mirror of
https://github.com/ollama/ollama.git
synced 2026-04-26 10:45:57 +02:00
* readme: add Ellama to list of community integrations (#9800) * readme: add screenpipe to community integrations (#9786) * Add support for ROCm gfx1151 (#9773) * conditionally enable parallel pipelines * sample: make mutations in transforms explicit (#9743) * updated minP to use early exit making use of sorted tokens * ml/backend/ggml: allocate memory with malloc when loading model (#9822) * runner: remove cache prompt flag from ollama runner (#9826) We do not need to bypass the prompt caching in the ollama runner yet, as only embedding models needed to bypass the prompt caching. When embedding models are implemented they can skip initializing this cache completely. * ollamarunner: Check for minBatch of context space when shifting Models can specify that a group of inputs need to be handled a single batch. However, context shifting didn't respect this and could trigger a break anyways. In this case, we should instead trigger a context shift earlier so that it occurs before the grouped batch. Note that there still some corner cases: - A long prompt that exceeds the context window can get truncated in the middle of an image. With the current models, this will result in the model not recognizing the image at all, which is pretty much the expected result with truncation. - The context window is set less than the minimum batch size. The only solution to this is to refuse to load the model with these settings. However, this can never occur with current models and default settings. Since users are unlikely to run into these scenarios, fixing them is left as a follow up. * Applied latest patches from McBane87 See this for details: https://github.com/whyvl/ollama-vulkan/issues/7#issuecomment-2708820861 Signed-off-by: Vadim Grinco <vadim@grinco.eu> * Add ability to enable flash attention on vulkan (#4) * discover: add flash attention handling for vulkan * envconfig: fix typo in config.go As part of the process some code was refactored and I added a new field FlashAttention to GpuInfo since the previous solution didn't allow for a granular check via vulkan extensions. As a side effect, this now allows for granular per-device FA support checking in other places --------- Signed-off-by: Vadim Grinco <vadim@grinco.eu> Co-authored-by: zeo <108888572+zeozeozeo@users.noreply.github.com> Co-authored-by: Louis Beaumont <louis.beaumont@gmail.com> Co-authored-by: Daniel Hiltgen <dhiltgen@users.noreply.github.com> Co-authored-by: Michael Yang <mxyng@pm.me> Co-authored-by: Parth Sareen <parth.sareen@ollama.com> Co-authored-by: Jeffrey Morgan <jmorganca@gmail.com> Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com> Co-authored-by: Jesse Gross <jesse@ollama.com> Co-authored-by: Nikita <50599445+nasrally@users.noreply.github.com>
314 lines
8.2 KiB
Go
314 lines
8.2 KiB
Go
package sample
|
|
|
|
import (
|
|
"math"
|
|
"math/rand/v2"
|
|
"testing"
|
|
)
|
|
|
|
// Helper to convert float32 slice to logit slice
|
|
func toTokens(values []float32) []token {
|
|
tokens := make([]token, len(values))
|
|
for i, v := range values {
|
|
tokens[i] = token{
|
|
id: int32(i),
|
|
value: v,
|
|
}
|
|
}
|
|
return tokens
|
|
}
|
|
|
|
// Helper to compare logit slices
|
|
func compareLogits(t *testing.T, name string, want []float32, got []token) {
|
|
t.Helper()
|
|
if len(want) != len(got) {
|
|
t.Errorf("%s: length mismatch: want %d, got %d", name, len(want), len(got))
|
|
return
|
|
}
|
|
for i := range want {
|
|
if math.Abs(float64(got[i].value-want[i])) > 1e-6 {
|
|
t.Errorf("%s: index %d: want %f, got %f", name, i, want[i], got[i].value)
|
|
}
|
|
}
|
|
}
|
|
|
|
func TestTemperature(t *testing.T) {
|
|
input := []float32{1.0, 4.0, -2.0, 0.0}
|
|
tokens := toTokens(input)
|
|
temperature(tokens, 0.5)
|
|
want := []float32{2.0, 8.0, -4.0, 0.0}
|
|
compareLogits(t, "temperature(0.5)", want, tokens)
|
|
|
|
input = []float32{1.0, 4.0, -2.0, 0.0}
|
|
tokens = toTokens(input)
|
|
temperature(tokens, 1.0)
|
|
want = []float32{1.0, 4.0, -2.0, 0.0}
|
|
compareLogits(t, "temperature(1)", want, tokens)
|
|
|
|
input = []float32{1.0, 4.0, -2.0, 0.0}
|
|
tokens = toTokens(input)
|
|
temperature(tokens, 0.0)
|
|
want = []float32{1e7, 4e7, -2e7, 0.0}
|
|
compareLogits(t, "temperature(0)", want, tokens)
|
|
}
|
|
|
|
func TestSoftmax(t *testing.T) {
|
|
tests := []struct {
|
|
name string
|
|
input []float32
|
|
expected []float32
|
|
}{
|
|
{
|
|
name: "correctness softmax",
|
|
input: []float32{1, -2, 3, 0},
|
|
expected: []float32{0.113550, 0.005653, 0.839024, 0.041773},
|
|
},
|
|
{
|
|
name: "normal distribution",
|
|
input: []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367},
|
|
},
|
|
{
|
|
name: "single value",
|
|
input: []float32{1.0},
|
|
},
|
|
{
|
|
name: "identical values",
|
|
input: []float32{0.9, 0.9, 0.9},
|
|
},
|
|
{
|
|
name: "large values",
|
|
input: []float32{1000.0, 2000.0, 3000.0},
|
|
},
|
|
{
|
|
name: "small values",
|
|
input: []float32{1e-6, 2e-6, 3e-6},
|
|
},
|
|
{
|
|
name: "negative values",
|
|
input: []float32{-1.0, -2.0, -3.0},
|
|
},
|
|
{
|
|
name: "mixed values",
|
|
input: []float32{-100.0, 0.0, 100.0},
|
|
},
|
|
}
|
|
|
|
for _, tt := range tests {
|
|
t.Run(tt.name, func(t *testing.T) {
|
|
tokens := toTokens(tt.input)
|
|
softmax(tokens)
|
|
|
|
if tt.expected != nil {
|
|
compareLogits(t, tt.name, tt.expected, tokens)
|
|
return
|
|
}
|
|
|
|
// Check probabilities sum to 1
|
|
var sum float32
|
|
for _, token := range tokens {
|
|
sum += token.value
|
|
if token.value < 0 || token.value > 1 {
|
|
t.Errorf("probability out of range [0,1]: got %f", token.value)
|
|
}
|
|
}
|
|
if math.Abs(float64(sum-1.0)) > 1e-6 {
|
|
t.Errorf("probabilities don't sum to 1: got %f", sum)
|
|
}
|
|
})
|
|
}
|
|
}
|
|
|
|
func TestTopK(t *testing.T) {
|
|
input := []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
|
|
tokens := toTokens(input)
|
|
tokens = topK(tokens, 5)
|
|
if len(tokens) != 5 {
|
|
t.Errorf("topK(5): wrong length: want 5, got %d", len(tokens))
|
|
}
|
|
want := []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154}
|
|
compareLogits(t, "topK(3)", want, tokens)
|
|
|
|
tokens = toTokens(input)
|
|
tokens = topK(tokens, 20)
|
|
if len(tokens) != len(input) {
|
|
t.Errorf("topK(20): wrong length: want %d, got %d", len(input), len(tokens))
|
|
}
|
|
|
|
input = []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
|
|
want = []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154, 0.043722924, 0.036774673, 0.026986899, 0.01681367, 0.0046718004, 0.00412893, 0.0030491839}
|
|
tokens = toTokens(input)
|
|
tokens = topK(tokens, -1)
|
|
if len(tokens) != len(input) {
|
|
t.Errorf("topK(-1): wrong length: want %d, got %d", len(input), len(tokens))
|
|
}
|
|
compareLogits(t, "topK(-1)", want, tokens)
|
|
|
|
input = []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
|
|
want = []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154, 0.043722924, 0.036774673, 0.026986899, 0.01681367, 0.0046718004, 0.00412893, 0.0030491839}
|
|
tokens = toTokens(input)
|
|
tokens = topK(tokens, 0)
|
|
if len(tokens) != len(input) {
|
|
t.Errorf("topK(-1): wrong length: want %d, got %d", len(input), len(tokens))
|
|
}
|
|
compareLogits(t, "topK(-1)", want, tokens)
|
|
|
|
input = []float32{-1e7, -2e7, -3e7, -4e7}
|
|
tokens = toTokens(input)
|
|
tokens = topK(tokens, 1)
|
|
if len(tokens) < 1 {
|
|
t.Error("topK should keep at least one token")
|
|
}
|
|
}
|
|
|
|
func TestTopP(t *testing.T) {
|
|
input := []float32{-3, -2, -1, 0, 1, 2, 4}
|
|
tokens := toTokens(input)
|
|
|
|
// First apply temperature and softmax to get probabilities
|
|
softmax(tokens)
|
|
tokens = topK(tokens, 20)
|
|
|
|
// Then apply topP
|
|
tokens = topP(tokens, 0.95)
|
|
|
|
// Should keep tokens until cumsum > 0.95
|
|
if len(tokens) > 3 {
|
|
t.Errorf("topP(0.95): kept too many tokens: got %d", len(tokens))
|
|
t.Logf("got: %v", tokens)
|
|
}
|
|
|
|
// Test edge case - ensure at least one token remains
|
|
input = []float32{-1e6, -1e6, -1e6} // One dominant token
|
|
tokens = toTokens(input)
|
|
softmax(tokens)
|
|
tokens = topP(tokens, 0.0) // Very small p
|
|
if len(tokens) < 1 {
|
|
t.Error("topP should keep at least one token")
|
|
}
|
|
}
|
|
|
|
func TestMinP(t *testing.T) {
|
|
input := []float32{-3, -2, -1, 0, 1, 2, 4, 3}
|
|
tokens := toTokens(input)
|
|
|
|
// First apply temperature and softmax
|
|
tokens = topK(tokens, 20)
|
|
softmax(tokens)
|
|
|
|
tokens = minP(tokens, 1.0)
|
|
|
|
if len(tokens) != 1 {
|
|
t.Errorf("minP(1.0): should keep all tokens, got %d, want %d", len(tokens), len(tokens))
|
|
}
|
|
|
|
// Test with normal p value
|
|
tokens = toTokens(input) // Reset tokens
|
|
tokens = topK(tokens, 20)
|
|
softmax(tokens)
|
|
tokens = minP(tokens, 0.2)
|
|
|
|
// Should keep tokens with prob >= 0.2 * max_prob
|
|
if len(tokens) > 3 {
|
|
t.Errorf("minP(0.2): kept too many tokens: got %d", len(tokens))
|
|
t.Logf("got: %v", tokens)
|
|
}
|
|
|
|
// Test with zero p value
|
|
tokens = toTokens(input) // Reset tokens
|
|
tokens = topK(tokens, 20)
|
|
softmax(tokens)
|
|
tokens = minP(tokens, 0.0)
|
|
|
|
// Should keep only the highest probability token
|
|
if len(tokens) != len(input) {
|
|
t.Errorf("minP(0.0): should keep only one token, got %d", len(tokens))
|
|
t.Logf("got: %v", tokens)
|
|
}
|
|
|
|
input = []float32{1e-10, 1e-10, 1e-10}
|
|
tokens = toTokens(input)
|
|
softmax(tokens)
|
|
tokens = minP(tokens, 1.0)
|
|
if len(tokens) < 1 {
|
|
t.Error("minP should keep at least one token even with extreme probabilities")
|
|
}
|
|
}
|
|
|
|
func TestSortLogits(t *testing.T) {
|
|
input := []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
|
|
tokens := toTokens(input)
|
|
|
|
tokens = topK(tokens, 20)
|
|
|
|
for i := 1; i < len(tokens); i++ {
|
|
if tokens[i].value > tokens[i-1].value {
|
|
t.Errorf("sortLogits: tokens not sorted in descending order at index %d: %f > %f",
|
|
i, tokens[i].value, tokens[i-1].value)
|
|
}
|
|
}
|
|
|
|
want := []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154, 0.043722924, 0.036774673, 0.026986899, 0.01681367, 0.0046718004, 0.00412893, 0.0030491839}
|
|
compareLogits(t, "sortLogits", want, tokens)
|
|
}
|
|
|
|
func BenchmarkTransforms(b *testing.B) {
|
|
// Generate random logits
|
|
tokens := make([]token, 1<<16)
|
|
for i := range tokens {
|
|
tokens[i] = token{
|
|
id: int32(i),
|
|
value: rand.Float32(),
|
|
}
|
|
}
|
|
|
|
tokensCopy := make([]token, len(tokens))
|
|
|
|
b.Run("Temperature", func(b *testing.B) {
|
|
b.ResetTimer()
|
|
for b.Loop() {
|
|
copy(tokensCopy, tokens)
|
|
temperature(tokensCopy, 0.5)
|
|
}
|
|
})
|
|
|
|
b.Run("Softmax", func(b *testing.B) {
|
|
b.ResetTimer()
|
|
for b.Loop() {
|
|
copy(tokensCopy, tokens)
|
|
softmax(tokensCopy)
|
|
}
|
|
})
|
|
|
|
b.Run("TopK", func(b *testing.B) {
|
|
b.ResetTimer()
|
|
for b.Loop() {
|
|
copy(tokensCopy, tokens)
|
|
tokens = topK(tokensCopy, 10)
|
|
}
|
|
})
|
|
|
|
b.Run("TopP", func(b *testing.B) {
|
|
b.ResetTimer()
|
|
for b.Loop() {
|
|
copy(tokensCopy, tokens)
|
|
tokens = topP(tokensCopy, 0.9)
|
|
}
|
|
})
|
|
|
|
b.Run("MinP", func(b *testing.B) {
|
|
b.ResetTimer()
|
|
for b.Loop() {
|
|
copy(tokensCopy, tokens)
|
|
tokens = minP(tokensCopy, 0.2)
|
|
}
|
|
})
|
|
|
|
b.Run("SortTokens", func(b *testing.B) {
|
|
b.ResetTimer()
|
|
for b.Loop() {
|
|
copy(tokensCopy, tokens)
|
|
tokens = topK(tokensCopy, 200000)
|
|
}
|
|
})
|
|
}
|