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- Removed outdated CI/CD workflows (tesla-k80-*.yml)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

173 lines
4.5 KiB
Go

package discover
import (
"bufio"
"fmt"
"io"
"log/slog"
"os"
"path/filepath"
"reflect"
"regexp"
"sort"
"strings"
"github.com/ollama/ollama/format"
)
func GetCPUMem() (memInfo, error) {
var mem memInfo
var total, available, free, buffers, cached, freeSwap uint64
f, err := os.Open("/proc/meminfo")
if err != nil {
return mem, err
}
defer f.Close()
s := bufio.NewScanner(f)
for s.Scan() {
line := s.Text()
switch {
case strings.HasPrefix(line, "MemTotal:"):
_, err = fmt.Sscanf(line, "MemTotal:%d", &total)
case strings.HasPrefix(line, "MemAvailable:"):
_, err = fmt.Sscanf(line, "MemAvailable:%d", &available)
case strings.HasPrefix(line, "MemFree:"):
_, err = fmt.Sscanf(line, "MemFree:%d", &free)
case strings.HasPrefix(line, "Buffers:"):
_, err = fmt.Sscanf(line, "Buffers:%d", &buffers)
case strings.HasPrefix(line, "Cached:"):
_, err = fmt.Sscanf(line, "Cached:%d", &cached)
case strings.HasPrefix(line, "SwapFree:"):
_, err = fmt.Sscanf(line, "SwapFree:%d", &freeSwap)
default:
continue
}
if err != nil {
return mem, err
}
}
mem.TotalMemory = total * format.KibiByte
mem.FreeSwap = freeSwap * format.KibiByte
if available > 0 {
mem.FreeMemory = available * format.KibiByte
} else {
mem.FreeMemory = (free + buffers + cached) * format.KibiByte
}
return mem, nil
}
const CpuInfoFilename = "/proc/cpuinfo"
type linuxCpuInfo struct {
ID string `cpuinfo:"processor"`
VendorID string `cpuinfo:"vendor_id"`
ModelName string `cpuinfo:"model name"`
PhysicalID string `cpuinfo:"physical id"`
Siblings string `cpuinfo:"siblings"`
CoreID string `cpuinfo:"core id"`
}
func GetCPUDetails() []CPU {
file, err := os.Open(CpuInfoFilename)
if err != nil {
slog.Warn("failed to get CPU details", "error", err)
return nil
}
defer file.Close()
return linuxCPUDetails(file)
}
func linuxCPUDetails(file io.Reader) []CPU {
reColumns := regexp.MustCompile("\t+: ")
scanner := bufio.NewScanner(file)
cpuInfos := []linuxCpuInfo{}
cpu := &linuxCpuInfo{}
for scanner.Scan() {
line := scanner.Text()
if sl := reColumns.Split(line, 2); len(sl) > 1 {
t := reflect.TypeOf(cpu).Elem()
s := reflect.ValueOf(cpu).Elem()
for i := range t.NumField() {
field := t.Field(i)
tag := field.Tag.Get("cpuinfo")
if tag == sl[0] {
s.FieldByName(field.Name).SetString(sl[1])
break
}
}
} else if strings.TrimSpace(line) == "" && cpu.ID != "" {
cpuInfos = append(cpuInfos, *cpu)
cpu = &linuxCpuInfo{}
}
}
if cpu.ID != "" {
cpuInfos = append(cpuInfos, *cpu)
}
// Process the sockets/cores/threads
socketByID := map[string]*CPU{}
coreBySocket := map[string]map[string]struct{}{}
threadsByCoreBySocket := map[string]map[string]int{}
for _, c := range cpuInfos {
if _, found := socketByID[c.PhysicalID]; !found {
socketByID[c.PhysicalID] = &CPU{
ID: c.PhysicalID,
VendorID: c.VendorID,
ModelName: c.ModelName,
}
coreBySocket[c.PhysicalID] = map[string]struct{}{}
threadsByCoreBySocket[c.PhysicalID] = map[string]int{}
}
if c.CoreID != "" {
coreBySocket[c.PhysicalID][c.PhysicalID+":"+c.CoreID] = struct{}{}
threadsByCoreBySocket[c.PhysicalID][c.PhysicalID+":"+c.CoreID]++
} else {
coreBySocket[c.PhysicalID][c.PhysicalID+":"+c.ID] = struct{}{}
threadsByCoreBySocket[c.PhysicalID][c.PhysicalID+":"+c.ID]++
}
}
// Tally up the values from the tracking maps
for id, s := range socketByID {
s.CoreCount = len(coreBySocket[id])
s.ThreadCount = 0
// This only works if HT is enabled, consider a more reliable model, maybe cache size comparisons?
efficiencyCoreCount := 0
for _, threads := range threadsByCoreBySocket[id] {
s.ThreadCount += threads
if threads == 1 {
efficiencyCoreCount++
}
}
if efficiencyCoreCount == s.CoreCount {
// 1:1 mapping means they're not actually efficiency cores, but regular cores
s.EfficiencyCoreCount = 0
} else {
s.EfficiencyCoreCount = efficiencyCoreCount
}
}
keys := make([]string, 0, len(socketByID))
result := make([]CPU, 0, len(socketByID))
for k := range socketByID {
keys = append(keys, k)
}
sort.Strings(keys)
for _, k := range keys {
result = append(result, *socketByID[k])
}
return result
}
func IsNUMA() bool {
ids := map[string]any{}
packageIds, _ := filepath.Glob("/sys/devices/system/cpu/cpu*/topology/physical_package_id")
for _, packageId := range packageIds {
id, err := os.ReadFile(packageId)
if err == nil {
ids[strings.TrimSpace(string(id))] = struct{}{}
}
}
return len(ids) > 1
}