich in features, cross-platform built with Flutter and designed for Ollama. Try the [web demo](https://hengkysteen.github.io/demo/ollamb/).)
 
 ### Cloud
 
@@ -434,6 +435,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [aichat](https://github.com/sigoden/aichat) All-in-one LLM CLI tool featuring Shell Assistant, Chat-REPL, RAG, AI tools & agents, with access to OpenAI, Claude, Gemini, Ollama, Groq, and more.
 - [PowershAI](https://github.com/rrg92/powershai) PowerShell module that brings AI to terminal on Windows, including support for Ollama
 - [orbiton](https://github.com/xyproto/orbiton) Configuration-free text editor and IDE with support for tab completion with Ollama.
+- [orca-cli](https://github.com/molbal/orca-cli) Ollama Registry CLI Application - Browse, pull and download models from Ollama Registry in your terminal.
 
 ### Apple Vision Pro
 

fs/ggml/ggml.go

@@ -413,7 +413,7 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 	}, offset, nil
 }
 
-func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialOffload, fullOffload uint64) {
+func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string) (kv []uint64, partialOffload, fullOffload uint64) {
 	embedding := f.KV().EmbeddingLength()
 	heads := f.KV().HeadCount()
 	headsKV := f.KV().HeadCountKV()
@@ -426,7 +426,10 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 	layers := f.Tensors().GroupLayers()
 
 	bytesPerElement := kvCacheBytesPerElement(kvCacheType)
-	kv = uint64(float64(context*f.KV().BlockCount()*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+	kv = make([]uint64, f.KV().BlockCount())
+	for i := range kv {
+		kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+	}
 
 	switch f.KV().Architecture() {
 	case "llama":
@@ -460,16 +463,14 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 	case "mllama":
 		var visionTokens, tiles uint64 = 1601, 4
 
-		if crossAttentionLayers, ok := f.KV()["mllama.attention.cross_attention_layers"].(*array); ok {
-			kv = headsKV *
-				(embeddingHeadsK + embeddingHeadsV) * // one for K, one for V
-				(2* // sizeof(float16)
-					(f.KV().BlockCount()-uint64(crossAttentionLayers.size))* // num non-cross attention layers
-					context +
-					4* // sizeof(float32)
-						uint64(crossAttentionLayers.size)* // num cross attention layers
-						visionTokens*
-						tiles)
+		crossAttentionLayers := f.KV().Uints("attention.cross_attention_layers")
+		for i := range kv {
+			if slices.Contains(crossAttentionLayers, uint32(i)) {
+				kv[i] = headsKV * (embeddingHeadsK + embeddingHeadsV) *
+					4 * // sizeof(float32)
+					visionTokens *
+					tiles
+			}
 		}
 
 		fullOffload = max(
@@ -505,6 +506,20 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 				4*embeddingHeadsK*context*8+
 				embedding*embeddingHeadsK*heads*9/16,
 		)
+
+		// Gemma2 also has sliding window attention but we only have an optimized implementation in the Ollama
+		// engine. Gemma3 always uses the Ollama engine.
+		if f.KV().Architecture() == "gemma3" {
+			const gemma3GlobalCacheCount = 6
+			slidingWindow := (uint64(numParallel) * uint64(f.KV().Uint("attention.sliding_window"))) + batch
+			for i := range kv {
+				// Every 6th layer is a global layer, which is the full context size that has already been set. The other
+				// layers are the smaller local (sliding) layers.
+				if (i+1)%gemma3GlobalCacheCount != 0 {
+					kv[i] = uint64(float64(slidingWindow*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+				}
+			}
+		}
 	case "command-r":
 		fullOffload = max(
 			4*batch*(embedding+vocab),

grammar/bench_test.go

@@ -1,22 +0,0 @@
-//go:build go1.24
-
-package grammar
-
-import "testing"
-
-func BenchmarkFromSchema(b *testing.B) {
-	for tt := range testCases(b) {
-		b.Run("", func(b *testing.B) {
-			s := []byte(tt.schema)
-
-			b.ReportAllocs()
-			for b.Loop() {
-				_, err := FromSchema(nil, s)
-				if err != nil {
-					b.Fatalf("GrammarFromSchema: %v", err)
-				}
-			}
-		})
-		return
-	}
-}

grammar/grammar.go

@@ -1,227 +0,0 @@
-package grammar
-
-import (
-	"bytes"
-	"encoding/json"
-	"fmt"
-	"iter"
-	"strconv"
-
-	"github.com/ollama/ollama/grammar/jsonschema"
-)
-
-const jsonTerms = `
-# Unicode
-#
-# Unicode characters can be specified directly in the grammar, for example
-# hiragana ::= [ぁ-ゟ], or with escapes: 8-bit (\xXX), 16-bit (\uXXXX) or 32-bit
-# (\UXXXXXXXX).
-unicode ::= \x{hex}{2} | \u{hex}{4} | \U{hex}{8}
-
-# JSON grammar from RFC 7159
-null    ::= "null"
-object  ::= "{" (kv ("," kv)*)? "}"
-array   ::= "[" (value ("," value)*)? "]"
-kv      ::= string ":" value
-integer ::= "0" | [1-9] [0-9]*
-number  ::= "-"? integer frac? exp?
-frac    ::= "." [0-9]+
-exp     ::= ("e" | "E") ("+" | "-") [0-9]+
-string  ::= "\"" char* "\""
-escape  ::= ["/" | "b" | "f" | "n" | "r" | "t" | unicode]
-char    ::= [^"\\] | escape
-space   ::= (" " | "\t" | "\n" | "\r")*
-hex     ::= [0-9] | [a-f] | [A-F]
-boolean ::= "true" | "false"
-value   ::= object | array | string | number | boolean | "null"
-
-# User-defined
-`
-
-// FromSchema generates a grammar from a JSON schema.
-func FromSchema(buf []byte, jsonSchema []byte) ([]byte, error) {
-	var s *jsonschema.Schema
-	if err := json.Unmarshal(jsonSchema, &s); err != nil {
-		return nil, err
-	}
-
-	var g builder
-
-	// "root" is the only rule that is guaranteed to exist, so we start
-	// with its length for padding, and then adjust it as we go.
-	g.pad = len("root")
-	for id := range dependencies("root", s) {
-		g.pad = max(g.pad, len(id))
-	}
-
-	g.b.WriteString(jsonTerms)
-
-	ids := make(map[*jsonschema.Schema]string)
-	for id, s := range dependencies("root", s) {
-		ids[s] = id
-		g.define(id)
-		if err := fromSchema(&g, ids, s); err != nil {
-			return nil, err
-		}
-	}
-	g.define("root")
-	if err := fromSchema(&g, ids, s); err != nil {
-		return nil, err
-	}
-	g.define("") // finalize the last rule
-	return g.b.Bytes(), nil
-}
-
-func fromSchema(g *builder, ids map[*jsonschema.Schema]string, s *jsonschema.Schema) error {
-	switch typ := s.EffectiveType(); typ {
-	case "array":
-		if len(s.PrefixItems) == 0 && s.Items == nil {
-			g.u("array")
-		} else {
-			g.q("[")
-			for i, s := range s.PrefixItems {
-				if i > 0 {
-					g.q(",")
-				}
-				g.u(ids[s])
-			}
-			if s.Items != nil {
-				g.u("(")
-				if len(s.PrefixItems) > 0 {
-					g.q(",")
-				}
-				g.u(ids[s.Items])
-				g.u(")*")
-			}
-			g.q("]")
-		}
-	case "object":
-		if len(s.Properties) == 0 {
-			g.u("object")
-		} else {
-			g.q("{")
-			for i, p := range s.Properties {
-				name := ids[p]
-				if i > 0 {
-					g.q(",")
-				}
-				g.q(p.Name)
-				g.q(":")
-				g.u(name)
-			}
-			g.q("}")
-		}
-	case "number":
-		buildConstrainedNumber(g, s)
-	case "string":
-		if len(s.Enum) == 0 {
-			g.u("string")
-		} else {
-			g.u("(")
-			for i, e := range s.Enum {
-				if i > 0 {
-					g.q("|")
-				}
-				g.q(string(e))
-			}
-			g.u(")")
-		}
-	case "boolean", "value", "null", "integer":
-		g.u(typ)
-	default:
-		return fmt.Errorf("%s: unsupported type %q", s.Name, typ)
-	}
-	return nil
-}
-
-// dependencies returns a sequence of all child dependencies of the schema in
-// post-order.
-//
-// The first value is the id/pointer to the dependency, and the second value
-// is the schema.
-func dependencies(id string, s *jsonschema.Schema) iter.Seq2[string, *jsonschema.Schema] {
-	return func(yield func(string, *jsonschema.Schema) bool) {
-		for i, p := range s.Properties {
-			id := fmt.Sprintf("%s_%d", id, i)
-			for did, d := range dependencies(id, p) {
-				if !yield(did, d) {
-					return
-				}
-			}
-			if !yield(id, p) {
-				return
-			}
-		}
-		for i, p := range s.PrefixItems {
-			id := fmt.Sprintf("tuple_%d", i)
-			for did, d := range dependencies(id, p) {
-				id := fmt.Sprintf("%s_%s", id, did)
-				if !yield(id, d) {
-					return
-				}
-			}
-			if !yield(id, p) {
-				return
-			}
-		}
-		if s.Items != nil {
-			id := fmt.Sprintf("%s_tuple_%d", id, len(s.PrefixItems))
-			for did, d := range dependencies(id, s.Items) {
-				if !yield(did, d) {
-					return
-				}
-			}
-			if !yield(id, s.Items) {
-				return
-			}
-		}
-	}
-}
-
-type builder struct {
-	b     bytes.Buffer
-	pad   int
-	rules int
-	items int
-}
-
-// define terminates the current rule, if any, and then either starts a new
-// rule or does nothing else if the name is empty.
-func (b *builder) define(name string) {
-	if b.rules > 0 {
-		b.b.WriteString(";\n")
-	}
-	if name == "" {
-		return
-	}
-	fmt.Fprintf(&b.b, "% -*s", b.pad, name)
-	b.b.WriteString(" ::=")
-	b.rules++
-	b.items = 0
-}
-
-// quote appends a terminal to the current rule.
-func (b *builder) q(s string) {
-	if b.items > 0 {
-		b.b.WriteString(" ")
-	}
-	b.b.WriteString(" ")
-	b.b.WriteString(strconv.Quote(s))
-}
-
-// u appends a non-terminal to the current rule.
-func (b *builder) u(s string) {
-	if b.items > 0 {
-		b.b.WriteString(" ")
-	}
-	b.b.WriteString(" ")
-	b.b.WriteString(s)
-}
-
-func buildConstrainedNumber(b *builder, s *jsonschema.Schema) {
-	if s.Minimum == 0 && s.Maximum == 0 {
-		b.u("TODO")
-	} else {
-		b.u("number")
-	}
-}

grammar/grammar_test.go

@@ -1,75 +0,0 @@
-package grammar
-
-import (
-	"bufio"
-	"cmp"
-	"iter"
-	"strings"
-	"testing"
-
-	_ "embed"
-
-	"github.com/ollama/ollama/grammar/internal/diff"
-)
-
-func TestFromSchema(t *testing.T) {
-	for tt := range testCases(t) {
-		t.Run(tt.name, func(t *testing.T) {
-			g, err := FromSchema(nil, []byte(tt.schema))
-			if err != nil {
-				t.Fatalf("FromSchema: %v", err)
-			}
-			got := string(g)
-			got = strings.TrimPrefix(got, jsonTerms)
-			if got != tt.want {
-				t.Logf("schema:\n%s", tt.schema)
-				t.Fatal(string(diff.Diff("got", []byte(got), "want", []byte(tt.want))))
-			}
-		})
-	}
-}
-
-type testCase struct {
-	name   string
-	schema string
-	want   string
-}
-
-//go:embed testdata/schemas.txt
-var tests string
-
-func testCases(t testing.TB) iter.Seq[testCase] {
-	t.Helper()
-	return func(yield func(testCase) bool) {
-		t.Helper()
-		sc := bufio.NewScanner(strings.NewReader(tests))
-		name := ""
-		for sc.Scan() {
-			line := strings.TrimSpace(sc.Text())
-			if line == "" {
-				name = ""
-				continue
-			}
-			if line[0] == '#' {
-				name = cmp.Or(name, strings.TrimSpace(line[1:]))
-				continue
-			}
-			s := sc.Text()
-			g := ""
-			for sc.Scan() {
-				line = strings.TrimSpace(sc.Text())
-				if line == "" || line[0] == '#' {
-					break
-				}
-				g += sc.Text() + "\n"
-			}
-			if !yield(testCase{name, s, g}) {
-				return
-			}
-			name = strings.TrimSpace(strings.TrimPrefix(line, "#"))
-		}
-		if err := sc.Err(); err != nil {
-			t.Fatalf("error reading tests: %v", err)
-		}
-	}
-}

grammar/internal/diff/diff.go

@@ -1,261 +0,0 @@
-// Copyright 2022 The Go Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-package diff
-
-import (
-	"bytes"
-	"fmt"
-	"sort"
-	"strings"
-)
-
-// A pair is a pair of values tracked for both the x and y side of a diff.
-// It is typically a pair of line indexes.
-type pair struct{ x, y int }
-
-// Diff returns an anchored diff of the two texts old and new
-// in the “unified diff” format. If old and new are identical,
-// Diff returns a nil slice (no output).
-//
-// Unix diff implementations typically look for a diff with
-// the smallest number of lines inserted and removed,
-// which can in the worst case take time quadratic in the
-// number of lines in the texts. As a result, many implementations
-// either can be made to run for a long time or cut off the search
-// after a predetermined amount of work.
-//
-// In contrast, this implementation looks for a diff with the
-// smallest number of “unique” lines inserted and removed,
-// where unique means a line that appears just once in both old and new.
-// We call this an “anchored diff” because the unique lines anchor
-// the chosen matching regions. An anchored diff is usually clearer
-// than a standard diff, because the algorithm does not try to
-// reuse unrelated blank lines or closing braces.
-// The algorithm also guarantees to run in O(n log n) time
-// instead of the standard O(n²) time.
-//
-// Some systems call this approach a “patience diff,” named for
-// the “patience sorting” algorithm, itself named for a solitaire card game.
-// We avoid that name for two reasons. First, the name has been used
-// for a few different variants of the algorithm, so it is imprecise.
-// Second, the name is frequently interpreted as meaning that you have
-// to wait longer (to be patient) for the diff, meaning that it is a slower algorithm,
-// when in fact the algorithm is faster than the standard one.
-func Diff(oldName string, old []byte, newName string, new []byte) []byte {
-	if bytes.Equal(old, new) {
-		return nil
-	}
-	x := lines(old)
-	y := lines(new)
-
-	// Print diff header.
-	var out bytes.Buffer
-	fmt.Fprintf(&out, "diff %s %s\n", oldName, newName)
-	fmt.Fprintf(&out, "--- %s\n", oldName)
-	fmt.Fprintf(&out, "+++ %s\n", newName)
-
-	// Loop over matches to consider,
-	// expanding each match to include surrounding lines,
-	// and then printing diff chunks.
-	// To avoid setup/teardown cases outside the loop,
-	// tgs returns a leading {0,0} and trailing {len(x), len(y)} pair
-	// in the sequence of matches.
-	var (
-		done  pair     // printed up to x[:done.x] and y[:done.y]
-		chunk pair     // start lines of current chunk
-		count pair     // number of lines from each side in current chunk
-		ctext []string // lines for current chunk
-	)
-	for _, m := range tgs(x, y) {
-		if m.x < done.x {
-			// Already handled scanning forward from earlier match.
-			continue
-		}
-
-		// Expand matching lines as far as possible,
-		// establishing that x[start.x:end.x] == y[start.y:end.y].
-		// Note that on the first (or last) iteration we may (or definitely do)
-		// have an empty match: start.x==end.x and start.y==end.y.
-		start := m
-		for start.x > done.x && start.y > done.y && x[start.x-1] == y[start.y-1] {
-			start.x--
-			start.y--
-		}
-		end := m
-		for end.x < len(x) && end.y < len(y) && x[end.x] == y[end.y] {
-			end.x++
-			end.y++
-		}
-
-		// Emit the mismatched lines before start into this chunk.
-		// (No effect on first sentinel iteration, when start = {0,0}.)
-		for _, s := range x[done.x:start.x] {
-			ctext = append(ctext, "-"+s)
-			count.x++
-		}
-		for _, s := range y[done.y:start.y] {
-			ctext = append(ctext, "+"+s)
-			count.y++
-		}
-
-		// If we're not at EOF and have too few common lines,
-		// the chunk includes all the common lines and continues.
-		const C = 3 // number of context lines
-		if (end.x < len(x) || end.y < len(y)) &&
-			(end.x-start.x < C || (len(ctext) > 0 && end.x-start.x < 2*C)) {
-			for _, s := range x[start.x:end.x] {
-				ctext = append(ctext, " "+s)
-				count.x++
-				count.y++
-			}
-			done = end
-			continue
-		}
-
-		// End chunk with common lines for context.
-		if len(ctext) > 0 {
-			n := end.x - start.x
-			if n > C {
-				n = C
-			}
-			for _, s := range x[start.x : start.x+n] {
-				ctext = append(ctext, " "+s)
-				count.x++
-				count.y++
-			}
-			done = pair{start.x + n, start.y + n}
-
-			// Format and emit chunk.
-			// Convert line numbers to 1-indexed.
-			// Special case: empty file shows up as 0,0 not 1,0.
-			if count.x > 0 {
-				chunk.x++
-			}
-			if count.y > 0 {
-				chunk.y++
-			}
-			fmt.Fprintf(&out, "@@ -%d,%d +%d,%d @@\n", chunk.x, count.x, chunk.y, count.y)
-			for _, s := range ctext {
-				out.WriteString(s)
-			}
-			count.x = 0
-			count.y = 0
-			ctext = ctext[:0]
-		}
-
-		// If we reached EOF, we're done.
-		if end.x >= len(x) && end.y >= len(y) {
-			break
-		}
-
-		// Otherwise start a new chunk.
-		chunk = pair{end.x - C, end.y - C}
-		for _, s := range x[chunk.x:end.x] {
-			ctext = append(ctext, " "+s)
-			count.x++
-			count.y++
-		}
-		done = end
-	}
-
-	return out.Bytes()
-}
-
-// lines returns the lines in the file x, including newlines.
-// If the file does not end in a newline, one is supplied
-// along with a warning about the missing newline.
-func lines(x []byte) []string {
-	l := strings.SplitAfter(string(x), "\n")
-	if l[len(l)-1] == "" {
-		l = l[:len(l)-1]
-	} else {
-		// Treat last line as having a message about the missing newline attached,
-		// using the same text as BSD/GNU diff (including the leading backslash).
-		l[len(l)-1] += "\n\\ No newline at end of file\n"
-	}
-	return l
-}
-
-// tgs returns the pairs of indexes of the longest common subsequence
-// of unique lines in x and y, where a unique line is one that appears
-// once in x and once in y.
-//
-// The longest common subsequence algorithm is as described in
-// Thomas G. Szymanski, “A Special Case of the Maximal Common
-// Subsequence Problem,” Princeton TR #170 (January 1975),
-// available at https://research.swtch.com/tgs170.pdf.
-func tgs(x, y []string) []pair {
-	// Count the number of times each string appears in a and b.
-	// We only care about 0, 1, many, counted as 0, -1, -2
-	// for the x side and 0, -4, -8 for the y side.
-	// Using negative numbers now lets us distinguish positive line numbers later.
-	m := make(map[string]int)
-	for _, s := range x {
-		if c := m[s]; c > -2 {
-			m[s] = c - 1
-		}
-	}
-	for _, s := range y {
-		if c := m[s]; c > -8 {
-			m[s] = c - 4
-		}
-	}
-
-	// Now unique strings can be identified by m[s] = -1+-4.
-	//
-	// Gather the indexes of those strings in x and y, building:
-	//	xi[i] = increasing indexes of unique strings in x.
-	//	yi[i] = increasing indexes of unique strings in y.
-	//	inv[i] = index j such that x[xi[i]] = y[yi[j]].
-	var xi, yi, inv []int
-	for i, s := range y {
-		if m[s] == -1+-4 {
-			m[s] = len(yi)
-			yi = append(yi, i)
-		}
-	}
-	for i, s := range x {
-		if j, ok := m[s]; ok && j >= 0 {
-			xi = append(xi, i)
-			inv = append(inv, j)
-		}
-	}
-
-	// Apply Algorithm A from Szymanski's paper.
-	// In those terms, A = J = inv and B = [0, n).
-	// We add sentinel pairs {0,0}, and {len(x),len(y)}
-	// to the returned sequence, to help the processing loop.
-	J := inv
-	n := len(xi)
-	T := make([]int, n)
-	L := make([]int, n)
-	for i := range T {
-		T[i] = n + 1
-	}
-	for i := range n {
-		k := sort.Search(n, func(k int) bool {
-			return T[k] >= J[i]
-		})
-		T[k] = J[i]
-		L[i] = k + 1
-	}
-	k := 0
-	for _, v := range L {
-		if k < v {
-			k = v
-		}
-	}
-	seq := make([]pair, 2+k)
-	seq[1+k] = pair{len(x), len(y)} // sentinel at end
-	lastj := n
-	for i := n - 1; i >= 0; i-- {
-		if L[i] == k && J[i] < lastj {
-			seq[k] = pair{xi[i], yi[J[i]]}
-			k--
-		}
-	}
-	seq[0] = pair{0, 0} // sentinel at start
-	return seq
-}

grammar/internal/diff/diff_test.go

@@ -1,44 +0,0 @@
-// Copyright 2022 The Go Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-package diff
-
-import (
-	"bytes"
-	"path/filepath"
-	"testing"
-
-	"golang.org/x/tools/txtar"
-)
-
-func clean(text []byte) []byte {
-	text = bytes.ReplaceAll(text, []byte("$\n"), []byte("\n"))
-	text = bytes.TrimSuffix(text, []byte("^D\n"))
-	return text
-}
-
-func Test(t *testing.T) {
-	files, _ := filepath.Glob("testdata/*.txt")
-	if len(files) == 0 {
-		t.Fatalf("no testdata")
-	}
-
-	for _, file := range files {
-		t.Run(filepath.Base(file), func(t *testing.T) {
-			a, err := txtar.ParseFile(file)
-			if err != nil {
-				t.Fatal(err)
-			}
-			if len(a.Files) != 3 || a.Files[2].Name != "diff" {
-				t.Fatalf("%s: want three files, third named \"diff\"", file)
-			}
-			diffs := Diff(a.Files[0].Name, clean(a.Files[0].Data), a.Files[1].Name, clean(a.Files[1].Data))
-			want := clean(a.Files[2].Data)
-			if !bytes.Equal(diffs, want) {
-				t.Fatalf("%s: have:\n%s\nwant:\n%s\n%s", file,
-					diffs, want, Diff("have", diffs, "want", want))
-			}
-		})
-	}
-}

grammar/internal/diff/testdata/allnew.txt

@@ -1,13 +0,0 @@
--- old --
--- new --
-a
-b
-c
--- diff --
-diff old new
---- old
-+++ new
-@@ -0,0 +1,3 @@
-+a
-+b
-+c

grammar/internal/diff/testdata/allold.txt

@@ -1,13 +0,0 @@
--- old --
-a
-b
-c
--- new --
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,3 +0,0 @@
--a
--b
--c

grammar/internal/diff/testdata/basic.txt

@@ -1,35 +0,0 @@
-Example from Hunt and McIlroy, “An Algorithm for Differential File Comparison.”
-https://www.cs.dartmouth.edu/~doug/diff.pdf
-
--- old --
-a
-b
-c
-d
-e
-f
-g
--- new --
-w
-a
-b
-x
-y
-z
-e
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,7 +1,7 @@
-+w
- a
- b
--c
--d
-+x
-+y
-+z
- e
--f
--g

grammar/internal/diff/testdata/dups.txt

@@ -1,40 +0,0 @@
--- old --
-a
-
-b
-
-c
-
-d
-
-e
-
-f
--- new --
-a
-
-B
-
-C
-
-d
-
-e
-
-f
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,8 +1,8 @@
- a
- $
--b
--
--c
-+B
-+
-+C
- $
- d
- $

grammar/internal/diff/testdata/end.txt

@@ -1,38 +0,0 @@
--- old --
-1
-2
-3
-4
-5
-6
-7
-eight
-nine
-ten
-eleven
--- new --
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
--- diff --
-diff old new
---- old
-+++ new
-@@ -5,7 +5,6 @@
- 5
- 6
- 7
--eight
--nine
--ten
--eleven
-+8
-+9
-+10

grammar/internal/diff/testdata/eof.txt

@@ -1,9 +0,0 @@
--- old --
-a
-b
-c^D
--- new --
-a
-b
-c^D
--- diff --

grammar/internal/diff/testdata/eof1.txt

@@ -1,18 +0,0 @@
--- old --
-a
-b
-c
--- new --
-a
-b
-c^D
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,3 +1,3 @@
- a
- b
--c
-+c
-\ No newline at end of file

grammar/internal/diff/testdata/eof2.txt

@@ -1,18 +0,0 @@
--- old --
-a
-b
-c^D
--- new --
-a
-b
-c
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,3 +1,3 @@
- a
- b
--c
-\ No newline at end of file
-+c

grammar/internal/diff/testdata/long.txt

@@ -1,62 +0,0 @@
--- old --
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
-11
-12
-13
-14
-14½
-15
-16
-17
-18
-19
-20
--- new --
-1
-2
-3
-4
-5
-6
-8
-9
-10
-11
-12
-13
-14
-17
-18
-19
-20
--- diff --
-diff old new
---- old
-+++ new
-@@ -4,7 +4,6 @@
- 4
- 5
- 6
--7
- 8
- 9
- 10
-@@ -12,9 +11,6 @@
- 12
- 13
- 14
--14½
--15
--16
- 17
- 18
- 19

grammar/internal/diff/testdata/same.txt

@@ -1,5 +0,0 @@
--- old --
-hello world
--- new --
-hello world
--- diff --

grammar/internal/diff/testdata/start.txt

@@ -1,34 +0,0 @@
--- old --
-e
-pi
-4
-5
-6
-7
-8
-9
-10
--- new --
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,5 +1,6 @@
--e
--pi
-+1
-+2
-+3
- 4
- 5
- 6

grammar/internal/diff/testdata/triv.txt

@@ -1,40 +0,0 @@
-Another example from Hunt and McIlroy,
-“An Algorithm for Differential File Comparison.”
-https://www.cs.dartmouth.edu/~doug/diff.pdf
-
-Anchored diff gives up on finding anything,
-since there are no unique lines.
-
--- old --
-a
-b
-c
-a
-b
-b
-a
--- new --
-c
-a
-b
-a
-b
-c
--- diff --
-diff old new
---- old
-+++ new
-@@ -1,7 +1,6 @@
--a
--b
--c
--a
--b
--b
--a
-+c
-+a
-+b
-+a
-+b
-+c

grammar/jsonschema/decode.go

@@ -1,171 +0,0 @@
-package jsonschema
-
-import (
-	"bytes"
-	"encoding/json"
-	"errors"
-)
-
-// Schema holds a JSON schema.
-type Schema struct {
-	// Name is the name of the property. For the parent/root property, this
-	// is "root". For child properties, this is the name of the property.
-	Name string `json:"-"`
-
-	// Type is the type of the property.
-	//
-	// TODO: Union types (e.g. make this a []string).
-	Type string
-
-	// PrefixItems is a list of schemas for each item in a tuple. By
-	// default, the tuple is "closed." unless Items is set to true or a
-	// valid Schema.
-	PrefixItems []*Schema
-
-	// Items is the schema for each item in a list.
-	//
-	// If it is missing, or its JSON value is "null" or "false", it is nil.
-	// If the JSON value is "true", it is set to the empty Schema. If the
-	// JSON value is an object, it will be decoded as a Schema.
-	Items *Schema
-
-	// MinItems specifies the minimum number of items allowed in a list.
-	MinItems int
-
-	// MaxItems specifies the maximum number of items allowed in a list.
-	MaxItems int
-
-	// Properties is the schema for each property of an object.
-	Properties []*Schema
-
-	// Format is the format of the property. This is used to validate the
-	// property against a specific format.
-	//
-	// It is the callers responsibility to validate the property against
-	// the format.
-	Format string
-
-	// Minimum specifies the minimum value for numeric properties.
-	Minimum float64
-
-	// Maximum specifies the maximum value for numeric properties.
-	Maximum float64
-
-	// Enum is a list of valid values for the property.
-	Enum []json.RawMessage
-}
-
-func (s *Schema) UnmarshalJSON(data []byte) error {
-	type S Schema
-	w := struct {
-		Properties props
-		Items      items
-		*S
-	}{
-		S: (*S)(s),
-	}
-	if err := json.Unmarshal(data, &w); err != nil {
-		return err
-	}
-	if w.Items.set {
-		s.Items = &w.Items.Schema
-	}
-	s.Properties = w.Properties
-	return nil
-}
-
-type items struct {
-	Schema
-	set bool
-}
-
-func (s *items) UnmarshalJSON(data []byte) error {
-	switch b := data[0]; b {
-	case 't':
-		*s = items{set: true}
-	case '{':
-		type I items
-		if err := json.Unmarshal(data, (*I)(s)); err != nil {
-			return err
-		}
-		s.set = true
-	case 'n', 'f':
-	default:
-		return errors.New("invalid Items")
-	}
-	return nil
-}
-
-// EffectiveType returns the effective type of the schema. If the Type field is
-// not empty, it is returned; otherwise:
-//
-//   - If the schema has both Properties and Items, it returns an empty string.
-//   - If the schema has Properties, it returns "object".
-//   - If the schema has Items, it returns "array".
-//   - If the schema has neither Properties nor Items, it returns "value".
-//
-// The returned string is never empty.
-func (d *Schema) EffectiveType() string {
-	if d.Type == "" {
-		if len(d.Properties) > 0 {
-			return "object"
-		}
-		if len(d.PrefixItems) > 0 || d.Items != nil {
-			return "array"
-		}
-		return "value"
-	}
-	return d.Type
-}
-
-// props is an ordered list of properties. The order of the properties
-// is the order in which they were defined in the schema.
-type props []*Schema
-
-var _ json.Unmarshaler = (*props)(nil)
-
-func (v *props) UnmarshalJSON(data []byte) error {
-	if len(data) == 0 {
-		return nil
-	}
-	if data[0] != '{' {
-		return errors.New("expected object")
-	}
-
-	d := json.NewDecoder(bytes.NewReader(data))
-
-	// TODO(bmizerany): Consider DisallowUnknownFields. Currently, we, like
-	// llama.cpp, ignore unknown fields, which could be lead to unexpected
-	// behavior for clients of this package, since they may not be aware
-	// that "additionalFields", "itemsPrefix", etc, are being ignored.
-	//
-	// For now, just do what llama.cpp does.
-
-	t, err := d.Token()
-	if err != nil {
-		return err
-	}
-	if t != json.Delim('{') {
-		return errors.New("expected object")
-	}
-	for d.More() {
-		// Use the first token (map key) as the property name, then
-		// decode the rest of the object fields into a Schema and
-		// append.
-		t, err := d.Token()
-		if err != nil {
-			return err
-		}
-		if t == json.Delim('}') {
-			return nil
-		}
-		s := &Schema{
-			Name: t.(string),
-		}
-		if err := d.Decode(s); err != nil {
-			return err
-		}
-		*v = append(*v, s)
-	}
-	return nil
-}

grammar/jsonschema/decode_test.go

@@ -1,104 +0,0 @@
-package jsonschema
-
-import (
-	"encoding/json"
-	"reflect"
-	"strings"
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-)
-
-const testSchemaBasic = `
-{
-  "properties": {
-    "tupleClosedEmpty":   { "prefixItems": [] },
-    "tupleClosedMissing": { "prefixItems": [{}] },
-    "tupleClosedNull":    { "prefixItems": [{}], "items": null },
-    "tupleClosedFalse":   { "prefixItems": [{}], "items": false },
-    "tupleOpenTrue":      { "prefixItems": [{}], "items": true },
-    "tupleOpenEmpty":     { "prefixItems": [{}], "items": {} },
-    "tupleOpenTyped":     { "prefixItems": [{}], "items": {"type": "boolean"} },
-    "tupleOpenMax":       { "prefixItems": [{}], "items": true, "maxItems": 3},
-
-    "array": { "items": {"type": "number"} },
-
-    "null": { "type": "null" },
-    "string": { "type": "string" },
-    "boolean": { "type": "boolean" }
-  }
-}
-`
-
-func TestSchemaUnmarshal(t *testing.T) {
-	var got *Schema
-	if err := json.Unmarshal([]byte(testSchemaBasic), &got); err != nil {
-		t.Fatalf("Unmarshal: %v", err)
-	}
-	want := &Schema{
-		Properties: []*Schema{
-			{Name: "tupleClosedEmpty", PrefixItems: []*Schema{}, Items: nil},
-			{Name: "tupleClosedMissing", PrefixItems: []*Schema{{}}, Items: nil},
-			{Name: "tupleClosedNull", PrefixItems: []*Schema{{}}, Items: nil},
-			{Name: "tupleClosedFalse", PrefixItems: []*Schema{{}}, Items: nil},
-
-			{Name: "tupleOpenTrue", PrefixItems: []*Schema{{}}, Items: &Schema{}},
-			{Name: "tupleOpenEmpty", PrefixItems: []*Schema{{}}, Items: &Schema{}},
-			{Name: "tupleOpenTyped", PrefixItems: []*Schema{{}}, Items: &Schema{Type: "boolean"}},
-			{Name: "tupleOpenMax", PrefixItems: []*Schema{{}}, Items: &Schema{}, MaxItems: 3},
-
-			{Name: "array", Items: &Schema{Type: "number"}},
-
-			{Name: "null", Type: "null"},
-			{Name: "string", Type: "string"},
-			{Name: "boolean", Type: "boolean"},
-		},
-	}
-
-	if diff := cmp.Diff(want, got); diff != "" {
-		t.Errorf("(-want, +got)\n%s", diff)
-	}
-}
-
-func TestEffectiveType(t *testing.T) {
-	const schema = `
-		{"properties": {
-			"o": {"type": "object"},
-			"a": {"type": "array"},
-			"n": {"type": "number"},
-			"s": {"type": "string"},
-			"z": {"type": "null"},
-			"b": {"type": "boolean"},
-
-			"t0": {"prefixItems": [{}], "items": {"type": "number"}},
-			"t1": {"items": {"type": "number"}, "maxItems": 3},
-
-			"v": {"maxItems": 3}
-		}}
-	`
-
-	var s *Schema
-	if err := json.Unmarshal([]byte(schema), &s); err != nil {
-		t.Fatalf("json.Unmarshal: %v", err)
-	}
-
-	var got []string
-	for _, p := range s.Properties {
-		got = append(got, p.EffectiveType())
-	}
-
-	want := strings.Fields(`
-		object
-		array
-		number
-		string
-		null
-		boolean
-		array
-		array
-		value
-	`)
-	if !reflect.DeepEqual(want, got) {
-		t.Errorf("\ngot:\n\t%v\nwant:\n\t%v", got, want)
-	}
-}

grammar/testdata/schemas.txt

@@ -1,76 +0,0 @@
-# This file holds tests for JSON schema to EBNF grammar conversions.
-#
-# The format is a JSON schema, followed by the expected EBNF grammar. Each test
-# MAY be preceded by a comment that describes the test (e.g. the test name), followed by
-# the JSON schema and the expected EBNF grammar. If no comment is present, the test
-# name the tests number in the file (e.g. "#0", "#1", etc.)
-#
-# Blank lines signify the end or start of a new test. Comments can be added
-# anywhere in the file, but they must be preceded by a '#' character and start at
-# the beginning of the line.
-
-# default
-{}
-root ::= value;
-
-{"properties": {}}
-root ::= value;
-
-# array
-{"properties": {"a": {"type": "array", "items": {"type": "string"}}}}
-root_0_tuple_0 ::= string;
-root_0         ::= "[" ( root_0_tuple_0 )* "]";
-root           ::= "{" "a" ":" root_0 "}";
-
-# array with nested array
-{"type": "array", "items": {"type": "array", "items": {"type": "string"}}}
-root_tuple_0_tuple_0 ::= string;
-root_tuple_0         ::= "[" ( root_tuple_0_tuple_0 )* "]";
-root                 ::= "[" ( root_tuple_0 )* "]";
-
-# object
-{"properties": {"e": {}}}
-root_0 ::= value;
-root   ::= "{" "e" ":" root_0 "}";
-
-# object with nested object
-{"properties": {"o": {"type": "object", "properties": {"e": {}}}}}
-root_0_0 ::= value;
-root_0   ::= "{" "e" ":" root_0_0 "}";
-root     ::= "{" "o" ":" root_0 "}";
-
-# boolean
-{"type": "boolean"}
-root ::= boolean;
-
-# number
-{"properties": {"n": {"type": "number", "minimum": 123, "maximum": 4567}}}
-root_0 ::= number;
-root   ::= "{" "n" ":" root_0 "}";
-
-# string
-{"type": "string"}
-root ::= string;
-
-# string with enum
-{"type": "string", "enum": ["a", "b", "c"]}
-root ::= ( "\"a\"" "|" "\"b\"" "|" "\"c\"" );
-
-# spaces in key
-{"properties": {"a b": {}}}
-root_0 ::= value;
-root   ::= "{" "a b" ":" root_0 "}";
-
-# issue7978
-{ "type": "object", "properties": { "steps": { "type": "array", "items": { "type": "object", "properties": { "explanation": { "type": "string" }, "output": { "type": "string" } }, "required": [ "explanation", "output" ], "additionalProperties": false } }, "final_answer": { "type": "string" } }, "required": [ "steps", "final_answer" ], "additionalProperties": false }
-root_0_tuple_0_0 ::= string;
-root_0_tuple_0_1 ::= string;
-root_0_tuple_0   ::= "{" "explanation" ":" root_0_tuple_0_0 "," "output" ":" root_0_tuple_0_1 "}";
-root_0           ::= "[" ( root_0_tuple_0 )* "]";
-root_1           ::= string;
-root             ::= "{" "steps" ":" root_0 "," "final_answer" ":" root_1 "}";
-
-# !! # special characters in key
-# !! {"properties": {"a!b": {}}}
-# !! !invalid character '!' in key
-# !! 

kvcache/causal.go

@@ -119,10 +119,10 @@ func (c *Causal) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity
 	}
 
 	var cacheSize int
-	if c.windowSize == math.MaxInt32 || capacity < int(c.windowSize)+maxBatch {
+	if c.windowSize == math.MaxInt32 || capacity < int(c.windowSize) {
 		cacheSize = maxSequences * capacity
 	} else {
-		cacheSize = maxSequences * (int(c.windowSize) + maxBatch)
+		cacheSize = (maxSequences * int(c.windowSize)) + maxBatch
 	}
 	cacheSize = roundUp(cacheSize, c.config.CachePadding)
 	c.cells = make([]cacheCell, cacheSize)

llm/memory.go

@@ -15,12 +15,12 @@ import (
 )
 
 // This algorithm looks for a complete fit to determine if we need to unload other models
-func PredictServerFit(allGpus discover.GpuInfoList, f *ggml.GGML, adapters, projectors []string, opts api.Options) (bool, uint64) {
+func PredictServerFit(allGpus discover.GpuInfoList, f *ggml.GGML, adapters, projectors []string, opts api.Options, numParallel int) (bool, uint64) {
 	// Split up the GPUs by type and try them
 	var estimatedVRAM uint64
 	for _, gpus := range allGpus.ByLibrary() {
 		var layerCount int
-		estimate := EstimateGPULayers(gpus, f, projectors, opts)
+		estimate := EstimateGPULayers(gpus, f, projectors, opts, numParallel)
 		layerCount, estimatedVRAM = estimate.Layers, estimate.VRAMSize
 		if opts.NumGPU < 0 {
 			if layerCount > 0 && layerCount >= int(f.KV().BlockCount()+1) {
@@ -71,7 +71,7 @@ type MemoryEstimate struct {
 
 // Given a model and one or more GPU targets, predict how many layers and bytes we can load, and the total size
 // The GPUs provided must all be the same Library
-func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []string, opts api.Options) MemoryEstimate {
+func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []string, opts api.Options, numParallel int) MemoryEstimate {
 	// Graph size for a partial offload, applies to all GPUs
 	var graphPartialOffload uint64
 
@@ -137,13 +137,19 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		}
 	}
 
-	kv, graphPartialOffload, graphFullOffload := f.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)), kvct)
+	kv, graphPartialOffload, graphFullOffload := f.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)), numParallel, kvct)
 
-	// KV is proportional to the number of layers
-	layerSize += kv / f.KV().BlockCount()
+	if len(kv) > 0 {
+		layerSize += kv[0]
+	}
+
+	var kvTotal uint64
+	for _, kvLayer := range kv {
+		kvTotal += kvLayer
+	}
 
 	if graphPartialOffload == 0 {
-		graphPartialOffload = f.KV().GQA() * kv / 6
+		graphPartialOffload = f.KV().GQA() * kvTotal / 6
 	}
 	if graphFullOffload == 0 {
 		graphFullOffload = graphPartialOffload
@@ -217,7 +223,7 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		// Some models have inconsistent layer sizes
 		if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
 			layerSize = blk.Size()
-			layerSize += kv / f.KV().BlockCount()
+			layerSize += kv[i]
 			memoryWeights += blk.Size()
 		}
 
@@ -315,7 +321,7 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		layersRequested:     opts.NumGPU,
 		layersModel:         int(f.KV().BlockCount()) + 1,
 		availableList:       availableList,
-		kv:                  kv,
+		kv:                  kvTotal,
 		allocationsList:     allocationsList,
 		memoryWeights:       memoryWeights,
 		memoryLayerOutput:   memoryLayerOutput,
@@ -374,7 +380,7 @@ func (m MemoryEstimate) LogValue() slog.Value {
 			slog.Group(
 				"weights",
 				// memory of the weights
-				"total", format.HumanBytes2(m.memoryWeights),
+				"total", format.HumanBytes2(m.memoryWeights+m.memoryLayerOutput),
 				// memory of repeating layers
 				"repeating", format.HumanBytes2(m.memoryWeights),
 				// memory of non-repeating layers

llm/memory_test.go

@@ -61,7 +61,7 @@ func TestEstimateGPULayers(t *testing.T) {
 	projectors := []string{}
 	opts := api.DefaultOptions()
 	t.Run("cpu", func(t *testing.T) {
-		estimate := EstimateGPULayers(gpus, ggml, projectors, opts)
+		estimate := EstimateGPULayers(gpus, ggml, projectors, opts, 1)
 		assert.Equal(t, 0, estimate.Layers)
 		assert.Equal(t, uint64(0), estimate.Graph)
 	})
@@ -112,7 +112,7 @@ func TestEstimateGPULayers(t *testing.T) {
 			gpus[1].FreeMemory += gpuMinimumMemory + layerSize + s.layer1*layerSize + 1
 			gpus[0].FreeMemory += max(graphFullOffload, graphPartialOffload)
 			gpus[1].FreeMemory += max(graphFullOffload, graphPartialOffload)
-			estimate := EstimateGPULayers(gpus, ggml, projectors, opts)
+			estimate := EstimateGPULayers(gpus, ggml, projectors, opts, 1)
 			assert.Equal(t, int(s.expect0+s.expect1), estimate.Layers, "scenario %d: %v", i, s)
 			assert.Equal(t, fmt.Sprintf("%d,%d", s.expect0, s.expect1), estimate.TensorSplit, "scenario %d: %v", i, s)
 			var layerSums uint64

llm/server.go

@@ -29,7 +29,6 @@ import (
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/fs/ggml"
-	"github.com/ollama/ollama/grammar"
 	"github.com/ollama/ollama/llama"
 	"github.com/ollama/ollama/model"
 )
@@ -110,7 +109,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		gpus = discover.GetCPUInfo()
 	}
 
-	estimate := EstimateGPULayers(gpus, f, projectors, opts)
+	estimate := EstimateGPULayers(gpus, f, projectors, opts, numParallel)
 	if len(gpus) > 1 || gpus[0].Library != "cpu" {
 		switch {
 		case gpus[0].Library == "metal" && estimate.VRAMSize > systemTotalMemory:
@@ -701,9 +700,9 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 			}
 
 			// User provided a JSON schema
-			g, err := grammar.FromSchema(nil, req.Format)
-			if err != nil {
-				return fmt.Errorf("invalid JSON schema in format: %w", err)
+			g := llama.SchemaToGrammar(req.Format)
+			if g == nil {
+				return fmt.Errorf("invalid JSON schema in format")
 			}
 			req.Grammar = string(g)
 		}
@@ -714,11 +713,6 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 		req.Options = &opts
 	}
 
-	if req.Options == nil {
-		opts := api.DefaultOptions()
-		req.Options = &opts
-	}
-
 	if err := s.sem.Acquire(ctx, 1); err != nil {
 		if errors.Is(err, context.Canceled) {
 			slog.Info("aborting completion request due to client closing the connection")
@@ -733,6 +727,7 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 	if req.Options.NumPredict < 0 || req.Options.NumPredict > 10*s.options.NumCtx {
 		req.Options.NumPredict = 10 * s.options.NumCtx
 	}
+
 	// Make sure the server is ready
 	status, err := s.getServerStatusRetry(ctx)
 	if err != nil {

model/process_text.go

@@ -32,8 +32,6 @@ type TextProcessor interface {
 	Encode(s string, addSpecial bool) ([]int32, error)
 	Decode([]int32) (string, error)
 	Is(int32, Special) bool
-
-	Vocab() *Vocabulary
 }
 
 type Vocabulary struct {

model/process_text_spm.go

@@ -49,10 +49,6 @@ func NewSentencePieceModel(pre string, vocab *Vocabulary) SentencePieceModel {
 	}
 }
 
-func (spm SentencePieceModel) Vocab() *Vocabulary {
-	return spm.vocab
-}
-
 func (spm SentencePieceModel) Is(id int32, special Special) bool {
 	return spm.vocab.Is(id, special)
 }

sample/state_machine.go

@@ -1,176 +0,0 @@
-package sample
-
-import (
-	"bytes"
-	"strings"
-
-	"github.com/ollama/ollama/model"
-)
-
-type Node struct {
-	TransitionEdges map[rune]*Node
-}
-
-type Graph struct {
-	proc        model.TextProcessor
-	decodedToks []string
-	curNode     *Node
-	grammar     []byte
-	rules       map[string]string
-}
-
-// baseRules is the set of rules that are used to parse the grammar
-// JSON grammar from RFC 7159
-var baseRules = map[string]string{
-	"object":  "\"{\" (kv (\",\" kv)*)? \"}\"",
-	"array":   "\"[\" (value (\",\" value)*)? \"]\"",
-	"string":  "\"\\\"\" char* \"\\\"\"",
-	"number":  "\"-\"? integer frac? exp?",
-	"kv":      "string \":\" value",
-	"integer": "\"0\" | [1-9] [0-9]*",
-	"frac":    "\".\" [0-9]+",
-	"exp":     "(\"e\" | \"E\") (\"+\" | \"-\") [0-9]+",
-	"escape":  "[\"/\" | \"b\" | \"f\" | \"n\" | \"r\" | \"t\" | unicode]",
-	"char":    "[^\"\\\\] | escape",
-	"space":   "(\" \" | \"\\t\" | \"\\n\" | \"\\r\")*",
-	"hex":     "[0-9] | [a-f] | [A-F]",
-	"boolean": "\"true\" | \"false\"",
-	"value":   "object | array | string | number | boolean | \"null\"",
-	"null":    "\"null\"",
-}
-
-func (g *Graph) BuildGraph(node *Node) error {
-	vocab := g.proc.Vocab()
-	decodedToks := make([]string, len(vocab.Values))
-	for i := range vocab.Values {
-		token, err := g.proc.Decode([]int32{int32(i)})
-		if err != nil {
-			return err
-		}
-		decodedToks[i] = token
-	}
-
-	g.decodedToks = decodedToks
-	g.rules = baseRules
-	g.rootPrefixes()
-	rootNode := &Node{
-		TransitionEdges: make(map[rune]*Node),
-	}
-	g.parseRule(g.rules["root"], rootNode)
-
-	return nil
-}
-
-// rootPrefixes extracts all root prefixes from the grammar
-// and parses the grammar string to extract root prefixes
-func (g *Graph) rootPrefixes() {
-	lines := bytes.Split(g.grammar, []byte("\n"))
-	for _, line := range lines {
-		line = bytes.TrimSpace(line)
-		if len(line) == 0 || bytes.HasPrefix(line, []byte("#")) {
-			continue
-		}
-
-		parts := bytes.SplitN(line, []byte("::="), 2)
-		if len(parts) != 2 {
-			continue
-		}
-
-		ruleName := string(bytes.TrimSpace(parts[0]))
-		if strings.HasPrefix(ruleName, "root") {
-			g.rules[ruleName] = string(bytes.TrimSpace(parts[1]))
-		}
-	}
-}
-
-// parseRule parses a grammar rule and returns a Node
-func (g *Graph) parseRule(rule string, curNode *Node) *Node {
-	/*
-		Here are the special characters in BNF grammar and their functions:
-		::= - Definition operator, means "is defined as"
-		| - Alternation, means "or"
-		* - Zero or more repetitions of preceding element
-		+ - One or more repetitions
-		? - Optional (zero or one occurrence)
-		[] - Character class, matches any single character within brackets
-		[^] - Negated character class, matches any character NOT listed
-		() - Grouping of elements
-		- - Range operator in character classes (e.g., [a-z])
-		"" - Literal string match
-	*/
-
-	// Split rule into tokens by whitespace
-	tokens := strings.Fields(rule)
-	if len(tokens) == 0 {
-		return &Node{
-			TransitionEdges: make(map[rune]*Node),
-		}
-	}
-
-	// Handle integer rule
-	if strings.Contains(rule, "[0-9]+") {
-		// Create node for first digit 1-9
-		firstDigitNode := &Node{
-			TransitionEdges: make(map[rune]*Node),
-		}
-		for r := '1'; r <= '9'; r++ {
-			curNode.TransitionEdges[r] = firstDigitNode
-		}
-
-		// Create node for subsequent digits 0-9
-		zeroToNineNode := &Node{
-			TransitionEdges: make(map[rune]*Node),
-		}
-		for r := '0'; r <= '9'; r++ {
-			// Loop back to same node for * operator
-			zeroToNineNode.TransitionEdges[r] = zeroToNineNode
-		}
-
-		// Connect first digit to subsequent digits
-		firstDigitNode.TransitionEdges = zeroToNineNode.TransitionEdges
-
-		// Also handle the "0" case
-		if strings.Contains(rule, "\"0\"") {
-			zeroNode := &Node{
-				TransitionEdges: make(map[rune]*Node),
-			}
-			curNode.TransitionEdges['0'] = zeroNode
-		}
-
-		return curNode
-	}
-
-	// recursive case
-	// grammar options
-	// TODO: handle left recursion
-	if strings.Contains(rule, "|") {
-		parts := strings.Split(rule, "|")
-		savedNode := curNode
-		for _, part := range parts {
-			// TODO: add correct transitions
-			g.parseRule(part, savedNode)
-		}
-	}
-
-	for _, token := range tokens {
-		if strings.HasPrefix(token, "\"") && strings.HasSuffix(token, "\"") {
-			token = strings.Trim(token, "\"")
-
-			for _, r := range token {
-				newNode := &Node{
-					TransitionEdges: make(map[rune]*Node),
-				}
-				curNode.TransitionEdges[r] = newNode
-				curNode = newNode
-			}
-			// strNode := &Node{
-			// 	TransitionEdges: make(map[rune]*Node),
-			// }
-
-			// TODO: length constraint
-			// to self
-		}
-	}
-
-	return curNode
-}

sample/structured_outputs.go

@@ -1,3 +0,0 @@
-package sample
-
-type StructuredOutput struct{}

sample/structured_outputs_test.go

@@ -1,194 +0,0 @@
-package sample
-
-import (
-	"testing"
-
-	"github.com/ollama/ollama/model"
-)
-
-func TestBuildGraph(t *testing.T) {
-	tests := []struct {
-		name    string
-		grammar []byte
-		wantErr bool
-	}{
-		{
-			name:    "empty grammar",
-			grammar: []byte{},
-			wantErr: false,
-		},
-		{
-			name: "valid grammar",
-			grammar: []byte(`root ::= value
-value ::= string | number`),
-			wantErr: false,
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			g := &Graph{
-				proc:    &mockProcessor{},
-				grammar: tt.grammar,
-				rules:   make(map[string]string),
-			}
-
-			node := &Node{
-				TransitionEdges: make(map[rune]*Node),
-			}
-
-			err := g.BuildGraph(node)
-			if (err != nil) != tt.wantErr {
-				t.Errorf("BuildGraph() error = %v, wantErr %v", err, tt.wantErr)
-			}
-
-			if !tt.wantErr {
-				if len(g.decodedToks) == 0 {
-					t.Error("Expected decoded tokens, got none")
-				}
-				if len(g.rules) == 0 {
-					t.Error("Expected rules to be populated")
-				}
-			}
-		})
-	}
-}
-
-func TestRootPrefixes(t *testing.T) {
-	tests := []struct {
-		name     string
-		grammar  []byte
-		expected map[string]string
-	}{
-		{
-			name:     "empty grammar",
-			grammar:  []byte{},
-			expected: map[string]string{},
-		},
-		{
-			name: "grammar with root prefix",
-			grammar: []byte(`root ::= value
-root_string ::= string`),
-			expected: map[string]string{
-				"root":        "value",
-				"root_string": "string",
-			},
-		},
-		{
-			name: "grammar with comments and empty lines",
-			grammar: []byte(`# comment
-root ::= value
-
-# another comment
-root_number ::= number`),
-			expected: map[string]string{
-				"root":        "value",
-				"root_number": "number",
-			},
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			g := &Graph{
-				grammar: tt.grammar,
-				rules:   make(map[string]string),
-			}
-
-			g.rootPrefixes()
-
-			for k, v := range tt.expected {
-				if actual, ok := g.rules[k]; !ok || actual != v {
-					t.Errorf("Expected rule %s = %s, got %s", k, v, actual)
-				}
-			}
-		})
-	}
-}
-
-func TestParseRule(t *testing.T) {
-	tests := []struct {
-		name     string
-		rule     string
-		expected string
-	}{
-		{
-			name:     "empty rule",
-			rule:     "",
-			expected: "",
-		},
-		{
-			name:     "simple string",
-			rule:     "root ::= \"test_string\"",
-			expected: "test_string",
-		},
-		{
-			name:     "simple string",
-			rule:     "root ::= \"test_string\" | \"test_string2\"",
-			expected: "test_stringtest_string2",
-		},
-		{
-			name: "integer",
-			rule: "root ::= [0-9]+",
-			// TODO: this is infinite acutally
-			expected: "0123456789",
-		},
-		// TODO: handle left recursion
-		// {
-		// 	name:     "left recursion",
-		// 	rule:     "root ::= root \"test_string\"",
-		// 	expected: "test_string",
-		// },
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			g := &Graph{
-				rules: make(map[string]string),
-			}
-
-			rootNode := &Node{
-				TransitionEdges: make(map[rune]*Node),
-			}
-			curNode := rootNode
-			g.parseRule(tt.rule, curNode)
-			sb := ""
-			for {
-				if len(curNode.TransitionEdges) == 0 {
-					break
-				}
-
-				for r, n := range curNode.TransitionEdges {
-					sb += string(r)
-					curNode = n
-				}
-				t.Logf("sb: %s", sb)
-			}
-
-			if sb != tt.expected {
-				t.Errorf("Expected %s, got %s", tt.expected, sb)
-			}
-		})
-	}
-}
-
-// mockProcessor implements the TextProcessor interface for testing
-type mockProcessor struct{}
-
-func (m *mockProcessor) Decode(tokens []int32) (string, error) {
-	return "test", nil
-}
-
-func (m *mockProcessor) Vocab() *model.Vocabulary {
-	return &model.Vocabulary{
-		Values: []string{"test1", "test2"},
-	}
-}
-
-func (m *mockProcessor) Encode(s string, addSpecial bool) ([]int32, error) {
-	return []int32{0, 1}, nil
-}
-
-func (m *mockProcessor) Is(token int32, special model.Special) bool {
-	return false
-}

server/sched.go

@@ -711,7 +711,7 @@ func pickBestFullFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuIn
 			req.opts.NumCtx = req.origNumCtx * p
 			if !envconfig.SchedSpread() {
 				for _, g := range sgl {
-					if ok, estimatedVRAM = llm.PredictServerFit([]discover.GpuInfo{g}, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+					if ok, estimatedVRAM = llm.PredictServerFit([]discover.GpuInfo{g}, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, p); ok {
 						slog.Info("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "parallel", p, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
 						*numParallel = p
 						return []discover.GpuInfo{g}
@@ -727,7 +727,7 @@ func pickBestFullFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuIn
 		// Now try all the GPUs
 		for _, p := range numParallelToTry {
 			req.opts.NumCtx = req.origNumCtx * p
-			if ok, estimatedVRAM = llm.PredictServerFit(sgl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+			if ok, estimatedVRAM = llm.PredictServerFit(sgl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, p); ok {
 				slog.Info("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "parallel", p, "required", format.HumanBytes2(estimatedVRAM))
 				*numParallel = p
 				return sgl
@@ -750,7 +750,7 @@ func pickBestPartialFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.Gp
 	var bestEstimate uint64
 	var bestFit int
 	for i, gl := range byLibrary {
-		_, estimatedVRAM := llm.PredictServerFit(gl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
+		_, estimatedVRAM := llm.PredictServerFit(gl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, *numParallel)
 		if estimatedVRAM > bestEstimate {
 			bestEstimate = estimatedVRAM
 			bestFit = i
@@ -825,7 +825,7 @@ func (s *Scheduler) expireRunner(model *Model) {
 // If not, pick a runner to unload, else return nil and the request can be loaded
 func (s *Scheduler) maybeFindCPURunnerToUnload(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList) *runnerRef {
 	slog.Debug("evaluating if CPU model load will fit in available system memory")
-	estimate := llm.EstimateGPULayers(gpus, f, req.model.ProjectorPaths, req.opts)
+	estimate := llm.EstimateGPULayers(gpus, f, req.model.ProjectorPaths, req.opts, req.opts.NumCtx/req.origNumCtx)
 	if estimate.TotalSize <= gpus[0].FreeMemory {
 		slog.Debug("cpu inference mode, model fits in available system memory", "model", format.HumanBytes2(estimate.TotalSize), "available", format.HumanBytes2(gpus[0].FreeMemory))
 		return nil