mirror of
https://github.com/dogkeeper886/ollama37.git
synced 2025-12-10 07:46:59 +00:00
ef14fb5b26
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with comprehensive build instructions - Documented toolchain constraints and CPU architecture trade-offs - Removed outdated CI/CD workflows (tesla-k80-*.yml) - Cleaned up temporary development artifacts ## Rationale This fork maintains Tesla K80 GPU support (compute 3.7) which was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>
94 lines
2.8 KiB
Go
94 lines
2.8 KiB
Go
package ggml
|
|
|
|
import (
|
|
"bytes"
|
|
"math/rand/v2"
|
|
"os"
|
|
"strings"
|
|
"testing"
|
|
|
|
"github.com/google/go-cmp/cmp"
|
|
)
|
|
|
|
func TestWriteGGUF(t *testing.T) {
|
|
b := bytes.NewBuffer(make([]byte, 2*3))
|
|
for range 8 {
|
|
t.Run("shuffle", func(t *testing.T) {
|
|
t.Parallel()
|
|
|
|
ts := []*Tensor{
|
|
{Name: "token_embd.weight", Shape: []uint64{2, 3}, WriterTo: b},
|
|
{Name: "blk.0.ffn_norm.weight", Shape: []uint64{2, 3}, WriterTo: b},
|
|
{Name: "blk.0.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: b},
|
|
{Name: "blk.1.ffn_up.weight", Shape: []uint64{2, 3}, WriterTo: b},
|
|
{Name: "blk.2.ffn_norm.weight", Shape: []uint64{2, 3}, WriterTo: b},
|
|
{Name: "blk.1.ffn_down.weight", Shape: []uint64{2, 3}, WriterTo: b},
|
|
{Name: "blk.0.attn_k.weight", Shape: []uint64{2, 3}, WriterTo: b},
|
|
{Name: "output_norm.weight", Shape: []uint64{3, 2}, WriterTo: b},
|
|
{Name: "output.weight", Shape: []uint64{3, 2}, WriterTo: b},
|
|
}
|
|
|
|
rand.Shuffle(len(ts), func(i, j int) {
|
|
ts[i], ts[j] = ts[j], ts[i]
|
|
})
|
|
|
|
w, err := os.CreateTemp(t.TempDir(), strings.ReplaceAll(t.Name(), "/", "_")+"*.bin")
|
|
if err != nil {
|
|
t.Fatal(err)
|
|
}
|
|
defer w.Close()
|
|
|
|
if err := WriteGGUF(w, KV{
|
|
"general.architecture": "test",
|
|
"general.alignment": uint32(16),
|
|
"test.key": "value",
|
|
"attention.key": "value2",
|
|
"tokenizer.key": "value3",
|
|
"adapter.key": "value4",
|
|
}, ts); err != nil {
|
|
t.Fatal(err)
|
|
}
|
|
|
|
r, err := os.Open(w.Name())
|
|
if err != nil {
|
|
t.Fatal(err)
|
|
}
|
|
defer r.Close()
|
|
|
|
ff, err := Decode(r, 0)
|
|
if err != nil {
|
|
t.Fatal(err)
|
|
}
|
|
|
|
if diff := cmp.Diff(KV{
|
|
"general.architecture": "test",
|
|
"general.alignment": uint32(16),
|
|
"general.parameter_count": uint64(54),
|
|
"test.key": "value",
|
|
"test.attention.key": "value2",
|
|
"tokenizer.key": "value3",
|
|
"adapter.key": "value4",
|
|
}, ff.KV()); diff != "" {
|
|
t.Errorf("Mismatch (-want +got):\n%s", diff)
|
|
}
|
|
|
|
if diff := cmp.Diff(Tensors{
|
|
Offset: 800,
|
|
items: []*Tensor{
|
|
{Name: "blk.0.attn_k.weight", Offset: 0, Shape: []uint64{2, 3}},
|
|
{Name: "blk.0.attn_norm.weight", Offset: 32, Shape: []uint64{2, 3}},
|
|
{Name: "blk.0.ffn_norm.weight", Offset: 64, Shape: []uint64{2, 3}},
|
|
{Name: "blk.1.ffn_down.weight", Offset: 96, Shape: []uint64{2, 3}},
|
|
{Name: "blk.1.ffn_up.weight", Offset: 128, Shape: []uint64{2, 3}},
|
|
{Name: "blk.2.ffn_norm.weight", Offset: 160, Shape: []uint64{2, 3}},
|
|
{Name: "output.weight", Offset: 192, Shape: []uint64{3, 2}},
|
|
{Name: "output_norm.weight", Offset: 224, Shape: []uint64{3, 2}},
|
|
{Name: "token_embd.weight", Offset: 256, Shape: []uint64{2, 3}},
|
|
},
|
|
}, ff.Tensors(), cmp.AllowUnexported(Tensors{})); diff != "" {
|
|
t.Errorf("Mismatch (-want +got):\n%s", diff)
|
|
}
|
|
})
|
|
}
|
|
}
|