Files
ollama37/runner/llamarunner/image.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- Removed outdated CI/CD workflows (tesla-k80-*.yml)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

150 lines
3.0 KiB
Go

package llamarunner
import (
"errors"
"fmt"
"hash/maphash"
"log/slog"
"sync"
"time"
"github.com/ollama/ollama/llama"
)
const imageCacheSize = 4
type ImageContext struct {
// mu is required to be held when generating embeddings or accessing the cache
mu sync.Mutex
mtmd *llama.MtmdContext
// cache of images to embeddings
images []imageCache
imageHash maphash.Hash
}
func NewImageContext(llamaContext *llama.Context, modelPath string) (*ImageContext, error) {
arch, err := llama.GetModelArch(modelPath)
if err != nil {
return nil, fmt.Errorf("unable to determine vision architecture: %w (%s)", err, modelPath)
}
var c ImageContext
if arch == "clip" {
c.mtmd, err = llama.NewMtmdContext(llamaContext, modelPath)
} else {
return nil, fmt.Errorf("unknown vision model architecture: %s", arch)
}
if err != nil {
return nil, err
}
c.images = make([]imageCache, imageCacheSize)
return &c, nil
}
func (c *ImageContext) Free(modelPath string) {
if c == nil {
return
}
if c.mtmd != nil {
c.mtmd.Free()
}
}
func (c *ImageContext) MultimodalTokenize(llamaContext *llama.Context, data []byte) ([]llama.MtmdChunk, error) {
if c == nil {
return nil, nil
}
if len(data) <= 0 {
return nil, errors.New("received zero length image")
}
hash := c.hashImage(data)
c.mu.Lock()
defer c.mu.Unlock()
chunks, err := c.findImage(hash)
if err != nil {
if c.mtmd != nil {
chunks, err = c.mtmd.MultimodalTokenize(llamaContext, data)
if err != nil {
return nil, err
}
} else {
return nil, errors.New("received image but vision model not loaded")
}
c.addImage(hash, chunks)
}
return chunks, nil
}
func (c *ImageContext) BatchSize(configuredBatchSize int) int {
// If images are not supported, we don't need to allocate embedding batches
if c == nil {
return 0
}
return configuredBatchSize
}
func (c *ImageContext) EmbedSize(llamaContext *llama.Context) int {
return llamaContext.Model().NEmbd()
}
type imageCache struct {
key uint64
val []llama.MtmdChunk
lastUsed time.Time
}
func (c *ImageContext) hashImage(image []byte) uint64 {
c.imageHash.Reset()
_, _ = c.imageHash.Write(image)
return c.imageHash.Sum64()
}
var errImageNotFound = errors.New("image not found in cache")
func (c *ImageContext) findImage(hash uint64) ([]llama.MtmdChunk, error) {
for i := range c.images {
if c.images[i].key == hash {
slog.Debug("loading image embeddings from cache", "entry", i)
c.images[i].lastUsed = time.Now()
return c.images[i].val, nil
}
}
return nil, errImageNotFound
}
func (c *ImageContext) addImage(hash uint64, embed []llama.MtmdChunk) {
best := time.Now()
var bestImage int
for i := range c.images {
if c.images[i].key == hash {
bestImage = i
break
}
if c.images[i].lastUsed.Compare(best) < 0 {
best = c.images[i].lastUsed
bestImage = i
}
}
slog.Debug("storing image embeddings in cache", "entry", bestImage, "used", c.images[bestImage].lastUsed)
c.images[bestImage].key = hash
c.images[bestImage].val = embed
c.images[bestImage].lastUsed = time.Now()
}