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with comprehensive build instructions - Documented toolchain constraints and CPU architecture trade-offs - Removed outdated CI/CD workflows (tesla-k80-*.yml) - Cleaned up temporary development artifacts ## Rationale This fork maintains Tesla K80 GPU support (compute 3.7) which was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>
2025-12-11 16:26:59 +00:00 · 2025-11-05 14:03:05 +08:00
parent fabe2c5cb7
commit ef14fb5b26
817 changed files with 241634 additions and 70888 deletions
--- a/fs/ggml/ggml.go
+++ b/fs/ggml/ggml.go
@@ -7,9 +7,11 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
+	"math"
 	"slices"
 	"strings"

+	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/fs/util/bufioutil"
 )

@@ -55,10 +57,28 @@ func (kv KV) EmbeddingLength() uint64 {
 	return uint64(kv.Uint("embedding_length"))
 }

+func (kv KV) HeadCount() []uint64 {
+	headCountDefault := uint32(1)
+	headCount := kv.UintOrArrayValueAsArray("attention.head_count", headCountDefault)
+	if len(headCount) == 1 {
+		headCountDefault = headCount[0]
+	}
+	nLayers := int(kv.BlockCount())
+	if len(headCount) > nLayers {
+		slog.Warn("got more elements of attention.head_count than layers", "len(headCount)", len(headCount), "layers", nLayers)
+	}
+	out := make([]uint64, nLayers)
+	for i := range nLayers {
+		if i >= len(headCount) {
+			out[i] = uint64(headCountDefault)
+		} else {
+			out[i] = uint64(headCount[i])
+		}
+	}
+	return out
+}
+
 func (kv KV) HeadCountMax() uint64 {
-	// TODO(drifkin): using the max value can cause an overestimation. In the
-	// future if array values become more popular, we can adapt the more invasive
-	// <https://github.com/ollama/ollama/pull/10225>
 	return uint64(kv.UintOrMaxArrayValue("attention.head_count", 1))
 }

@@ -66,6 +86,27 @@ func (kv KV) HeadCountMin() uint64 {
 	return uint64(kv.UintOrMinArrayValue("attention.head_count", 1))
 }

+func (kv KV) HeadCountKV() []uint64 {
+	headCountKVDefault := uint32(1)
+	headCountKV := kv.UintOrArrayValueAsArray("attention.head_count_kv", headCountKVDefault)
+	if len(headCountKV) == 1 {
+		headCountKVDefault = headCountKV[0]
+	}
+	nLayers := int(kv.BlockCount())
+	if len(headCountKV) > nLayers {
+		slog.Warn("got more elements of attention.head_count than layers", "len(headCountKV)", len(headCountKV), "layers", nLayers)
+	}
+	out := make([]uint64, nLayers)
+	for i := range nLayers {
+		if i >= len(headCountKV) {
+			out[i] = uint64(headCountKVDefault)
+		} else {
+			out[i] = uint64(headCountKV[i])
+		}
+	}
+	return out
+}
+
 func (kv KV) HeadCountKVMax() uint64 {
 	return uint64(kv.UintOrMaxArrayValue("attention.head_count_kv", 1))
 }
@@ -98,6 +139,26 @@ func (kv KV) ChatTemplate() string {
 	return kv.String("tokenizer.chat_template")
 }

+// ssm architecture parameters
+
+func (kv KV) SSMConvKernel() uint64 {
+	return uint64(kv.Uint("ssm.conv_kernel"))
+}
+
+func (kv KV) SSMInnerSize() uint64 {
+	return uint64(kv.Uint("ssm.inner_size"))
+}
+
+func (kv KV) SSMStateSize() uint64 {
+	return uint64(kv.Uint("ssm.state_size"))
+}
+
+func (kv KV) SSMGroupCount() uint64 {
+	return uint64(kv.Uint("ssm.group_count"))
+}
+
+// general types
+
 func (kv KV) String(key string, defaultValue ...string) string {
 	val, _ := keyValue(kv, key, append(defaultValue, "")...)
 	return val
@@ -129,22 +190,27 @@ func (kv KV) UintOrMinArrayValue(key string, defaultValue uint32) uint32 {
 }

 func (kv KV) UintOrArrayValue(key string, defaultValue uint32) (uint32, uint32) {
+	arrVal := kv.UintOrArrayValueAsArray(key, defaultValue)
+	return slices.Min(arrVal), slices.Max(arrVal)
+}
+
+func (kv KV) UintOrArrayValueAsArray(key string, defaultValue uint32) []uint32 {
 	if u32, ok := keyValue(kv, key, uint32(0)); ok {
-		return u32, u32
+		return []uint32{u32}
 	} else if u32s, ok := keyValue(kv, key, &array[uint32]{}); ok {
-		min := slices.Min(u32s.values)
-		max := slices.Max(u32s.values)
-		return min, max
+		return u32s.values
 	} else if i32s, ok := keyValue(kv, key, &array[int32]{}); ok {
-		min := slices.Min(i32s.values)
-		max := slices.Max(i32s.values)
-		if min < 0 || max < 0 {
-			slog.Warn("array values are unexpectedly negative", "key", key, "min", min, "max", max)
+		dst := make([]uint32, len(i32s.values))
+		for i, v := range i32s.values {
+			if v < 0 {
+				slog.Warn("array values are unexpectedly negative", "key", key, "i", i, "v", v)
+			}
+			dst[i] = uint32(v)
 		}
-		return uint32(min), uint32(max)
+		return dst
 	}

-	return defaultValue, defaultValue
+	return []uint32{defaultValue}
 }

 func (kv KV) Strings(key string, defaultValue ...[]string) []string {
@@ -176,11 +242,13 @@ func (kv KV) OllamaEngineRequired() bool {
 	return slices.Contains([]string{
 		"gemma3",
 		"gemma3n",
-		"mistral3",
+		"gptoss", "gpt-oss",
 		"llama4",
+		"mistral3",
 		"mllama",
 		"qwen25vl",
-		"gptoss",
+		"qwen3", "qwen3moe",
+		"qwen3vl", "qwen3vlmoe",
 	}, kv.Architecture())
 }

@@ -275,7 +343,7 @@ type Tensor struct {

 func (t Tensor) block() (n int) {
 	if _, err := fmt.Sscanf(t.Name, "blk.%d.", &n); err != nil {
-		return -1
+		return math.MaxInt
 	}

 	return
@@ -288,24 +356,24 @@ func (t Tensor) blockSize() uint64 {
 func (t TensorType) BlockSize() uint64 {
 	switch t {
 	case
-		0,  // F32
-		1,  // F16
-		24, // I8
-		25, // I16
-		26, // I32
-		27, // I64
-		28, // F64
-		30: // BF16
+		TensorTypeF32,
+		TensorTypeF16,
+		TensorTypeI8,
+		TensorTypeI16,
+		TensorTypeI32,
+		TensorTypeI64,
+		TensorTypeF64,
+		TensorTypeBF16:
 		return 1
 	case
-		2,  // Q4_0
-		3,  // Q4_1
-		4,  // MXFP4
-		6,  // Q5_0
-		7,  // Q5_1
-		8,  // Q8_0
-		9,  // Q8_1
-		20: // IQ4_NL
+		TensorTypeQ4_0,
+		TensorTypeQ4_1,
+		TensorTypeQ5_0,
+		TensorTypeQ5_1,
+		TensorTypeQ8_0,
+		TensorTypeQ8_1,
+		tensorTypeIQ4_NL,
+		4, TensorTypeMXFP4:
 		return 32
 	default:
 		return 256
@@ -328,8 +396,6 @@ func (t TensorType) TypeSize() uint64 {
 		return 2 + blockSize/2
 	case TensorTypeQ4_1:
 		return 2 + 2 + blockSize/2
-	case TensorTypeMXFP4:
-		return 1 + blockSize/2
 	case TensorTypeQ5_0:
 		return 2 + 4 + blockSize/2
 	case TensorTypeQ5_1:
@@ -380,6 +446,8 @@ func (t TensorType) TypeSize() uint64 {
 		return blockSize/8 + blockSize/16 + blockSize/32
 	case TensorTypeBF16:
 		return 2
+	case 4, TensorTypeMXFP4:
+		return 1 + blockSize/2
 	default:
 		return 0
 	}
@@ -479,10 +547,14 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, error) {
 	}, nil
 }

-func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string) (kv []uint64, partialOffload, fullOffload uint64) {
+func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string, useFlashAttention bool) (kv []uint64, partialOffload, fullOffload uint64) {
+	context *= uint64(numParallel)
+
 	embedding := f.KV().EmbeddingLength()
 	heads := f.KV().HeadCountMax()
+	headsArr := f.KV().HeadCount()
 	headsKV := f.KV().HeadCountKVMax()
+	headsKVArr := f.KV().HeadCountKV()
 	vocab := uint64(f.KV()["tokenizer.ggml.tokens"].(*array[string]).size)

 	embeddingHeads := f.KV().EmbeddingHeadCountMax()
@@ -492,12 +564,51 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 	layers := f.Tensors().GroupLayers()

 	bytesPerElement := kvCacheBytesPerElement(kvCacheType)
+
+	// Default for models unless special-cased below. These defaults mirror the
+	// cache usage in llama.cpp under the assumption that models without special
+	// cases below will use the llamarunner and caching will be handled by the
+	// llama.cpp layer.
+	//
+	// This also assumes that a layer without heads or headsKV set is recurrent
+	// which is usually the case. Some models (eg nemotronh) use "blocks" in
+	// place of layers where some are MLP blocks that don't have any cache.
+	// Models like this will need a special case below to be accurately
+	// estimated.
 	var kvTotal uint64
 	kv = make([]uint64, f.KV().BlockCount())
+	kvSizeAttn := uint64(0)
+	kvSizeRecurrent := uint64(0)
 	for i := range kv {
-		kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+		headsL := headsArr[i]
+		headsKVL := headsKVArr[i]
+		if headsL > 0 && headsKVL > 0 {
+			// full attention layer
+			// NOTE: Assumes uniform values for all attn layers
+			kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKVL) * bytesPerElement)
+			kvSizeAttn += kv[i]
+		} else {
+			// recurrent layer
+			ssmDConv := f.KV().SSMConvKernel()
+			ssmDState := f.KV().SSMStateSize()
+			ssmDInner := f.KV().SSMInnerSize()
+			ssmNGroups := f.KV().SSMGroupCount()
+			nEmbdR := uint64(0)
+			if ssmDConv > 0 {
+				nEmbdR = (ssmDConv - 1) * (ssmDInner + 2*ssmNGroups*ssmDState)
+			}
+			nEmbdS := ssmDState * ssmDInner
+
+			// recurrent always uses F32 in llama.cpp backend
+			// https://github.com/ggml-org/llama.cpp/blob/master/src/llama-model.cpp#L18644
+			bytesPerElementRecurrent := kvCacheBytesPerElement("f32")
+
+			kv[i] = (nEmbdR + nEmbdS) * uint64(bytesPerElementRecurrent)
+			kvSizeRecurrent += kv[i]
+		}
 		kvTotal += kv[i]
 	}
+	slog.Debug("default cache size estimate", "attention MiB", float32(kvSizeAttn)/(1024.*1024.), "attention bytes", kvSizeAttn, "recurrent MiB", float32(kvSizeRecurrent)/(1024.*1024.), "recurrent bytes", kvSizeRecurrent)

 	switch f.KV().Architecture() {
 	case "llama", "llama4":
@@ -665,7 +776,7 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 					4*qkvBias.Shape[0],
 			)
 		}
-	case "gptoss":
+	case "gptoss", "gpt-oss":
 		kv = make([]uint64, f.KV().BlockCount())
 		for i := range kv {
 			kv[i] = uint64(float64((embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
@@ -675,8 +786,12 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 				kv[i] *= context
 			}
 		}
-		fullOffload = 4 * f.KV().HeadCountMax() / cmp.Or(f.KV().HeadCountKVMin(), 1) * kvTotal / 6
-		partialOffload = fullOffload
+
+		partialOffload = 2 * f.KV().HeadCountMax() / cmp.Or(f.KV().HeadCountKVMin(), 1) * kvTotal / 6
+		if useFlashAttention {
+			// rough estimate of graph size with flash attention on
+			partialOffload = (4*uint64(numParallel) + context>>10 + 110) * format.MebiByte
+		}
 	}

 	return
@@ -751,7 +866,11 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {

 // SupportsKVCacheType checks if the requested cache type is supported
 func (f GGML) SupportsKVCacheType(cacheType string) bool {
-	return slices.Contains([]string{"f16", "q8_0", "q4_0"}, cacheType)
+	if cacheType == "" || cacheType == "f16" {
+		return true
+	}
+
+	return slices.Contains([]string{"q8_0", "q4_0"}, cacheType)
 }

 // SupportsFlashAttention checks if the model supports flash attention
@@ -761,7 +880,7 @@ func (f GGML) SupportsFlashAttention() bool {
 		return false
 	}

-	if f.KV().Architecture() == "gptoss" {
+	if arch := f.KV().Architecture(); slices.Contains([]string{"gemma2"}, arch) {
 		return false
 	}

@@ -771,6 +890,16 @@ func (f GGML) SupportsFlashAttention() bool {
 	return headCountK != 0 && headCountV != 0 && headCountK == headCountV
 }

+// FlashAttention checks if the model should enable flash attention
+func (f GGML) FlashAttention() bool {
+	return slices.Contains([]string{
+		"gemma3",
+		"gptoss", "gpt-oss",
+		"qwen3", "qwen3moe",
+		"qwen3vl", "qwen3vlmoe",
+	}, f.KV().String("general.architecture"))
+}
+
 // kvCacheBytesPerElement returns the number of bytes per element for a given KV cache type
 func kvCacheBytesPerElement(cacheType string) float64 {
 	switch cacheType {
@@ -778,6 +907,8 @@ func kvCacheBytesPerElement(cacheType string) float64 {
 		return 1 // 1/2 of fp16
 	case "q4_0":
 		return 0.5 // 1/4 of fp16
+	case "f32":
+		return 4 // f32 (default for recurrent)
 	default:
 		return 2 // f16 (default)
 	}