Files
ollama37/model/vocabulary.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- Removed outdated CI/CD workflows (tesla-k80-*.yml)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

113 lines
2.0 KiB
Go

package model
import (
"log/slog"
"slices"
"sync"
)
type Special int32
const (
SpecialBOS Special = iota
SpecialEOS
)
type Vocabulary struct {
Values []string
Types []int32
Scores []float32
Merges []string
BOS, EOS []int32
AddBOS, AddEOS bool
specialOnce sync.Once
special []string
valuesOnce sync.Once
values map[string]int32
mergeOnce sync.Once
merge map[string]int32
}
func (v *Vocabulary) Is(id int32, special Special) bool {
switch special {
case SpecialBOS:
return slices.Contains(v.BOS, id)
case SpecialEOS:
return slices.Contains(v.EOS, id)
default:
return false
}
}
func (v *Vocabulary) addSpecials(ids []int32) []int32 {
if v.AddBOS && len(v.BOS) > 0 {
if slices.Contains(v.BOS, ids[0]) {
slog.Warn("adding bos token to prompt which already has it", "id", v.BOS)
}
slog.Debug("adding bos token to prompt", "id", v.BOS[0])
ids = append([]int32{v.BOS[0]}, ids...)
}
if v.AddEOS && len(v.EOS) > 0 {
if slices.Contains(v.BOS, ids[len(ids)-1]) {
slog.Warn("adding eos token to prompt which already has it", "id", v.EOS)
}
slog.Debug("adding eos token to prompt", "id", v.EOS[0])
ids = append(ids, v.EOS[0])
}
return ids
}
func (v *Vocabulary) Encode(s string) int32 {
v.valuesOnce.Do(func() {
v.values = make(map[string]int32, len(v.Values))
for i, value := range v.Values {
v.values[value] = int32(i)
}
})
if id, ok := v.values[s]; ok {
return id
}
return -1
}
func (v *Vocabulary) Decode(id int32) string {
return v.Values[id]
}
func (v *Vocabulary) SpecialVocabulary() []string {
v.specialOnce.Do(func() {
for i := range v.Values {
if v.Types[i] == TOKEN_TYPE_CONTROL || v.Types[i] == TOKEN_TYPE_USER_DEFINED {
v.special = append(v.special, v.Values[i])
}
}
})
return v.special
}
func (v *Vocabulary) Merge(left, right string) int {
v.mergeOnce.Do(func() {
v.merge = make(map[string]int32, len(v.Merges))
for i, merge := range v.Merges {
v.merge[merge] = int32(i)
}
})
if id, ok := v.merge[left+" "+right]; ok {
return int(id)
}
return -1
}