Files
ollama37/server/prompt_test.go
Shang Chieh Tseng ef14fb5b26 Sync with upstream ollama/ollama and restore Tesla K80 (compute 3.7) support
This commit represents a complete rework after pulling the latest changes from
official ollama/ollama repository and re-applying Tesla K80 compatibility patches.

## Key Changes

### CUDA Compute Capability 3.7 Support (Tesla K80)
- Added sm_37 (compute 3.7) to CMAKE_CUDA_ARCHITECTURES in CMakeLists.txt
- Updated CMakePresets.json to include compute 3.7 in "CUDA 11" preset
- Using 37-virtual (PTX with JIT compilation) for maximum compatibility

### Legacy Toolchain Compatibility
- **NVIDIA Driver**: 470.256.02 (last version supporting Kepler/K80)
- **CUDA Version**: 11.4.4 (last CUDA 11.x supporting compute 3.7)
- **GCC Version**: 10.5.0 (required by CUDA 11.4 host_config.h)

### CPU Architecture Trade-offs
Due to GCC 10.5 limitation, sacrificed newer CPU optimizations:
- Alderlake CPU variant enabled WITHOUT AVX_VNNI (requires GCC 11+)
- Still supports: SSE4.2, AVX, F16C, AVX2, BMI2, FMA
- Performance impact: ~3-7% on newer CPUs (acceptable for K80 compatibility)

### Build System Updates
- Modified ml/backend/ggml/ggml/src/ggml-cuda/CMakeLists.txt for compute 3.7
- Added -Wno-deprecated-gpu-targets flag to suppress warnings
- Updated ml/backend/ggml/ggml/src/CMakeLists.txt for Alderlake without AVX_VNNI

### Upstream Sync
Merged latest llama.cpp changes including:
- Enhanced KV cache management with ISWA and hybrid memory support
- Improved multi-modal support (mtmd framework)
- New model architectures (Gemma3, Llama4, Qwen3, etc.)
- GPU backend improvements for CUDA, Metal, and ROCm
- Updated quantization support and GGUF format handling

### Documentation
- Updated CLAUDE.md with comprehensive build instructions
- Documented toolchain constraints and CPU architecture trade-offs
- Removed outdated CI/CD workflows (tesla-k80-*.yml)
- Cleaned up temporary development artifacts

## Rationale

This fork maintains Tesla K80 GPU support (compute 3.7) which was dropped in
official Ollama due to legacy driver/CUDA requirements. The toolchain constraint
creates a deadlock:
- K80 → Driver 470 → CUDA 11.4 → GCC 10 → No AVX_VNNI

We accept the loss of cutting-edge CPU optimizations to enable running modern
LLMs on legacy but still capable Tesla K80 hardware (12GB VRAM per GPU).

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-05 14:03:05 +08:00

267 lines
8.2 KiB
Go

package server
import (
"bytes"
"testing"
"github.com/google/go-cmp/cmp"
"github.com/ollama/ollama/api"
"github.com/ollama/ollama/template"
)
func TestChatPrompt(t *testing.T) {
type expect struct {
prompt string
images [][]byte
error error
}
tmpl, err := template.Parse(`
{{- if .System }}{{ .System }} {{ end }}
{{- if .Prompt }}{{ .Prompt }} {{ end }}
{{- if .Response }}{{ .Response }} {{ end }}`)
if err != nil {
t.Fatal(err)
}
visionModel := Model{Template: tmpl, ProjectorPaths: []string{"vision"}}
cases := []struct {
name string
model Model
limit int
truncate bool
msgs []api.Message
expect
}{
{
name: "messages",
model: visionModel,
limit: 64,
truncate: true,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "assistant", Content: "I-I'm a what?"},
{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
},
expect: expect{
prompt: "You're a test, Harry! I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
},
},
{
name: "truncate messages",
model: visionModel,
limit: 1,
truncate: true,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "assistant", Content: "I-I'm a what?"},
{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
},
expect: expect{
prompt: "A test. And a thumping good one at that, I'd wager. ",
},
},
{
name: "truncate messages with image",
model: visionModel,
limit: 64,
truncate: true,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "assistant", Content: "I-I'm a what?"},
{Role: "user", Content: "A test. And a thumping good one at that, I'd wager.", Images: []api.ImageData{[]byte("something")}},
},
expect: expect{
prompt: "[img-0]A test. And a thumping good one at that, I'd wager. ",
images: [][]byte{
[]byte("something"),
},
},
},
{
name: "truncate messages with images",
model: visionModel,
limit: 64,
truncate: true,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!", Images: []api.ImageData{[]byte("something")}},
{Role: "assistant", Content: "I-I'm a what?"},
{Role: "user", Content: "A test. And a thumping good one at that, I'd wager.", Images: []api.ImageData{[]byte("somethingelse")}},
},
expect: expect{
prompt: "[img-0]A test. And a thumping good one at that, I'd wager. ",
images: [][]byte{
[]byte("somethingelse"),
},
},
},
{
name: "messages with images",
model: visionModel,
limit: 2048,
truncate: true,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!", Images: []api.ImageData{[]byte("something")}},
{Role: "assistant", Content: "I-I'm a what?"},
{Role: "user", Content: "A test. And a thumping good one at that, I'd wager.", Images: []api.ImageData{[]byte("somethingelse")}},
},
expect: expect{
prompt: "[img-0]You're a test, Harry! I-I'm a what? [img-1]A test. And a thumping good one at that, I'd wager. ",
images: [][]byte{
[]byte("something"),
[]byte("somethingelse"),
},
},
},
{
name: "message with image tag",
model: visionModel,
limit: 2048,
truncate: true,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry! [img]", Images: []api.ImageData{[]byte("something")}},
{Role: "assistant", Content: "I-I'm a what?"},
{Role: "user", Content: "A test. And a thumping good one at that, I'd wager.", Images: []api.ImageData{[]byte("somethingelse")}},
},
expect: expect{
prompt: "You're a test, Harry! [img-0] I-I'm a what? [img-1]A test. And a thumping good one at that, I'd wager. ",
images: [][]byte{
[]byte("something"),
[]byte("somethingelse"),
},
},
},
{
name: "messages with interleaved images",
model: visionModel,
limit: 2048,
truncate: true,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "user", Images: []api.ImageData{[]byte("something")}},
{Role: "user", Images: []api.ImageData{[]byte("somethingelse")}},
{Role: "assistant", Content: "I-I'm a what?"},
{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
},
expect: expect{
prompt: "You're a test, Harry!\n\n[img-0]\n\n[img-1] I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
images: [][]byte{
[]byte("something"),
[]byte("somethingelse"),
},
},
},
{
name: "truncate message with interleaved images",
model: visionModel,
limit: 1024,
truncate: true,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "user", Images: []api.ImageData{[]byte("something")}},
{Role: "user", Images: []api.ImageData{[]byte("somethingelse")}},
{Role: "assistant", Content: "I-I'm a what?"},
{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
},
expect: expect{
prompt: "[img-0] I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
images: [][]byte{
[]byte("somethingelse"),
},
},
},
{
name: "message with system prompt",
model: visionModel,
limit: 2048,
truncate: true,
msgs: []api.Message{
{Role: "system", Content: "You are the Test Who Lived."},
{Role: "user", Content: "You're a test, Harry!"},
{Role: "assistant", Content: "I-I'm a what?"},
{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
},
expect: expect{
prompt: "You are the Test Who Lived. You're a test, Harry! I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
},
},
{
name: "out of order system",
model: visionModel,
limit: 2048,
truncate: true,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "assistant", Content: "I-I'm a what?"},
{Role: "system", Content: "You are the Test Who Lived."},
{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
},
expect: expect{
prompt: "You're a test, Harry! I-I'm a what? You are the Test Who Lived. A test. And a thumping good one at that, I'd wager. ",
},
},
{
name: "multiple images same prompt",
model: visionModel,
limit: 2048,
truncate: true,
msgs: []api.Message{
{Role: "user", Content: "Compare these two pictures of hotdogs", Images: []api.ImageData{[]byte("one hotdog"), []byte("two hotdogs")}},
},
expect: expect{
prompt: "[img-0][img-1]Compare these two pictures of hotdogs ",
images: [][]byte{[]byte("one hotdog"), []byte("two hotdogs")},
},
},
{
name: "no truncate with limit exceeded",
model: visionModel,
limit: 10,
truncate: false,
msgs: []api.Message{
{Role: "user", Content: "You're a test, Harry!"},
{Role: "assistant", Content: "I-I'm a what?"},
{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
},
expect: expect{
prompt: "You're a test, Harry! I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
},
},
}
for _, tt := range cases {
t.Run(tt.name, func(t *testing.T) {
model := tt.model
opts := api.Options{Runner: api.Runner{NumCtx: tt.limit}}
think := false
prompt, images, err := chatPrompt(t.Context(), &model, mockRunner{}.Tokenize, &opts, tt.msgs, nil, &api.ThinkValue{Value: think}, tt.truncate)
if tt.error == nil && err != nil {
t.Fatal(err)
} else if tt.error != nil && err != tt.error {
t.Fatalf("expected err '%q', got '%q'", tt.error, err)
}
if diff := cmp.Diff(prompt, tt.prompt); diff != "" {
t.Errorf("mismatch (-got +want):\n%s", diff)
}
if len(images) != len(tt.images) {
t.Fatalf("expected %d images, got %d", len(tt.images), len(images))
}
for i := range images {
if images[i].ID != i {
t.Errorf("expected ID %d, got %d", i, images[i].ID)
}
if len(model.Config.ModelFamilies) == 0 {
if !bytes.Equal(images[i].Data, tt.images[i]) {
t.Errorf("expected %q, got %q", tt.images[i], images[i].Data)
}
}
}
})
}
}