matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-18 19:56:59 +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- Removed outdated CI/CD workflows (tesla-k80-*.yml)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

39
model/models/llama4/model.go
View File

@@ -18,7 +18,7 @@ type Model struct {
model.BytePairEncoding
ImageProcessor
*VisionModel `gguf:"v,vision"`
*VisionModel `gguf:"v"`
*Projector `gguf:"mm"`
*TextModel
}
@@ -34,8 +34,6 @@ func (p *Projector) Forward(ctx ml.Context, visionOutputs ml.Tensor) ml.Tensor {
func New(c fs.Config) (model.Model, error) {
m := Model{
BytePairEncoding: model.NewBytePairEncoding(
c.String("tokenizer.ggml.pretokenizer",
`[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]*[\p{Ll}\p{Lm}\p{Lo}\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]+[\p{Ll}\p{Lm}\p{Lo}\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n/]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
&model.Vocabulary{
Values: c.Strings("tokenizer.ggml.tokens"),
Types: c.Ints("tokenizer.ggml.token_type"),
@@ -48,6 +46,7 @@ func New(c fs.Config) (model.Model, error) {
c.Ints("tokenizer.ggml.eos_token_ids")...,
),
},
`[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]*[\p{Ll}\p{Lm}\p{Lo}\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]+[\p{Ll}\p{Lm}\p{Lo}\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n/]*|\s*[\r\n]+|\s+(?!\S)|\s+`,
),
ImageProcessor: newImageProcessor(c),
VisionModel: newVisionModel(c),
@@ -77,7 +76,7 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input
return nil, err
}
tilesLocal := ctx.Input().FromFloatSlice(pixelsLocal, size.X, size.Y, m.numChannels)
tilesLocal := ctx.Input().FromFloats(pixelsLocal, size.X, size.Y, m.numChannels)
ratioW, ratioH := size.X/m.imageSize, size.Y/m.imageSize
@@ -88,7 +87,7 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input
pixelValues := tilesLocal
if len(pixelsGlobal) > 0 {
tilesGlobal := ctx.Input().FromFloatSlice(pixelsGlobal, m.imageSize, m.imageSize, m.numChannels)
tilesGlobal := ctx.Input().FromFloats(pixelsGlobal, m.imageSize, m.imageSize, m.numChannels)
pixelValues = pixelValues.Concat(ctx, tilesGlobal, 3)
}
@@ -134,16 +133,16 @@ type separator struct {
y bool
}
func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
var result []input.Input
func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
var result []*input.Input
for _, inp := range inputs {
if len(inp.Multimodal) == 0 {
result = append(result, inp)
continue
}
var imageInputs []input.Input
imageInputs = append(imageInputs, input.Input{Token: 200080}) // <|image_start|>
var imageInputs []*input.Input
imageInputs = append(imageInputs, &input.Input{Token: 200080}) // <|image_start|>
for i, mm := range inp.Multimodal {
patchesPerChunk := mm.Tensor.Dim(1)
@@ -151,20 +150,20 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
if i < len(inp.Multimodal)-1 {
separator := mm.Data.(*separator)
imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
imageInputs = append(imageInputs, slices.Repeat([]input.Input{{Token: 200092}}, patchesPerChunk-1)...)
imageInputs = append(imageInputs, &input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
imageInputs = append(imageInputs, slices.Repeat([]*input.Input{{Token: 200092}}, patchesPerChunk-1)...)
if separator.x {
imageInputs = append(imageInputs, input.Input{Token: 200084}) // <|tile_x_separator|>
imageInputs = append(imageInputs, &input.Input{Token: 200084}) // <|tile_x_separator|>
}
if separator.y {
imageInputs = append(imageInputs, input.Input{Token: 200085}) // <|tile_y_separator|>
imageInputs = append(imageInputs, &input.Input{Token: 200085}) // <|tile_y_separator|>
}
} else {
imageInputs = append(imageInputs, input.Input{Token: 200090}) // <|image|>
imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
imageInputs = append(imageInputs, slices.Repeat([]input.Input{{Token: 200092}}, patchesPerChunk-1)...)
imageInputs = append(imageInputs, input.Input{Token: 200080}) // <|image_end|>
imageInputs = append(imageInputs, &input.Input{Token: 200090}) // <|image|>
imageInputs = append(imageInputs, &input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
imageInputs = append(imageInputs, slices.Repeat([]*input.Input{{Token: 200092}}, patchesPerChunk-1)...)
imageInputs = append(imageInputs, &input.Input{Token: 200080}) // <|image_end|>
}
}
@@ -175,10 +174,8 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
}
func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
return m.TextModel.Forward(ctx, batch.Inputs, positions, outputs, batch, m.Cache), nil
positions := ctx.Input().FromInts(batch.Positions, len(batch.Positions))
return m.TextModel.Forward(ctx, batch.Inputs, positions, batch.Outputs, batch, m.Cache), nil
}
func init() {

38
model/models/llama4/model_text.go
View File

@@ -33,8 +33,8 @@ func (sa *TextAttention) Forward(ctx ml.Context, hiddenStates, positions, attent
value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
if useRope {
query = fast.RoPE(ctx, query, positions, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
key = fast.RoPE(ctx, key, positions, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
query = fast.RoPE(ctx, query, positions, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
key = fast.RoPE(ctx, key, positions, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
}
if opts.useQKNorm {
@@ -58,14 +58,14 @@ type TextMLP struct {
}
func (mlp *TextMLP) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOptions) ml.Tensor {
hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenStates))
hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
return mlp.Down.Forward(ctx, hiddenStates)
}
type TextExperts struct {
Gate *nn.Linear `gguf:"ffn_gate_exps"`
Up *nn.Linear `gguf:"ffn_up_exps"`
Down *nn.Linear `gguf:"ffn_down_exps"`
Gate *nn.LinearBatch `gguf:"ffn_gate_exps"`
Up *nn.LinearBatch `gguf:"ffn_up_exps"`
Down *nn.LinearBatch `gguf:"ffn_down_exps"`
}
func (e *TextExperts) Forward(ctx ml.Context, hiddenStates, routerLogits ml.Tensor, opts *TextOptions) ml.Tensor {
@@ -76,9 +76,9 @@ func (e *TextExperts) Forward(ctx ml.Context, hiddenStates, routerLogits ml.Tens
hiddenStates = hiddenStates.Repeat(ctx, 1, opts.numExpertsUsed)
hiddenStates = hiddenStates.Mul(ctx, scores)
upStates := e.Up.Weight.MulmatID(ctx, hiddenStates, experts)
gateStates := e.Gate.Weight.MulmatID(ctx, hiddenStates, experts)
downStates := e.Down.Weight.MulmatID(ctx, upStates.Mul(ctx, gateStates.SILU(ctx)), experts)
upStates := e.Up.Forward(ctx, hiddenStates, experts)
gateStates := e.Gate.Forward(ctx, hiddenStates, experts)
downStates := e.Down.Forward(ctx, upStates.Mul(ctx, gateStates.SILU(ctx)), experts)
nextStates := downStates.View(ctx, 0, hiddenStates.Dim(0), downStates.Stride(2), hiddenStates.Dim(2))
for i := 1; i < opts.numExpertsUsed; i++ {
@@ -88,22 +88,10 @@ func (e *TextExperts) Forward(ctx ml.Context, hiddenStates, routerLogits ml.Tens
return nextStates
}
// TextSharedExpert is TextMLP with different tensor names
type TextSharedExpert struct {
Gate *nn.Linear `gguf:"ffn_gate_shexp"`
Up *nn.Linear `gguf:"ffn_up_shexp"`
Down *nn.Linear `gguf:"ffn_down_shexp"`
}
func (mlp *TextSharedExpert) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOptions) ml.Tensor {
hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenStates))
return mlp.Down.Forward(ctx, hiddenStates)
}
type TextMOE struct {
Router *nn.Linear `gguf:"ffn_gate_inp"`
Experts *TextExperts
SharedExpert *TextSharedExpert
SharedExpert *TextMLP `gguf:",suf:_shexp"`
}
func (moe *TextMOE) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOptions) ml.Tensor {
@@ -196,7 +184,7 @@ func newTextModel(c fs.Config) *TextModel {
numExpertsUsed: int(c.Uint("expert_used_count")),
ropeDim: int(c.Uint("rope.dimension_count")),
ropeBase: c.Float("rope.freq_base"),
ropeScale: c.Float("rope.freq_scale", 1),
ropeScale: c.Float("rope.scaling.factor", 1),
eps: c.Float("attention.layer_norm_rms_epsilon"),
interleaveLayerStep: int(c.Uint("interleave_moe_layer_step", 1)),
noRopeInterval: int(c.Uint("no_rope_interval", 4)),
@@ -223,7 +211,7 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
scales[i] = float32(math.Log(math.Floor(((float64(p)+1.0)/float64(m.attentionFloorScale))+1.0))*m.attentionScale + 1.0)
}
attentionScales = ctx.Input().FromFloatSlice(scales, 1, 1, len(scales))
attentionScales = ctx.Input().FromFloats(scales, 1, 1, len(scales))
}
for i, layer := range m.Layers {
@@ -248,5 +236,5 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
}
func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, m.ropeScale, rope.WithFactors(m.Layers[layer].Attention.RopeFactors)), nil
return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, 1./m.ropeScale, rope.WithFactors(m.Layers[layer].Attention.RopeFactors)), nil
}

2
model/models/llama4/model_vision.go
View File

@@ -245,7 +245,7 @@ func (m *VisionModel) rotaryEmbedding(ctx ml.Context) (ml.Tensor, ml.Tensor) {
}
}
ropeFreqs := ctx.Input().FromFloatSlice(freqs, freqDim/2, numPatches, 2)
ropeFreqs := ctx.Input().FromFloats(freqs, freqDim/2, numPatches, 2)
ropeFreqs = ropeFreqs.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
ropeFreqs = ropeFreqs.Reshape(ctx, freqDim, 1, numPatches)

6
model/models/llama4/process_image.go
View File

@@ -73,7 +73,7 @@ func (p ImageProcessor) bestResolution(img image.Point, possibleResolutions []im
for i, res := range possibleResolutions {
scaleW := float64(res.X) / float64(w)
scaleH := float64(res.Y) / float64(h)
scale := math.Min(scaleW, scaleH)
scale := min(scaleW, scaleH)
scales[i] = scale
}
@@ -124,11 +124,11 @@ func (p ImageProcessor) maxResolution(imageRes, targetRes image.Point) image.Poi
if scaleW < scaleH {
newRes = image.Point{
targetRes.X,
int(math.Min(math.Floor(float64(imageRes.Y)*scaleW), float64(targetRes.Y))),
int(min(math.Floor(float64(imageRes.Y)*scaleW), float64(targetRes.Y))),
}
} else {
newRes = image.Point{
int(math.Min(math.Floor(float64(imageRes.X)*scaleH), float64(targetRes.X))),
int(min(math.Floor(float64(imageRes.X)*scaleH), float64(targetRes.X))),
targetRes.Y,
}
}