deprecate ggml

- remove ggml runner
- automatically pull gguf models when ggml detected
- tell users to update to gguf in the case automatic pull fails

Co-Authored-By: Jeffrey Morgan <jmorganca@gmail.com>

.dockerignore

@@ -3,7 +3,6 @@ ollama
 app
 dist
 scripts
-llm/llama.cpp/ggml
 llm/llama.cpp/gguf
 .env
 .cache

.gitmodules

@@ -1,8 +1,3 @@
-[submodule "llm/llama.cpp/ggml"]
-    path = llm/llama.cpp/ggml
-    url = https://github.com/ggerganov/llama.cpp.git
-    ignore = dirty
-    shallow = true
 [submodule "llm/llama.cpp/gguf"]
     path = llm/llama.cpp/gguf
     url = https://github.com/ggerganov/llama.cpp.git

cmd/cmd.go

@@ -571,12 +571,32 @@ func generate(cmd *cobra.Command, opts generateOptions) error {
 		Images:   images,
 	}
 
+	if err := client.Generate(ctx, &request, fn); err != nil {
+		switch {
+		case errors.Is(err, context.Canceled):
+			return nil
+		case strings.Contains(err.Error(), "unsupported model format"):
+			// pull and retry to see if the model has been updated
+			parts := strings.Split(opts.Model, string(os.PathSeparator))
+			if len(parts) == 1 {
+				// this is a library model, log some info
+				fmt.Fprintln(os.Stderr, "This model is no longer compatible with Ollama. Pulling a new version...")
+			}
+			if err := PullHandler(cmd, []string{opts.Model}); err != nil {
+				fmt.Printf("Error: %s\n", err)
+				return fmt.Errorf("unsupported model, please update this model to gguf format") // relay the original error
+			}
+			// retry
 			if err := client.Generate(ctx, &request, fn); err != nil {
 				if errors.Is(err, context.Canceled) {
 					return nil
 				}
 				return err
 			}
+		default:
+			return err
+		}
+	}
 	if opts.Prompt != "" {
 		fmt.Println()
 		fmt.Println()

docs/modelfile.md

@@ -188,7 +188,7 @@ SYSTEM """<system message>"""
 
 ### ADAPTER
 
-The `ADAPTER` instruction specifies the LoRA adapter to apply to the base model. The value of this instruction should be an absolute path or a path relative to the Modelfile and the file must be in a GGML file format. The adapter should be tuned from the base model otherwise the behaviour is undefined.
+The `ADAPTER` instruction specifies the LoRA adapter to apply to the base model. The value of this instruction should be an absolute path or a path relative to the Modelfile and the file must be in a GGUF file format. The adapter should be tuned from the base model otherwise the behaviour is undefined.
 
 ```modelfile
 ADAPTER ./ollama-lora.bin

llm/ggml.go

@@ -86,74 +86,6 @@ type container interface {
 	Decode(*readSeekOffset) (model, error)
 }
 
-type containerGGML struct{}
-
-func (c *containerGGML) Name() string {
-	return "ggml"
-}
-
-func (c *containerGGML) Decode(ro *readSeekOffset) (model, error) {
-	// file contents aren't decoded
-	ro.Seek(0, io.SeekEnd)
-	return nil, nil
-}
-
-type containerGGMF struct {
-	version uint32
-}
-
-func (c *containerGGMF) Name() string {
-	return "ggmf"
-}
-
-func (c *containerGGMF) Decode(ro *readSeekOffset) (model, error) {
-	var version uint32
-	binary.Read(ro, binary.LittleEndian, &version)
-
-	switch version {
-	case 1:
-	default:
-		return nil, errors.New("invalid version")
-	}
-
-	c.version = version
-
-	// remaining file contents aren't decoded
-	ro.Seek(0, io.SeekEnd)
-
-	return nil, nil
-}
-
-type containerGGJT struct {
-	version uint32
-}
-
-func (c *containerGGJT) Name() string {
-	return "ggjt"
-}
-
-func (c *containerGGJT) Decode(ro *readSeekOffset) (model, error) {
-	var version uint32
-	binary.Read(ro, binary.LittleEndian, &version)
-
-	switch version {
-	case 1, 2, 3:
-	default:
-		return nil, errors.New("invalid version")
-	}
-
-	c.version = version
-
-	// different model types may have different layouts for hyperparameters
-	var llama llamaModel
-	binary.Read(ro, binary.LittleEndian, &llama.hyperparameters)
-
-	// remaining file contents aren't decoded
-	ro.Seek(0, io.SeekEnd)
-
-	return &llama, nil
-}
-
 type containerLORA struct {
 	version uint32
 }
@@ -194,6 +126,8 @@ const (
 	FILE_MAGIC_GGUF_BE = 0x47475546
 )
 
+var ErrUnsupportedFormat = errors.New("unsupported model format")
+
 func DecodeGGML(r io.ReadSeeker) (*GGML, error) {
 	ro := readSeekOffset{ReadSeeker: r}
 
@@ -204,12 +138,8 @@ func DecodeGGML(r io.ReadSeeker) (*GGML, error) {
 
 	var c container
 	switch magic {
-	case FILE_MAGIC_GGML:
+	case FILE_MAGIC_GGML, FILE_MAGIC_GGMF, FILE_MAGIC_GGJT:
-		c = &containerGGML{}
+		return nil, ErrUnsupportedFormat
-	case FILE_MAGIC_GGMF:
-		c = &containerGGMF{}
-	case FILE_MAGIC_GGJT:
-		c = &containerGGJT{}
 	case FILE_MAGIC_GGLA:
 		c = &containerLORA{}
 	case FILE_MAGIC_GGUF_LE:

llm/llama.cpp/generate_darwin_amd64.go

@@ -2,15 +2,6 @@ package llm
 
 //go:generate git submodule init
 
-//go:generate git submodule update --force ggml
-//go:generate git -C ggml apply ../patches/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../patches/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../patches/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
-//go:generate git -C ggml apply ../patches/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
-//go:generate cmake -S ggml -B ggml/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_NAME=Darwin -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
-//go:generate cmake --build ggml/build/cpu --target server --config Release
-//go:generate mv ggml/build/cpu/bin/server ggml/build/cpu/bin/ollama-runner
-
 //go:generate git submodule update --force gguf
 //go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
 //go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_METAL=off -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_NAME=Darwin -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0 -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=on

llm/llama.cpp/generate_darwin_arm64.go

@@ -2,15 +2,6 @@ package llm
 
 //go:generate git submodule init
 
-//go:generate git submodule update --force ggml
-//go:generate git -C ggml apply ../patches/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../patches/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../patches/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
-//go:generate git -C ggml apply ../patches/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
-//go:generate cmake -S ggml -B ggml/build/metal -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
-//go:generate cmake --build ggml/build/metal --target server --config Release
-//go:generate mv ggml/build/metal/bin/server ggml/build/metal/bin/ollama-runner
-
 //go:generate git submodule update --force gguf
 //go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
 //go:generate cmake -S gguf -B gguf/build/metal -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0

llm/llama.cpp/generate_linux.go

@@ -2,15 +2,6 @@ package llm
 
 //go:generate git submodule init
 
-//go:generate git submodule update --force ggml
-//go:generate git -C ggml apply ../patches/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../patches/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../patches/0005-ggml-support-CUDA-s-half-type-for-aarch64-1455-2670.patch
-//go:generate git -C ggml apply ../patches/0001-copy-cuda-runtime-libraries.patch
-//go:generate cmake -S ggml -B ggml/build/cpu -DLLAMA_K_QUANTS=on
-//go:generate cmake --build ggml/build/cpu --target server --config Release
-//go:generate mv ggml/build/cpu/bin/server ggml/build/cpu/bin/ollama-runner
-
 //go:generate git submodule update --force gguf
 //go:generate git -C gguf apply ../patches/0001-copy-cuda-runtime-libraries.patch
 //go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
@@ -18,9 +9,6 @@ package llm
 //go:generate cmake --build gguf/build/cpu --target server --config Release
 //go:generate mv gguf/build/cpu/bin/server gguf/build/cpu/bin/ollama-runner
 
-//go:generate cmake -S ggml -B ggml/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on
-//go:generate cmake --build ggml/build/cuda --target server --config Release
-//go:generate mv ggml/build/cuda/bin/server ggml/build/cuda/bin/ollama-runner
 //go:generate cmake -S gguf -B gguf/build/cuda -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DLLAMA_CUDA_PEER_MAX_BATCH_SIZE=0
 //go:generate cmake --build gguf/build/cuda --target server --config Release
 //go:generate mv gguf/build/cuda/bin/server gguf/build/cuda/bin/ollama-runner

llm/llama.cpp/generate_windows.go

@@ -2,13 +2,6 @@ package llm
 
 //go:generate git submodule init
 
-//go:generate git submodule update --force ggml
-//go:generate git -C ggml apply ../patches/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../patches/0002-34B-model-support.patch
-//go:generate cmake -S ggml -B ggml/build/cpu -DLLAMA_K_QUANTS=on
-//go:generate cmake --build ggml/build/cpu --target server --config Release
-//go:generate cmd /c move ggml\build\cpu\bin\Release\server.exe ggml\build\cpu\bin\Release\ollama-runner.exe
-
 //go:generate git submodule update --force gguf
 //go:generate git -C gguf apply ../patches/0001-update-default-log-target.patch
 //go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off

llm/llama.cpp/ggml

@@ -1 +0,0 @@
-Subproject commit 9e232f0234073358e7031c1b8d7aa45020469a3b

llm/llama.cpp/patches/0001-add-detokenize-endpoint.patch

@@ -1,51 +0,0 @@
-From 032ef7ff2423f5117bb59d42fb71be9cebf0a2de Mon Sep 17 00:00:00 2001
-From: Bruce MacDonald <brucewmacdonald@gmail.com>
-Date: Mon, 28 Aug 2023 18:08:12 -0400
-Subject: [PATCH] add detokenize endpoint
-
----
- examples/server/server.cpp | 21 +++++++++++++++++++++
- 1 file changed, 21 insertions(+)
-
-diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index 9966045..5014691 100644
---- a/examples/server/server.cpp
-+++ b/examples/server/server.cpp
-@@ -1075,6 +1075,12 @@ static json format_tokenizer_response(const std::vector<llama_token> &tokens)
-         {"tokens", tokens}};
- }
- 
-+static json format_detokenized_response(std::string content)
-+{
-+    return json{
-+        {"content", content}};
-+}
-+
- static void parse_options_completion(const json &body, llama_server_context &llama)
- {
-     gpt_params default_params;
-@@ -1361,6 +1367,21 @@ int main(int argc, char **argv)
-         const json data = format_tokenizer_response(tokens);
-         return res.set_content(data.dump(), "application/json"); });
- 
-+    svr.Post("/detokenize", [&llama](const Request &req, Response &res)
-+             {
-+        auto lock = llama.lock();
-+
-+        const json body = json::parse(req.body);
-+        std::string content;
-+        if (body.count("tokens") != 0)
-+        {
-+            const std::vector<llama_token> tokens = body["tokens"];
-+            content = tokens_to_str(llama.ctx, tokens.cbegin(), tokens.cend());
-+        }
-+
-+        const json data = format_detokenized_response(content);
-+        return res.set_content(data.dump(), "application/json"); });
-+
-     svr.Post("/embedding", [&llama](const Request &req, Response &res)
-              {
-         auto lock = llama.lock();
--- 
-2.39.2 (Apple Git-143)
-

llm/llama.cpp/patches/0002-34B-model-support.patch

@@ -1,89 +0,0 @@
-From 6145068a6613c37bb43a7408b5496524bdcfc402 Mon Sep 17 00:00:00 2001
-From: Bruce MacDonald <brucewmacdonald@gmail.com>
-Date: Mon, 28 Aug 2023 18:08:53 -0400
-Subject: [PATCH] 34B model support
-
----
- llama.cpp | 10 ++++++++++
- 1 file changed, 10 insertions(+)
-
-diff --git a/llama.cpp b/llama.cpp
-index f2cbe76..62c5cdf 100644
---- a/llama.cpp
-+++ b/llama.cpp
-@@ -79,6 +79,7 @@ enum e_model {
-     MODEL_7B,
-     MODEL_13B,
-     MODEL_30B,
-+    MODEL_34B,
-     MODEL_65B,
-     MODEL_70B,
- };
-@@ -122,6 +123,7 @@ static std::map<e_model, size_t> MEM_REQ_SCRATCH0(int n_ctx)
-         { MODEL_7B,   ((size_t) n_ctx / 16ull + 100ull) * MB },
-         { MODEL_13B,  ((size_t) n_ctx / 12ull + 120ull) * MB },
-         { MODEL_30B,  ((size_t) n_ctx /  9ull + 160ull) * MB },
-+        { MODEL_34B,  ((size_t) n_ctx / 9ull + 160ull) * MB },
-         { MODEL_65B,  ((size_t) n_ctx /  6ull + 256ull) * MB }, // guess
-         { MODEL_70B,  ((size_t) n_ctx /  7ull + 164ull) * MB },
-     };
-@@ -135,6 +137,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
-         { MODEL_7B,  160ull * MB },
-         { MODEL_13B, 192ull * MB },
-         { MODEL_30B, 256ull * MB },
-+        { MODEL_34B, 256ull * MB },
-         { MODEL_65B, 384ull * MB }, // guess
-         { MODEL_70B, 304ull * MB },
-     };
-@@ -149,6 +152,7 @@ static const std::map<e_model, size_t> & MEM_REQ_EVAL()
-         { MODEL_7B,  10ull * MB },
-         { MODEL_13B, 12ull * MB },
-         { MODEL_30B, 16ull * MB },
-+        { MODEL_34B, 16ull * MB },
-         { MODEL_65B, 24ull * MB }, // guess
-         { MODEL_70B, 24ull * MB },
-     };
-@@ -164,6 +168,7 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
-         { MODEL_7B,   512ull * kB },
-         { MODEL_13B,  640ull * kB },
-         { MODEL_30B,  768ull * kB },
-+        { MODEL_34B,  768ull * kB },
-         { MODEL_65B, 1280ull * kB },
-         { MODEL_70B, 1280ull * kB },
-     };
-@@ -179,6 +184,7 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
-         { MODEL_7B,  128ull },
-         { MODEL_13B, 160ull },
-         { MODEL_30B, 208ull },
-+        { MODEL_34B, 208ull },
-         { MODEL_65B, 256ull },
-         { MODEL_70B, 256ull },
-     };
-@@ -1027,6 +1033,7 @@ static const char * llama_model_type_name(e_model type) {
-         case MODEL_7B: return "7B";
-         case MODEL_13B: return "13B";
-         case MODEL_30B: return "30B";
-+        case MODEL_34B: return "34B";
-         case MODEL_65B: return "65B";
-         case MODEL_70B: return "70B";
-         default: LLAMA_ASSERT(false);
-@@ -1074,6 +1081,7 @@ static void llama_model_load_internal(
-             case 26: model.type = e_model::MODEL_3B; break;
-             case 32: model.type = e_model::MODEL_7B; break;
-             case 40: model.type = e_model::MODEL_13B; break;
-+            case 48: model.type = e_model::MODEL_34B; break;
-             case 60: model.type = e_model::MODEL_30B; break;
-             case 80: model.type = e_model::MODEL_65B; break;
-             default:
-@@ -1094,6 +1102,8 @@ static void llama_model_load_internal(
-             LLAMA_LOG_WARN("%s: warning: assuming 70B model based on GQA == %d\n", __func__, n_gqa);
-             model.type = e_model::MODEL_70B;
-             hparams.f_ffn_mult = 1.3f; // from the params.json of the 70B model
-+        } else if (model.type == e_model::MODEL_34B && n_gqa == 8) {
-+            hparams.f_ffn_mult = 1.0f; // from the params.json of the 34B model
-         }
- 
-         hparams.rope_freq_base  = rope_freq_base;
--- 
-2.39.2 (Apple Git-143)
-

llm/llama.cpp/patches/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch

@@ -1,30 +0,0 @@
-From dadbed99e65252d79f81101a392d0d6497b86caa Mon Sep 17 00:00:00 2001
-From: Shouzheng Liu <lshzh.hi@gmail.com>
-Date: Mon, 21 Aug 2023 06:59:29 -0400
-Subject: [PATCH] metal : fix synchronization in new matrix multiplication
- kernel (#2686)
-
----
- ggml-metal.metal | 3 ++-
- 1 file changed, 2 insertions(+), 1 deletion(-)
-
-diff --git a/ggml-metal.metal b/ggml-metal.metal
-index 3f31252..88d48f6 100644
---- a/ggml-metal.metal
-+++ b/ggml-metal.metal
-@@ -1898,10 +1898,11 @@ kernel void kernel_mul_mm(device const  uchar * src0,
-         threadgroup float *temp_str = ((threadgroup float *)shared_memory) \
-                                       + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
-         for (int i = 0; i < 8; i++) {
-+            threadgroup_barrier(mem_flags::mem_device);
-             simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
-         }
- 
--        threadgroup_barrier(mem_flags::mem_threadgroup);
-+        threadgroup_barrier(mem_flags::mem_device);
-         device float *C = dst + BLOCK_SIZE_M * r0 + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
-         if (sgitg==0) {
-             for (int i = 0; i < n_rows; i++) {
--- 
-2.41.0
-

llm/llama.cpp/patches/0004-metal-add-missing-barriers-for-mul-mat-2699.patch

@@ -1,41 +0,0 @@
-From 14b1d7e6f720dee41ce5a826376df738096d9033 Mon Sep 17 00:00:00 2001
-From: Shouzheng Liu <lshzh.hi@gmail.com>
-Date: Tue, 22 Aug 2023 02:18:40 -0400
-Subject: [PATCH] metal : add missing barriers for mul-mat (#2699)
-
----
- ggml-metal.metal | 5 +++--
- 1 file changed, 3 insertions(+), 2 deletions(-)
-
-diff --git a/ggml-metal.metal b/ggml-metal.metal
-index 88d48f6..ce3541f 100644
---- a/ggml-metal.metal
-+++ b/ggml-metal.metal
-@@ -1850,6 +1850,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
-         //load data and store to threadgroup memory
-         half4x4 temp_a;
-         dequantize_func(x, il, temp_a);
-+        threadgroup_barrier(mem_flags::mem_threadgroup);
-         #pragma unroll(16)
-         for (int i = 0; i < 16; i++) {
-             *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
-@@ -1895,14 +1896,14 @@ kernel void kernel_mul_mm(device const  uchar * src0,
-         }
-     } else {
-         // block is smaller than 64x32, we should avoid writing data outside of the matrix
-+        threadgroup_barrier(mem_flags::mem_threadgroup);
-         threadgroup float *temp_str = ((threadgroup float *)shared_memory) \
-                                       + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
-         for (int i = 0; i < 8; i++) {
--            threadgroup_barrier(mem_flags::mem_device);
-             simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
-         }
- 
--        threadgroup_barrier(mem_flags::mem_device);
-+        threadgroup_barrier(mem_flags::mem_threadgroup);
-         device float *C = dst + BLOCK_SIZE_M * r0 + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
-         if (sgitg==0) {
-             for (int i = 0; i < n_rows; i++) {
--- 
-2.41.0
-

llm/llama.cpp/patches/0005-ggml-support-CUDA-s-half-type-for-aarch64-1455-2670.patch

@@ -1,32 +0,0 @@
-From 1e3bc523d8053a77df3ac7126a84d0297ee97ef6 Mon Sep 17 00:00:00 2001
-From: Kylin <56434533+KyL0N@users.noreply.github.com>
-Date: Tue, 22 Aug 2023 15:14:23 +0800
-Subject: [PATCH] ggml : support CUDA's half type for aarch64(#1455) (#2670)
-
-* ggml: support CUDA's half type for aarch64(#1455)
-support CUDA's half type for aarch64 in ggml_fp16_t definition
-
-* ggml: use __CUDACC__ to recognise nvcc compiler
----
- ggml.h | 5 +++--
- 1 file changed, 3 insertions(+), 2 deletions(-)
-
-diff --git a/ggml.h b/ggml.h
-index 544ad2d..0ec7ec5 100644
---- a/ggml.h
-+++ b/ggml.h
-@@ -259,8 +259,9 @@
- extern "C" {
- #endif
- 
--#ifdef __ARM_NEON
--    // we use the built-in 16-bit float type
-+#if defined(__ARM_NEON) && defined(__CUDACC__)
-+    typedef half ggml_fp16_t;
-+#elif defined(__ARM_NEON)
-     typedef __fp16 ggml_fp16_t;
- #else
-     typedef uint16_t ggml_fp16_t;
--- 
-2.39.2 (Apple Git-143)
-

llm/llama.go

@@ -59,13 +59,12 @@ ws ::= ([ \t\n] ws)?
 var llamaCppEmbed embed.FS
 
 type ModelRunner struct {
-	Type        string // "gguf" or "ggml"
 	Path        string // path to the model runner executable
 	Accelerated bool
 }
 
-func chooseRunners(workDir, runnerType string) []ModelRunner {
+func chooseRunners(workDir string) []ModelRunner {
-	buildPath := path.Join("llama.cpp", runnerType, "build")
+	buildPath := path.Join("llama.cpp", "gguf", "build")
 	var runners []ModelRunner
 
 	// set the runners based on the OS
@@ -73,25 +72,25 @@ func chooseRunners(workDir, runnerType string) []ModelRunner {
 	switch runtime.GOOS {
 	case "darwin":
 		if runtime.GOARCH == "arm64" {
-			runners = []ModelRunner{{Type: runnerType, Path: path.Join(buildPath, "metal", "bin", "ollama-runner")}}
+			runners = []ModelRunner{{Path: path.Join(buildPath, "metal", "bin", "ollama-runner")}}
 		} else {
-			runners = []ModelRunner{{Type: runnerType, Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")}}
+			runners = []ModelRunner{{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")}}
 		}
 	case "linux":
 		runners = []ModelRunner{
-			{Type: runnerType, Path: path.Join(buildPath, "cuda", "bin", "ollama-runner"), Accelerated: true},
+			{Path: path.Join(buildPath, "cuda", "bin", "ollama-runner"), Accelerated: true},
-			{Type: runnerType, Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
+			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
 		}
 	case "windows":
 		// TODO: select windows GPU runner here when available
 		runners = []ModelRunner{
-			{Type: runnerType, Path: path.Join(buildPath, "cuda", "bin", "Release", "ollama-runner.exe"), Accelerated: true},
+			{Path: path.Join(buildPath, "cuda", "bin", "Release", "ollama-runner.exe"), Accelerated: true},
-			{Type: runnerType, Path: path.Join(buildPath, "cpu", "bin", "Release", "ollama-runner.exe")},
+			{Path: path.Join(buildPath, "cpu", "bin", "Release", "ollama-runner.exe")},
 		}
 	default:
 		log.Printf("unknown OS, running on CPU: %s", runtime.GOOS)
 		runners = []ModelRunner{
-			{Type: runnerType, Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
+			{Path: path.Join(buildPath, "cpu", "bin", "ollama-runner")},
 		}
 	}
 
@@ -141,7 +140,7 @@ func chooseRunners(workDir, runnerType string) []ModelRunner {
 		}
 	}
 	if !runnerAvailable {
-		log.Fatalf("%s runner not found", runnerType)
+		log.Fatalf("gguf runner not found")
 	}
 
 	// return the runners to try in priority order
@@ -149,7 +148,6 @@ func chooseRunners(workDir, runnerType string) []ModelRunner {
 	for _, r := range runners {
 		// clean the ModelRunner paths so that they match the OS we are running on
 		localRunnersByPriority = append(localRunnersByPriority, ModelRunner{
-			Type:        r.Type,
 			Path:        filepath.Clean(path.Join(workDir, r.Path)),
 			Accelerated: r.Accelerated,
 		})
@@ -350,6 +348,7 @@ func newLlama(model string, adapters, projectors []string, runners []ModelRunner
 		"--batch-size", fmt.Sprintf("%d", opts.NumBatch),
 		"--n-gpu-layers", fmt.Sprintf("%d", numGPU),
 		"--embedding",
+		"--parallel", "2",
 	}
 
 	if opts.MainGPU > 0 {

llm/llm.go

@@ -76,16 +76,8 @@ func New(workDir, model string, adapters, projectors []string, opts api.Options)
 		}
 	}
 
-	switch ggml.Name() {
-	case "gguf":
-		// TODO: gguf will load these options automatically from the model binary
 	opts.NumGQA = 0
 	opts.RopeFrequencyBase = 0.0
 	opts.RopeFrequencyScale = 0.0
-		return newLlama(model, adapters, projectors, chooseRunners(workDir, "gguf"), ggml.NumLayers(), opts)
+	return newLlama(model, adapters, projectors, chooseRunners(workDir), ggml.NumLayers(), opts)
-	case "ggml", "ggmf", "ggjt", "ggla":
-		return newLlama(model, adapters, projectors, chooseRunners(workDir, "ggml"), ggml.NumLayers(), opts)
-	default:
-		return nil, fmt.Errorf("unknown ggml type: %s", ggml.ModelFamily())
-	}
 }

server/images.go

@@ -418,6 +418,27 @@ func CreateModel(ctx context.Context, name, modelFileDir string, commands []pars
 					return err
 				}
 
+				// if the model is not in gguf format, pull the base model to try and get it in gguf format
+				if fromConfig.ModelFormat != "gguf" {
+					fn(api.ProgressResponse{Status: "updating base model"})
+					if err := PullModel(ctx, c.Args, &RegistryOptions{}, fn); err != nil {
+						log.Printf("error pulling model: %v", err)
+					}
+					// Reset the file pointer to the beginning of the file
+					_, err = fromConfigFile.Seek(0, 0)
+					if err != nil {
+						return fmt.Errorf("update from config after pull: %w", err)
+					}
+					if err := json.NewDecoder(fromConfigFile).Decode(&fromConfig); err != nil {
+						return err
+					}
+				}
+
+				// if the model is still not in gguf format, error out
+				if fromConfig.ModelFormat != "gguf" {
+					return fmt.Errorf("%s is not in gguf format, this base model is not compatible with this version of ollama", c.Args)
+				}
+
 				config.SetModelFormat(fromConfig.ModelFormat)
 				config.SetModelFamily(append(fromConfig.ModelFamilies, fromConfig.ModelFamily)...)
 				config.SetModelType(fromConfig.ModelType)
@@ -456,16 +477,22 @@ func CreateModel(ctx context.Context, name, modelFileDir string, commands []pars
 			defer bin.Close()
 
 			var offset int64
+		CREATE:
 			for {
 				fn(api.ProgressResponse{Status: "creating model layer"})
 
 				bin.Seek(offset, io.SeekStart)
 				ggml, err := llm.DecodeGGML(bin)
-				if errors.Is(err, io.EOF) {
+				if err != nil {
-					break
+					switch {
-				} else if err != nil {
+					case errors.Is(err, io.EOF):
+						break CREATE
+					case errors.Is(err, llm.ErrUnsupportedFormat):
+						return fmt.Errorf("model binary specified in FROM field is not a valid gguf format model, %w", err)
+					default:
 						return err
 					}
+				}
 
 				config.SetModelFormat(ggml.Name())
 				config.SetModelFamily(ggml.ModelFamily())

server/routes.go

@@ -114,7 +114,7 @@ func load(c *gin.Context, modelName string, reqOpts map[string]interface{}, sess
 			// some older models are not compatible with newer versions of llama.cpp
 			// show a generalized compatibility error until there is a better way to
 			// check for model compatibility
-			if strings.Contains(err.Error(), "failed to load model") {
+			if errors.Is(llm.ErrUnsupportedFormat, err) || strings.Contains(err.Error(), "failed to load model") {
 				err = fmt.Errorf("%v: this model may be incompatible with your version of Ollama. If you previously pulled this model, try updating it by running `ollama pull %s`", err, model.ShortName)
 			}