matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-18 11:47:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

Sync with upstream ollama/ollama and restore Tesla K80 (compute 3.7) support

Browse Source 
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- Removed outdated CI/CD workflows (tesla-k80-*.yml)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>
This commit is contained in:
Shang Chieh Tseng
2025-11-05 14:03:05 +08:00
parent fabe2c5cb7
commit ef14fb5b26
817 changed files with 241634 additions and 70888 deletions
Download Patch File Download Diff File Expand all files Collapse all files

121
kvcache/causal_test.go
View File

@@ -96,6 +96,86 @@ func TestSWA(t *testing.T) {
testCache(t, backend, cache, tests)
}
func TestSWASeparateBatches(t *testing.T) {
backend := &testBackend{}
cache := NewSWACache(1, nil)
defer cache.Close()
cache.Init(backend, ml.DTypeF16, 2, 16, 2)
x := float32(math.Inf(-1))
tests := []testCase{
{
name: "First seq 0",
in: []float32{1, 2},
inShape: []int{1, 1, 2},
seqs: []int{0, 0},
pos: []int32{0, 1},
expected: []float32{1, 2},
expectedShape: []int{1, 1, 2},
expectedMask: []float32{
0, x,
0, 0,
},
},
{
name: "Second seq 0",
in: []float32{3, 4},
inShape: []int{1, 1, 2},
seqs: []int{0, 0},
pos: []int32{2, 3},
expected: []float32{2, 3, 4},
expectedShape: []int{1, 1, 3},
expectedMask: []float32{
0, 0, x,
x, 0, 0,
},
},
{
name: "First seq 1",
in: []float32{5, 6},
inShape: []int{1, 1, 2},
seqs: []int{1, 1},
pos: []int32{0, 1},
expected: []float32{5, 6},
expectedShape: []int{1, 1, 2},
expectedMask: []float32{
0, x,
0, 0,
},
},
{
name: "Second seq 1",
in: []float32{7, 8},
inShape: []int{1, 1, 2},
seqs: []int{1, 1},
pos: []int32{2, 3},
expected: []float32{6, 3, 4, 7, 8},
expectedShape: []int{1, 1, 5},
expectedMask: []float32{
0, x, x, 0, x,
x, x, x, 0, 0,
},
},
{
name: "Third seq 0",
in: []float32{9, 10},
inShape: []int{1, 1, 2},
seqs: []int{0, 0},
pos: []int32{4, 5},
expected: []float32{9, 10, 3, 4},
expectedShape: []int{1, 1, 4},
expectedMask: []float32{
0, x, x, 0,
0, 0, x, x,
},
},
}
testCache(t, backend, cache, tests)
}
func TestSWAMem(t *testing.T) {
backend := &testBackend{}
cache := NewSWAMemCache(1, 3, nil)
@@ -397,7 +477,7 @@ func testCache(t *testing.T, backend ml.Backend, cache Cache, tests []testCase)
}
cache.SetLayer(0)
tensor := context.FromFloatSlice(test.in, test.inShape...)
tensor := context.FromFloats(test.in, test.inShape...)
cache.Put(context, tensor, tensor)
out, _, mask := cache.Get(context)
@@ -431,15 +511,15 @@ func TestCanResume(t *testing.T) {
defer context.Close()
err := cache.StartForward(context, input.Batch{
Positions: []int32{0, 1, 2, 3},
Sequences: []int{0, 0, 0, 0},
Positions: []int32{0, 1, 2, 3, 4},
Sequences: []int{0, 0, 0, 0, 0},
}, false)
if err != nil {
t.Fatalf("StartForward failed: %v", err)
}
cache.SetLayer(0)
tensor := context.FromFloatSlice([]float32{1, 2, 3, 4}, 1, 1, 4)
tensor := context.FromFloats([]float32{1, 2, 3, 4, 5}, 1, 1, 5)
cache.Put(context, tensor, tensor)
// with window size 4, nothing has slid out of the window yet
@@ -455,18 +535,21 @@ func TestCanResume(t *testing.T) {
if !cache.CanResume(0, 3) {
t.Errorf("CanResume(0, 3) = false, want true (latest position)")
}
if !cache.CanResume(0, 4) {
t.Errorf("CanResume(0, 4) = false, want true (latest position)")
}
// shift window by adding position 4
// shift window by adding position 5
err = cache.StartForward(context, input.Batch{
Positions: []int32{4, 5},
Sequences: []int{0, 0},
Positions: []int32{5},
Sequences: []int{0},
}, false)
if err != nil {
t.Fatalf("StartForward failed: %v", err)
}
cache.SetLayer(0)
tensor = context.FromFloatSlice([]float32{5, 6}, 1, 1, 2)
tensor = context.FromFloats([]float32{6}, 1, 1, 1)
cache.Put(context, tensor, tensor)
// only the latest position has overlapping windows
@@ -503,28 +586,28 @@ func TestCanResumeSWAMem(t *testing.T) {
defer context.Close()
err := cache.StartForward(context, input.Batch{
Positions: []int32{0, 1, 2, 3, 4, 5},
Sequences: []int{0, 0, 0, 0, 0, 0},
Positions: []int32{0, 1, 2, 3, 4, 5, 6},
Sequences: []int{0, 0, 0, 0, 0, 0, 0},
}, false)
if err != nil {
t.Fatalf("StartForward failed: %v", err)
}
cache.SetLayer(0)
tensor := context.FromFloatSlice([]float32{1, 2, 3, 4, 5, 6}, 1, 1, 6)
tensor := context.FromFloats([]float32{1, 2, 3, 4, 5, 6, 7}, 1, 1, 7)
cache.Put(context, tensor, tensor)
// shift window by adding position 6
// shift window by adding position 7
err = cache.StartForward(context, input.Batch{
Positions: []int32{6, 7},
Sequences: []int{0, 0},
Positions: []int32{7},
Sequences: []int{0},
}, false)
if err != nil {
t.Fatalf("StartForward failed: %v", err)
}
cache.SetLayer(0)
tensor = context.FromFloatSlice([]float32{7, 8}, 1, 1, 2)
tensor = context.FromFloats([]float32{8}, 1, 1, 1)
cache.Put(context, tensor, tensor)
// only the latest position has overlapping windows
@@ -587,7 +670,7 @@ func (c *testContext) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
return c.Empty(dtype, shape...)
}
func (c *testContext) FromFloatSlice(s []float32, shape ...int) ml.Tensor {
func (c *testContext) FromFloats(s []float32, shape ...int) ml.Tensor {
t := c.Empty(ml.DTypeF32, shape...).(*testTensor)
copy(t.data, s)
@@ -595,13 +678,13 @@ func (c *testContext) FromFloatSlice(s []float32, shape ...int) ml.Tensor {
return t
}
func (c *testContext) FromIntSlice(s []int32, shape ...int) ml.Tensor {
func (c *testContext) FromInts(s []int32, shape ...int) ml.Tensor {
f := make([]float32, len(s))
for i := range f {
f[i] = float32(s[i])
}
out := c.FromFloatSlice(f, shape...)
out := c.FromFloats(f, shape...)
out.(*testTensor).dtype = ml.DTypeI32
return out
@@ -613,7 +696,7 @@ func (c *testContext) Arange(start, stop, step float32, dtype ml.DType) ml.Tenso
s = append(s, i)
}
out := c.FromFloatSlice(s, len(s))
out := c.FromFloats(s, len(s))
out.(*testTensor).dtype = dtype
return out
}