model: add a test for model forward pass during implementation

Adds a new test file to verify model forward pass behavior through
JSON-specified test cases. The framework loads model files (.gguf) and their
corresponding test specifications to validate expected outputs using greedy
sampling.

.github/workflows/release.yaml

@@ -329,9 +329,7 @@ jobs:
           done
         working-directory: dist/${{ matrix.os }}-${{ matrix.arch }}
       - run: |
-          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in; do
-            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | pigz -9vc >$(basename ${ARCHIVE//.*/}.tgz);
-          done
+          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in; do tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE | pigz -9vc >$(basename ${ARCHIVE//.*/}.tgz); done
       - uses: actions/upload-artifact@v4
         with:
           name: dist-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.target }}

.gitignore

@@ -14,3 +14,6 @@ test_data
 __debug_bin*
 llama/build
 llama/vendor
+model/testdata/models/*
+!model/testdata/models/*.md
+!model/testdata/models/*.json

CMakeLists.txt

@@ -24,7 +24,7 @@ set(GGML_LLAMAFILE ON)
 set(GGML_CUDA_PEER_MAX_BATCH_SIZE 128)
 set(GGML_CUDA_GRAPHS ON)
 
-if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
+if((NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
     OR (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm|aarch64|ARM64|ARMv[0-9]+"))
     set(GGML_CPU_ALL_VARIANTS ON)
 endif()

README.md

@@ -381,7 +381,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [ChibiChat](https://github.com/CosmicEventHorizon/ChibiChat) (Kotlin-based Android app to chat with Ollama and Koboldcpp API endpoints)
 - [LocalLLM](https://github.com/qusaismael/localllm) (Minimal Web-App to run ollama models on it with a GUI)
 - [Ollamazing](https://github.com/buiducnhat/ollamazing) (Web extension to run Ollama models)
-- [OpenDeepResearcher-via-searxng](https://github.com/benhaotang/OpenDeepResearcher-via-searxng) (A Deep Research equivent endpoint with Ollama support for running locally)
 
 ### Cloud
 
@@ -549,7 +548,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Alfred Ollama](https://github.com/zeitlings/alfred-ollama) (Alfred Workflow)
 - [TextLLaMA](https://github.com/adarshM84/TextLLaMA) A Chrome Extension that helps you write emails, correct grammar, and translate into any language
 - [Simple-Discord-AI](https://github.com/zyphixor/simple-discord-ai)
-- [LLM Telegram Bot](https://github.com/innightwolfsleep/llm_telegram_bot) (telegram bot, primary for RP. Oobabooga-like buttons, [A1111](https://github.com/AUTOMATIC1111/stable-diffusion-webui) API integration e.t.c)
 
 ### Supported backends
 

api/client.go

@@ -126,8 +126,7 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 			return err
 		}
 	}
-
-	return ctx.Err()
+	return nil
 }
 
 const maxBufferSize = 512 * format.KiloByte
@@ -190,7 +189,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 		}
 	}
 
-	return ctx.Err()
+	return nil
 }
 
 // GenerateResponseFunc is a function that [Client.Generate] invokes every time

cmd/cmd.go

@@ -15,11 +15,13 @@ import (
 	"net"
 	"net/http"
 	"os"
+	"os/signal"
 	"path/filepath"
 	"runtime"
 	"strconv"
 	"strings"
 	"sync/atomic"
+	"syscall"
 	"time"
 
 	"github.com/containerd/console"
@@ -328,7 +330,6 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 			if err := PullHandler(cmd, []string{name}); err != nil {
 				return nil, err
 			}
-
 			return client.Show(cmd.Context(), &api.ShowRequest{Name: name})
 		}
 		return info, err
@@ -857,6 +858,17 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 	spinner := progress.NewSpinner("")
 	p.Add("", spinner)
 
+	cancelCtx, cancel := context.WithCancel(cmd.Context())
+	defer cancel()
+
+	sigChan := make(chan os.Signal, 1)
+	signal.Notify(sigChan, syscall.SIGINT)
+
+	go func() {
+		<-sigChan
+		cancel()
+	}()
+
 	var state *displayResponseState = &displayResponseState{}
 	var latest api.ChatResponse
 	var fullResponse strings.Builder
@@ -891,7 +903,10 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		req.KeepAlive = opts.KeepAlive
 	}
 
-	if err := client.Chat(cmd.Context(), req, fn); err != nil {
+	if err := client.Chat(cancelCtx, req, fn); err != nil {
+		if errors.Is(err, context.Canceled) {
+			return nil, nil
+		}
 		return nil, err
 	}
 
@@ -931,6 +946,17 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 		generateContext = []int{}
 	}
 
+	ctx, cancel := context.WithCancel(cmd.Context())
+	defer cancel()
+
+	sigChan := make(chan os.Signal, 1)
+	signal.Notify(sigChan, syscall.SIGINT)
+
+	go func() {
+		<-sigChan
+		cancel()
+	}()
+
 	var state *displayResponseState = &displayResponseState{}
 
 	fn := func(response api.GenerateResponse) error {
@@ -966,7 +992,10 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 		KeepAlive: opts.KeepAlive,
 	}
 
-	if err := client.Generate(cmd.Context(), &request, fn); err != nil {
+	if err := client.Generate(ctx, &request, fn); err != nil {
+		if errors.Is(err, context.Canceled) {
+			return nil
+		}
 		return err
 	}
 
@@ -988,7 +1017,8 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 		latest.Summary()
 	}
 
-	cmd.SetContext(context.WithValue(cmd.Context(), generateContextKey("context"), latest.Context))
+	ctx = context.WithValue(cmd.Context(), generateContextKey("context"), latest.Context)
+	cmd.SetContext(ctx)
 
 	return nil
 }

docs/windows.md

@@ -55,7 +55,7 @@ Here's a quick example showing API access from `powershell`
 ## Troubleshooting
 
 Ollama on Windows stores files in a few different locations.  You can view them in
-the explorer window by hitting `<Ctrl>+R` and type in:
+the explorer window by hitting `<cmd>+R` and type in:
 - `explorer %LOCALAPPDATA%\Ollama` contains logs, and downloaded updates
     - *app.log* contains most resent logs from the GUI application
     - *server.log* contains the most recent server logs

format/format_test.go

@@ -12,9 +12,6 @@ func TestHumanNumber(t *testing.T) {
 
 	testCases := []testCase{
 		{0, "0"},
-		{999, "999"},
-		{1000, "1K"},
-		{1001, "1K"},
 		{1000000, "1M"},
 		{125000000, "125M"},
 		{500500000, "500.50M"},

kvcache/causal_test.go

@@ -305,10 +305,6 @@ func (b *testBackend) NewContext() ml.Context {
 	return &testContext{}
 }
 
-func (b *testBackend) SystemInfo() string {
-	return "not implemented"
-}
-
 type testContext struct{}
 
 func (c *testContext) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
@@ -434,7 +430,7 @@ func (t *testTensor) Conv2D(ctx ml.Context, weight ml.Tensor, s0, s1, p0, p1, d0
 	panic("not implemented")
 }
 
-func (t *testTensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, dim uint32, base, scale float32) ml.Tensor {
+func (t *testTensor) RoPE(ctx ml.Context, rc ml.RopeConfig) ml.Tensor {
 	panic("not implemented")
 }
 

main.go

@@ -2,8 +2,6 @@ package main
 
 import (
 	"context"
-	"os"
-	"os/signal"
 
 	"github.com/spf13/cobra"
 
@@ -11,15 +9,5 @@ import (
 )
 
 func main() {
-	ctx, cancel := context.WithCancel(context.Background())
-	defer cancel()
-
-	sigChan := make(chan os.Signal, 1)
-	signal.Notify(sigChan, os.Interrupt)
-	go func() {
-		<-sigChan
-		cancel()
-	}()
-
-	cobra.CheckErr(cmd.NewCLI().ExecuteContext(ctx))
+	cobra.CheckErr(cmd.NewCLI().ExecuteContext(context.Background()))
 }

ml/backend.go

@@ -23,7 +23,6 @@ type Backend interface {
 	Config() Config
 	Get(name string) Tensor
 	NewContext() Context
-	SystemInfo() string
 }
 
 var backends = make(map[string]func(*os.File) (Backend, error))
@@ -44,6 +43,42 @@ func NewBackend(f *os.File) (Backend, error) {
 	return nil, fmt.Errorf("unsupported backend")
 }
 
+// RopeType specifies the type of RoPE (Rotary Position Embedding) to use, these types are implemented in the backend
+type RopeType int
+
+const (
+	RopeTypeStandard RopeType = iota
+	_                         // not yet used
+	RopeTypeNeoX
+)
+
+// RopeConfig contains all configuration for the RoPE (Rotary Position Embedding) operation
+type RopeConfig struct {
+	// PositionIDs contains the position indices for each token in the sequence
+	// These indices are used to calculate the rotary embeddings
+	PositionIDs Tensor
+
+	// RopeFactors is an optional tensor containing pre-computed rotation factors
+	RopeFactors Tensor
+
+	// RopeDim specifies the dimension size for the rotary embeddings
+	RopeDim uint32
+
+	// RopeType indicates which RoPE variant to use (e.g. normal or neox)
+	RopeType RopeType
+
+	// OrigCtxLen stores the original context length the model was trained with
+	OrigCtxLen int
+
+	// RopeBase is the base value used in the frequency calculation
+	RopeBase float32
+
+	// RopeScale is a scaling factor applied to position indices
+	RopeScale float32
+
+	// YaRN parameters can be added here if they need to be configurable
+}
+
 type Context interface {
 	Zeros(dtype DType, shape ...int) Tensor
 	FromFloatSlice(s []float32, shape ...int) (Tensor, error)
@@ -76,7 +111,7 @@ type Tensor interface {
 	Scale(ctx Context, s float64) Tensor
 
 	Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
-	RoPE(ctx Context, positionIDs, ropeFactors Tensor, dim uint32, base, scale float32) Tensor
+	RoPE(ctx Context, rc RopeConfig) Tensor
 
 	Tanh(ctx Context) Tensor
 	GELU(ctx Context) Tensor

ml/backend/ggml/ggml.go

@@ -1,27 +1,11 @@
 package ggml
 
-/*
-#cgo CPPFLAGS: -I${SRCDIR}/ggml/include
-#include <stdlib.h>
-#include <stdint.h>
-#include "ggml.h"
-#include "ggml-cpu.h"
-#include "ggml-backend.h"
-static struct ggml_backend_feature * getBackendFeatures(void *fp, ggml_backend_reg_t reg) {return ((ggml_backend_get_features_t)(fp))(reg);}
-static struct ggml_backend_feature * getNextBackendFeatures(struct ggml_backend_feature * feature) { return &feature[1];}
-
-typedef enum {COMP_UNKNOWN,COMP_GCC,COMP_CLANG} COMPILER;
-COMPILER inline get_compiler() {
-#if defined(__clang__)
-	return COMP_CLANG;
-#elif defined(__GNUC__)
-	return COMP_GCC;
-#else
-	return UNKNOWN_COMPILER;
-#endif
-}
-
-*/
+// #cgo CPPFLAGS: -I${SRCDIR}/ggml/include
+// #include <stdlib.h>
+// #include <stdint.h>
+// #include "ggml.h"
+// #include "ggml-cpu.h"
+// #include "ggml-backend.h"
 import "C"
 
 import (
@@ -595,13 +579,9 @@ func (t *Tensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 	}
 }
 
-const (
-	ropeTypeNorm C.int = iota
-)
-
-func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDim uint32, ropeBase, ropeScale float32) ml.Tensor {
-	if ropeFactors == nil {
-		ropeFactors = &Tensor{}
+func (t *Tensor) RoPE(ctx ml.Context, rc ml.RopeConfig) ml.Tensor {
+	if rc.RopeFactors == nil {
+		rc.RopeFactors = &Tensor{}
 	}
 
 	dequant := t.t
@@ -611,12 +591,15 @@ func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDi
 
 	return &Tensor{
 		t: C.ggml_rope_ext(
-			ctx.(*Context).ctx, dequant, positionIDs.(*Tensor).t, ropeFactors.(*Tensor).t,
-			C.int(ropeDim),
-			131072,       // YaRN n_ctx_train
-			ropeTypeNorm, // ROPE_TYPE_NORM
-			C.float(ropeBase),
-			C.float(ropeScale),
+			ctx.(*Context).ctx,
+			dequant,
+			rc.PositionIDs.(*Tensor).t,
+			rc.RopeFactors.(*Tensor).t,
+			C.int(rc.RopeDim),
+			C.int(rc.RopeType),
+			C.int(rc.OrigCtxLen),
+			C.float(rc.RopeBase),
+			C.float(rc.RopeScale),
 			0.,  // YaRN ext_factor
 			1.,  // YaRN attn_factor
 			32., // YaRN beta_fast
@@ -642,34 +625,3 @@ func (t *Tensor) Conv2D(ctx ml.Context, t2 ml.Tensor, s0, s1, p0, p1, d0, d1 int
 		t: C.ggml_conv_2d(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.int(s0), C.int(s1), C.int(p0), C.int(p1), C.int(d0), C.int(d1)),
 	}
 }
-
-func (b *Backend) SystemInfo() string {
-	var compiler string
-	switch C.get_compiler() {
-	case C.COMP_UNKNOWN:
-		compiler = "cgo(unknown_compiler)"
-	case C.COMP_GCC:
-		compiler = "cgo(gcc)"
-	case C.COMP_CLANG:
-		compiler = "cgo(clang)"
-	}
-
-	var s string
-	for i := range C.ggml_backend_reg_count() {
-		reg := C.ggml_backend_reg_get(i)
-		fName := C.CString("ggml_backend_get_features")
-		defer C.free(unsafe.Pointer(fName))
-		get_features_fn := C.ggml_backend_reg_get_proc_address(reg, fName)
-		if get_features_fn != nil {
-			s += C.GoString(C.ggml_backend_reg_name(reg))
-			s += " : "
-			for features := C.getBackendFeatures(get_features_fn, reg); features.name != nil; features = C.getNextBackendFeatures(features) {
-				s += C.GoString(features.name)
-				s += " = "
-				s += C.GoString(features.value)
-				s += " | "
-			}
-		}
-	}
-	return s + compiler
-}

model/model.go

@@ -21,7 +21,6 @@ import (
 	_ "github.com/ollama/ollama/ml/backend"
 )
 
-// Options contains the inputs for a model forward pass
 type Options struct {
 	Inputs    []int32
 	Positions []int32
@@ -35,13 +34,11 @@ type config struct {
 	Cache kvcache.Cache
 }
 
-// Base implements the common fields and methods for all models
 type Base struct {
 	b ml.Backend
 	config
 }
 
-// Backend returns the underlying backend that will run the model
 func (m *Base) Backend() ml.Backend {
 	return m.b
 }
@@ -50,7 +47,6 @@ func (m *Base) Config() config {
 	return m.config
 }
 
-// Model implements a specific model architecture, defining the forward pass and any model-specific configuration
 type Model interface {
 	Forward(ml.Context, Options) (ml.Tensor, error)
 
@@ -60,7 +56,6 @@ type Model interface {
 
 var models = make(map[string]func(ml.Config) (Model, error))
 
-// Register registers a model constructor for the given architecture
 func Register(name string, f func(ml.Config) (Model, error)) {
 	if _, ok := models[name]; ok {
 		panic("model: model already registered")
@@ -69,9 +64,8 @@ func Register(name string, f func(ml.Config) (Model, error)) {
 	models[name] = f
 }
 
-// New initializes a new model instance with the provided configuration based on the metadata in the model file
-func New(modelPath string) (Model, error) {
-	r, err := os.Open(modelPath)
+func New(s string) (Model, error) {
+	r, err := os.Open(s)
 	if err != nil {
 		return nil, err
 	}

model/model_external_test.go

@@ -0,0 +1,138 @@
+// Package model_test provides external tests for the model package.
+// This test file specifically tests the forward pass functionality on models.
+// It is in a separate package (model_test) to avoid import cycles while still
+// being able to test the public API of the model package.
+package model_test
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/sample"
+
+	_ "github.com/ollama/ollama/model/models"
+)
+
+type modelTest struct {
+	Prompt            string   `json:"prompt"`
+	OutputContainsOne []string `json:"output_contains_one"`
+}
+
+func TestForwardSimple(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping in short mode")
+	}
+
+	// Read all JSON files from testdata/models
+	files, err := os.ReadDir("testdata/models")
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	for _, file := range files {
+		if !strings.HasSuffix(file.Name(), ".json") {
+			continue
+		}
+
+		jsonPath := filepath.Join("testdata/models", file.Name())
+		ggufPath := filepath.Join("testdata/models", strings.TrimSuffix(file.Name(), ".json")+".gguf")
+
+		// Skip if no corresponding .gguf file exists
+		if _, err := os.Stat(ggufPath); err != nil {
+			t.Logf("skipping %s: no corresponding GGUF file found", file.Name())
+			continue
+		}
+
+		data, err := os.ReadFile(jsonPath)
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		var test modelTest
+		if err := json.Unmarshal(data, &test); err != nil {
+			t.Fatal(err)
+		}
+
+		t.Run(strings.TrimSuffix(file.Name(), ".json"), func(t *testing.T) {
+			m, err := model.New(ggufPath)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			m.Config().Cache.Init(m.Backend(), ml.DTypeF32, 2048)
+
+			inputs, err := m.(model.TextProcessor).Encode(test.Prompt)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			var result []string
+			for len(result) < 100 { // Limit to 100 tokens max
+				options := model.Options{
+					Inputs:    inputs,
+					Positions: make([]int32, len(inputs)),
+					Sequences: make([]int, len(inputs)),
+					Outputs:   []int32{int32(len(inputs) - 1)},
+				}
+				for i := range options.Positions {
+					options.Positions[i] = int32(i)
+					options.Sequences[i] = 0
+				}
+
+				ctx := m.Backend().NewContext()
+
+				modelOutput, err := model.Forward(ctx, m, options)
+				if err != nil {
+					ctx.Close()
+					t.Fatal(fmt.Errorf("forward pass failed: %v", err))
+				}
+
+				f32s := modelOutput.Floats()
+				logits := make([]float64, len(f32s))
+				for i, f32 := range f32s {
+					logits[i] = float64(f32)
+				}
+
+				token, err := sample.Sample(logits, sample.Greedy())
+				if err != nil {
+					ctx.Close()
+					t.Fatal(fmt.Errorf("sampling failed: %v", err))
+				}
+
+				ctx.Close()
+
+				// Greedy sampling: take the token with the highest logit
+				nextToken := int32(token[0])
+				if m.(model.TextProcessor).Is(nextToken, model.SpecialEOS) {
+					break
+				}
+
+				piece, err := m.(model.TextProcessor).Decode([]int32{nextToken})
+				if err != nil {
+					t.Fatal(err)
+				}
+
+				result = append(result, piece)
+				output := strings.Join(result, "")
+
+				for _, expectedOutput := range test.OutputContainsOne {
+					if strings.Contains(output, expectedOutput) {
+						t.Logf("Test passed with output: %q (matched expected: %q)", output, expectedOutput)
+						return
+					}
+				}
+
+				// Maintain full context by appending new token
+				inputs = append(inputs, nextToken)
+			}
+
+			t.Fatalf("Expected output containing one of %q but got: %q", test.OutputContainsOne, strings.Join(result, ""))
+		})
+	}
+}

model/models/llama/model.go

@@ -10,10 +10,10 @@ import (
 )
 
 type Options struct {
-	RopeFactors                      ml.Tensor `gguf:"rope_freqs.weight"`
-	hiddenSize, numHeads, numKVHeads int
-	eps, ropeBase, ropeScale         float32
-	ropeDim                          uint32
+	RopeFactors                              ml.Tensor `gguf:"rope_freqs.weight"`
+	ctxLen, hiddenSize, numHeads, numKVHeads int
+	eps, ropeBase, ropeScale                 float32
+	ropeDim                                  uint32
 }
 
 type Model struct {
@@ -46,6 +46,7 @@ func New(c ml.Config) (model.Model, error) {
 			numHeads:   int(c.Uint("attention.head_count")),
 			numKVHeads: int(c.Uint("attention.head_count_kv")),
 			eps:        c.Float("attention.layer_norm_rms_epsilon"),
+			ctxLen:     int(c.Uint("context_length")),
 			ropeBase:   c.Float("rope.freq_base"),
 			ropeScale:  c.Float("rope.freq_scale", 1),
 			ropeDim:    c.Uint("rope.dimension_count"),
@@ -67,14 +68,23 @@ type SelfAttention struct {
 func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
 	headDim := opts.hiddenSize / opts.numHeads
+	rc := ml.RopeConfig{
+		PositionIDs: positionIDs,
+		RopeFactors: opts.RopeFactors,
+		RopeDim:     opts.ropeDim,
+		RopeType:    ml.RopeTypeStandard,
+		OrigCtxLen:  opts.ctxLen,
+		RopeBase:    opts.ropeBase,
+		RopeScale:   opts.ropeScale,
+	}
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	q = q.RoPE(ctx, positionIDs, opts.RopeFactors, opts.ropeDim, opts.ropeBase, opts.ropeScale)
+	q = q.RoPE(ctx, rc)
 
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	k = k.RoPE(ctx, positionIDs, opts.RopeFactors, opts.ropeDim, opts.ropeBase, opts.ropeScale)
+	k = k.RoPE(ctx, rc)
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -99,7 +109,18 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 }
 
 func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return key.RoPE(ctx, shift, m.Options.RopeFactors, m.Options.ropeDim, m.Options.ropeBase, m.Options.ropeScale), nil
+	return key.RoPE(
+		ctx,
+		ml.RopeConfig{
+			PositionIDs: shift,
+			RopeFactors: m.Options.RopeFactors,
+			RopeDim:     m.Options.ropeDim,
+			RopeType:    ml.RopeTypeStandard,
+			OrigCtxLen:  m.Options.ctxLen,
+			RopeBase:    m.Options.ropeBase,
+			RopeScale:   m.Options.ropeScale,
+		},
+	), nil
 }
 
 type MLP struct {

model/models/mllama/model_text.go

@@ -19,14 +19,23 @@ type TextSelfAttention struct {
 func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions, _ ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
 	headDim := opts.hiddenSize / opts.numHeads
+	rc := ml.RopeConfig{
+		PositionIDs: positions,
+		RopeFactors: opts.RopeFactors,
+		RopeDim:     opts.ropeDim,
+		RopeType:    ml.RopeTypeStandard,
+		OrigCtxLen:  opts.ctxLen,
+		RopeBase:    opts.ropeBase,
+		RopeScale:   opts.ropeScale,
+	}
 
 	query := sa.Query.Forward(ctx, hiddenState)
 	query = query.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	query = query.RoPE(ctx, positions, opts.RopeFactors, opts.ropeDim, opts.ropeBase, opts.ropeScale)
+	query = query.RoPE(ctx, rc)
 
 	key := sa.Key.Forward(ctx, hiddenState)
 	key = key.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	key = key.RoPE(ctx, positions, opts.RopeFactors, opts.ropeDim, opts.ropeBase, opts.ropeScale)
+	key = key.RoPE(ctx, rc)
 
 	value := sa.Value.Forward(ctx, hiddenState)
 	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -52,7 +61,18 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions, _ m
 
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
 	// This will only get called for layers in the cache, which are just the self attention layers
-	return key.RoPE(ctx, shift, m.RopeFactors, m.ropeDim, m.ropeBase, m.ropeScale), nil
+	return key.RoPE(
+		ctx,
+		ml.RopeConfig{
+			PositionIDs: shift,
+			RopeFactors: m.RopeFactors,
+			RopeDim:     m.ropeDim,
+			RopeType:    ml.RopeTypeStandard,
+			OrigCtxLen:  m.ctxLen,
+			RopeBase:    m.ropeBase,
+			RopeScale:   m.ropeScale,
+		},
+	), nil
 }
 
 type TextMLP struct {
@@ -189,9 +209,9 @@ func (d *TextDecoder) Forward(ctx ml.Context, hiddenState, positionIDs, mask, cr
 type TextModelOptions struct {
 	RopeFactors ml.Tensor `gguf:"rope_freqs.weight"`
 
-	hiddenSize, numHeads, numKVHeads int
-	eps, ropeBase, ropeScale         float32
-	ropeDim                          uint32
+	ctxLen, hiddenSize, numHeads, numKVHeads int
+	eps, ropeBase, ropeScale                 float32
+	ropeDim                                  uint32
 
 	crossAttentionLayers []uint32
 }

model/models/models.go

@@ -3,4 +3,5 @@ package models
 import (
 	_ "github.com/ollama/ollama/model/models/llama"
 	_ "github.com/ollama/ollama/model/models/mllama"
+	_ "github.com/ollama/ollama/model/models/qwen2"
 )

model/models/qwen2/model.go

@@ -0,0 +1,222 @@
+package qwen2
+
+import (
+	"math"
+
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/model"
+)
+
+type Options struct {
+	RopeFactors    ml.Tensor `gguf:"rope_freqs.weight"`
+	contextLength  int
+	hiddenSize     int
+	numAttnHeads   int
+	numKVHeads     int
+	modelEpsilon   float32
+	ropeBaseFreq   float32
+	ropeFreqScale  float32
+	ropeDimensions uint32
+}
+
+type Model struct {
+	model.Base
+	model.BytePairEncoding
+
+	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 ml.Config) (model.Model, error) {
+	m := &Model{
+		BytePairEncoding: model.NewBytePairEncoding(
+			c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
+			&model.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Types:  c.Uints("tokenizer.ggml.token_type"),
+				Merges: c.Strings("tokenizer.ggml.merges"),
+				BOS:    int32(c.Uint("tokenizer.ggml.bos_token_id")),
+				EOS:    int32(c.Uint("tokenizer.ggml.eos_token_id")),
+			},
+		),
+		Layers: make([]Layer, c.Uint("block_count")),
+		Options: &Options{
+			hiddenSize:     int(c.Uint("embedding_length")),
+			numAttnHeads:   int(c.Uint("attention.head_count")),
+			numKVHeads:     int(c.Uint("attention.head_count_kv")),
+			modelEpsilon:   c.Float("attention.layer_norm_rms_epsilon"),
+			contextLength:  int(c.Uint("context_length")),
+			ropeBaseFreq:   c.Float("rope.freq_base"),
+			ropeFreqScale:  c.Float("rope.freq_scale", 1),
+			ropeDimensions: c.Uint("rope.dimension_count", 64),
+		},
+	}
+
+	m.Cache = kvcache.NewCausalCache(m.Shift)
+
+	return m, nil
+}
+
+// Shift applies rotary position embeddings to the key tensor for causal attention caching
+func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	return key.RoPE(
+		ctx,
+		ml.RopeConfig{
+			PositionIDs: shift,
+			RopeFactors: m.Options.RopeFactors,
+			RopeDim:     m.Options.ropeDimensions,
+			RopeType:    ml.RopeTypeNeoX,
+			OrigCtxLen:  m.Options.contextLength,
+			RopeBase:    m.Options.ropeBaseFreq,
+			RopeScale:   m.Options.ropeFreqScale,
+		},
+	), nil
+}
+
+// SelfAttention implements the multi-head self-attention mechanism
+// with separate projections for query, key, value and output transformations
+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"`
+}
+
+func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, inputPositions ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	// Initialize dimensions and configuration
+	batchSize := hiddenState.Dim(1)
+	headDimension := opts.hiddenSize / opts.numAttnHeads
+	ropeConfig := ml.RopeConfig{
+		PositionIDs: inputPositions,
+		RopeFactors: nil,
+		RopeDim:     opts.ropeDimensions,
+		RopeType:    ml.RopeTypeNeoX,
+		OrigCtxLen:  opts.contextLength,
+		RopeBase:    opts.ropeBaseFreq,
+		RopeScale:   opts.ropeFreqScale,
+	}
+
+	// Project and reshape query states with rotary embeddings
+	queryStates := sa.Query.Forward(ctx, hiddenState)
+	queryStates = queryStates.Reshape(ctx, headDimension, opts.numAttnHeads, batchSize)
+	queryStates = queryStates.RoPE(ctx, ropeConfig)
+
+	// Project and reshape key states with rotary embeddings
+	keyStates := sa.Key.Forward(ctx, hiddenState)
+	keyStates = keyStates.Reshape(ctx, headDimension, opts.numKVHeads, batchSize)
+	keyStates = keyStates.RoPE(ctx, ropeConfig)
+
+	// Project and reshape value states
+	valueStates := sa.Value.Forward(ctx, hiddenState)
+	valueStates = valueStates.Reshape(ctx, headDimension, opts.numKVHeads, batchSize)
+
+	// Update and retrieve from KV cache
+	cache.Put(ctx, keyStates, valueStates)
+	keyStates, valueStates, attentionMask := cache.Get(ctx)
+
+	// Prepare tensors for attention computation
+	queryStates = queryStates.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
+	keyStates = keyStates.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
+	valueStates = valueStates.Permute(ctx, 1, 2, 0, 3).Contiguous(ctx)
+
+	// Apply scaling and attention mask to scores
+	attentionScores := keyStates.MulmatFullPrec(ctx, queryStates)
+	attentionScores = attentionScores.Scale(ctx, 1.0/math.Sqrt(float64(headDimension)))
+	attentionScores = attentionScores.Add(ctx, attentionMask)
+	// Compute scaled dot-product attention
+	attentionProbs := attentionScores.Softmax(ctx)
+
+	// Apply attention weights and reshape
+	weightedStates := valueStates.Mulmat(ctx, attentionProbs)
+	weightedStates = weightedStates.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
+	weightedStates = weightedStates.Reshape(ctx, opts.hiddenSize, batchSize)
+
+	// Project to output dimension
+	return sa.Output.Forward(ctx, weightedStates)
+}
+
+// MLP implements the feed-forward network component with SwiGLU activation
+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, opts *Options) ml.Tensor {
+	// Apply SwiGLU activation gating
+	gateActivation := mlp.Gate.Forward(ctx, hiddenState).SILU(ctx)
+	upProjection := mlp.Up.Forward(ctx, hiddenState)
+	intermediateStates := gateActivation.Mul(ctx, upProjection)
+
+	// Project back to hidden dimension
+	return mlp.Down.Forward(ctx, intermediateStates)
+}
+
+// Layer represents a single transformer layer combining self-attention and feed-forward components
+type Layer struct {
+	AttentionNorm *nn.RMSNorm `gguf:"attn_norm"`
+	SelfAttention *SelfAttention
+	MLPNorm       *nn.RMSNorm `gguf:"ffn_norm"`
+	MLP           *MLP
+}
+
+func (l *Layer) Forward(ctx ml.Context, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	// Self-attention branch with residual connection
+	residual := hiddenState
+
+	normalizedAttention := l.AttentionNorm.Forward(ctx, hiddenState, opts.modelEpsilon)
+	attentionOutput := l.SelfAttention.Forward(ctx, normalizedAttention, positionIDs, cache, opts)
+	hiddenState = attentionOutput.Add(ctx, residual)
+
+	// Feed-forward branch with residual connection
+	residual = hiddenState
+	normalizedMLP := l.MLPNorm.Forward(ctx, hiddenState, opts.modelEpsilon)
+	mlpOutput := l.MLP.Forward(ctx, normalizedMLP, opts)
+	output := mlpOutput.Add(ctx, residual)
+
+	return output
+}
+
+func (m *Model) Forward(ctx ml.Context, opts model.Options) (ml.Tensor, error) {
+	// Convert input tokens and positions to tensors
+	inputTensor, err := ctx.FromIntSlice(opts.Inputs, len(opts.Inputs))
+	if err != nil {
+		return nil, err
+	}
+
+	positionsTensor, err := ctx.FromIntSlice(opts.Positions, len(opts.Positions))
+	if err != nil {
+		return nil, err
+	}
+
+	// Initial token embedding
+	hiddenStates := m.TokenEmbedding.Forward(ctx, inputTensor)
+
+	// Process through transformer layers
+	for i, layer := range m.Layers {
+		m.Cache.SetLayer(i)
+		hiddenStates = layer.Forward(ctx, hiddenStates, positionsTensor, m.Cache, m.Options)
+	}
+
+	// Final layer normalization and output projection
+	normalizedOutput := m.OutputNorm.Forward(ctx, hiddenStates, m.modelEpsilon)
+	logits := m.Output.Forward(ctx, normalizedOutput)
+
+	// Extract requested output token positions
+	outputsTensor, err := ctx.FromIntSlice(opts.Outputs, len(opts.Outputs))
+	if err != nil {
+		return nil, err
+	}
+
+	return logits.Rows(ctx, outputsTensor), nil
+}
+
+func init() {
+	model.Register("qwen2", New)
+}

model/testdata/models/README.md

@@ -0,0 +1,10 @@
+# Test Model Directory
+
+This directory is used for storing model files (like `.gguf` files) that are required to run the tests in `model_external_test.go`. 
+
+## Usage
+
+- Place any model files you need for testing in this directory
+- The test file will look for any model files here (e.g., `llama3.gguf`)
+- All non-markdown files in this directory are git-ignored to prevent large model files from being committed to the repository
+- Only `.md` files (like this README) will be tracked in git

model/testdata/models/qwen2_5.json

@@ -0,0 +1,7 @@
+{
+  "prompt": "<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n<|im_start|>user\nhi<|im_end|>\n<|im_start|>assistant\n",
+  "output_contains_one": [
+    "Hello",
+    "Hi"
+  ]
+}

progress/progress.go

@@ -1,7 +1,6 @@
 package progress
 
 import (
-	"bufio"
 	"fmt"
 	"io"
 	"sync"
@@ -14,8 +13,7 @@ type State interface {
 
 type Progress struct {
 	mu sync.Mutex
-	// buffer output to minimize flickering on all terminals
-	w *bufio.Writer
+	w  io.Writer
 
 	pos int
 
@@ -24,7 +22,7 @@ type Progress struct {
 }
 
 func NewProgress(w io.Writer) *Progress {
-	p := &Progress{w: bufio.NewWriter(w)}
+	p := &Progress{w: w}
 	go p.start()
 	return p
 }
@@ -49,29 +47,26 @@ func (p *Progress) stop() bool {
 func (p *Progress) Stop() bool {
 	stopped := p.stop()
 	if stopped {
-		fmt.Fprintln(p.w)
+		fmt.Fprint(p.w, "\n")
 	}
-
-	// show cursor
-	fmt.Fprint(p.w, "\033[?25h")
-	p.w.Flush()
 	return stopped
 }
 
 func (p *Progress) StopAndClear() bool {
+	fmt.Fprint(p.w, "\033[?25l")
+	defer fmt.Fprint(p.w, "\033[?25h")
+
 	stopped := p.stop()
 	if stopped {
 		// clear all progress lines
-		for range p.pos - 1 {
-			fmt.Fprint(p.w, "\033[A")
+		for i := range p.pos {
+			if i > 0 {
+				fmt.Fprint(p.w, "\033[A")
+			}
+			fmt.Fprint(p.w, "\033[2K\033[1G")
 		}
-
-		fmt.Fprint(p.w, "\033[2K", "\033[1G")
 	}
 
-	// show cursor
-	fmt.Fprint(p.w, "\033[?25h")
-	p.w.Flush()
 	return stopped
 }
 
@@ -86,31 +81,30 @@ func (p *Progress) render() {
 	p.mu.Lock()
 	defer p.mu.Unlock()
 
-	fmt.Fprint(p.w, "\033[?2026h")
-	defer fmt.Fprint(p.w, "\033[?2026l")
+	fmt.Fprint(p.w, "\033[?25l")
+	defer fmt.Fprint(p.w, "\033[?25h")
 
-	for range p.pos - 1 {
-		fmt.Fprint(p.w, "\033[A")
+	// clear already rendered progress lines
+	for i := range p.pos {
+		if i > 0 {
+			fmt.Fprint(p.w, "\033[A")
+		}
+		fmt.Fprint(p.w, "\033[2K\033[1G")
 	}
 
-	fmt.Fprint(p.w, "\033[1G")
-
 	// render progress lines
 	for i, state := range p.states {
-		fmt.Fprint(p.w, state.String(), "\033[K")
+		fmt.Fprint(p.w, state.String())
 		if i < len(p.states)-1 {
 			fmt.Fprint(p.w, "\n")
 		}
 	}
 
 	p.pos = len(p.states)
-	p.w.Flush()
 }
 
 func (p *Progress) start() {
 	p.ticker = time.NewTicker(100 * time.Millisecond)
-	// hide cursor
-	fmt.Fprint(p.w, "\033[?25l")
 	for range p.ticker.C {
 		p.render()
 	}

runner/ollamarunner/runner.go

@@ -813,8 +813,6 @@ func (s *Server) loadModel(
 		panic(err)
 	}
 
-	slog.Info("system", "info", s.model.Backend().SystemInfo() /* "threads", *threads */)
-
 	// TODO(jessegross): LoRA loading
 	if lpath.String() != "" {
 		panic("loras are not yet implemented")
@@ -883,6 +881,7 @@ func Execute(args []string) error {
 	})
 	slog.SetDefault(slog.New(handler))
 	slog.Info("starting ollama engine")
+	// TODO(jessegross): Some system info would be useful
 
 	server := &Server{
 		batchSize: *batchSize,