Merge remote-tracking branch 'upstream/master' into ntkv2

Author: cebtenzzre

CMakeLists.txt

@@ -67,7 +67,9 @@ endif()
 option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
 option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
 set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
-option(LLAMA_CUBLAS                          "llama: use cuBLAS"                                OFF)
+option(LLAMA_CUBLAS                          "llama: use CUDA"                                  OFF)
+#option(LLAMA_CUDA_CUBLAS                     "llama: use cuBLAS for prompt processing"          OFF)
+set(LLAMA_CUDA_MMQ_Y       "64" CACHE STRING "llama: y tile size for mmq CUDA kernels")
 option(LLAMA_CUDA_FORCE_DMMV                 "llama: use dmmv instead of mmvq CUDA kernels"     OFF)
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
@@ -251,6 +253,10 @@ if (LLAMA_CUBLAS)
         set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
 
         add_compile_definitions(GGML_USE_CUBLAS)
+#        if (LLAMA_CUDA_CUBLAS)
+#            add_compile_definitions(GGML_CUDA_CUBLAS)
+#        endif()
+        add_compile_definitions(GGML_CUDA_MMQ_Y=${LLAMA_CUDA_MMQ_Y})
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
@@ -271,10 +277,14 @@ if (LLAMA_CUBLAS)
         endif()
 
     if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+        # 52 == lowest CUDA 12 standard
+        # 60 == f16 CUDA intrinsics
+        # 61 == integer CUDA intrinsics
+        # 70 == (assumed) compute capability at which unrolling a loop in mul_mat_q kernels is faster
         if (LLAMA_CUDA_DMMV_F16)
-            set(CMAKE_CUDA_ARCHITECTURES "60;61") # needed for f16 CUDA intrinsics
+            set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
         else()
-            set(CMAKE_CUDA_ARCHITECTURES "52;61") # lowest CUDA 12 standard + lowest for integer intrinsics
+            set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
@@ -497,6 +507,8 @@ endif()
 add_library(ggml OBJECT
             ggml.c
             ggml.h
+            ggml-alloc.c
+            ggml-alloc.h
             ${GGML_SOURCES_CUDA}
             ${GGML_SOURCES_OPENCL}
             ${GGML_SOURCES_METAL}

Makefile

@@ -194,7 +194,7 @@ ifdef LLAMA_CUBLAS
 	CXXFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
 	LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
 	OBJS      += ggml-cuda.o
-	NVCCFLAGS = --forward-unknown-to-host-compiler
+	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(LLAMA_CUDA_NVCC)
 else
@@ -220,14 +220,25 @@ else ifdef LLAMA_CUDA_DMMV_Y
 else
 	NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
 endif # LLAMA_CUDA_MMV_Y
+ifdef LLAMA_CUDA_F16
+	NVCCFLAGS += -DGGML_CUDA_F16
+endif # LLAMA_CUDA_F16
 ifdef LLAMA_CUDA_DMMV_F16
-	NVCCFLAGS += -DGGML_CUDA_DMMV_F16
+	NVCCFLAGS += -DGGML_CUDA_F16
 endif # LLAMA_CUDA_DMMV_F16
 ifdef LLAMA_CUDA_KQUANTS_ITER
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
 else
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
+ifdef LLAMA_CUDA_MMQ_Y
+	NVCCFLAGS += -DGGML_CUDA_MMQ_Y=$(LLAMA_CUDA_MMQ_Y)
+else
+	NVCCFLAGS += -DGGML_CUDA_MMQ_Y=64
+endif # LLAMA_CUDA_MMQ_Y
+#ifdef LLAMA_CUDA_CUBLAS
+#	NVCCFLAGS += -DGGML_CUDA_CUBLAS
+#endif # LLAMA_CUDA_CUBLAS
 ifdef LLAMA_CUDA_CCBIN
 	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
@@ -318,7 +329,12 @@ $(info )
 ggml.o: ggml.c ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-llama.o: llama.cpp ggml.h ggml-cuda.h ggml-metal.h llama.h llama-util.h
+ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+OBJS += ggml-alloc.o
+
+llama.o: llama.cpp ggml.h ggml-alloc.h ggml-cuda.h ggml-metal.h llama.h llama-util.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 common.o: examples/common.cpp examples/common.h

README.md

@@ -77,6 +77,7 @@ as the main playground for developing new features for the [ggml](https://github
 **Supported models:**
 
 - [X] LLaMA 🦙
+- [x] LLaMA 2 🦙🦙
 - [X] [Alpaca](https://github.com/ggerganov/llama.cpp#instruction-mode-with-alpaca)
 - [X] [GPT4All](https://github.com/ggerganov/llama.cpp#using-gpt4all)
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca)
@@ -399,12 +400,16 @@ Building the program with BLAS support may lead to some performance improvements
 
   The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance:
 
+<!---
+  | LLAMA_CUDA_CUBLAS       | Boolean                |   false | Use cuBLAS instead of custom CUDA kernels for prompt processing. Faster for all quantization formats except for q4_0 and q8_0, especially for k-quants. Increases VRAM usage (700 MiB for 7b, 970 MiB for 13b, 1430 MiB for 33b). |
+--->
   | Option                  | Legal values           | Default | Description |
   |-------------------------|------------------------|---------|-------------|
+  | LLAMA_CUDA_MMQ_Y        | Positive integer >= 32 |      64 | Tile size in y direction when using the custom CUDA kernels for prompt processing. Higher values can be faster depending on the amount of shared memory available. Power of 2 heavily recommended. |
   | LLAMA_CUDA_FORCE_DMMV   | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
   | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_MMV_Y       | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_DMMV_F16     | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels. Can improve performance on relatively recent GPUs. |
+  | LLAMA_CUDA_MMV_Y        | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_F16          | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
   | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
 
 - #### CLBlast
@@ -650,6 +655,19 @@ python3 convert.py pygmalion-7b/ --outtype q4_1
 - The LLaMA models are officially distributed by Facebook and will **never** be provided through this repository.
 - Refer to [Facebook's LLaMA repository](https://github.com/facebookresearch/llama/pull/73/files) if you need to request access to the model data.
 
+### Obtaining and using the Facebook LLaMA 2 model
+
+- Refer to [Facebook's LLaMA download page](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) if you want to access the model data.
+- Alternatively, if you want to save time and space, you can download already converted and quantized models from [TheBloke](https://huggingface.co/TheBloke), including:
+  - [LLaMA 2 7B base](https://huggingface.co/TheBloke/Llama-2-7B-GGML)
+  - [LLaMA 2 13B base](https://huggingface.co/TheBloke/Llama-2-13B-GGML)
+  - [LLaMA 2 70B base](https://huggingface.co/TheBloke/Llama-2-70B-GGML)
+  - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGML)
+  - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGML)
+  - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGML)
+- Specify `-eps 1e-5` for best generation quality
+- Specify `-gqa 8` for 70B models to work
+
 ### Verifying the model files
 
 Please verify the [sha256 checksums](SHA256SUMS) of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files.

convert.py

@@ -133,19 +133,20 @@ def make_tensors_list() -> List[str]:
 
 def find_n_mult(n_ff: int, n_embd: int) -> int:
     # hardcoded magic range
-    for n_mult in range(256, 1, -1):
+    for n_mult in range(8192, 1, -1):
         calc_ff = (((8*n_embd) // 3 + n_mult - 1) // n_mult)*n_mult
         if calc_ff == n_ff:
             return n_mult
     raise Exception(f"failed to find n_mult for (n_ff={n_ff}, n_embd={n_embd}).")
 
 @dataclass
 class Params:
-    n_vocab: int
-    n_embd:  int
-    n_mult:  int
-    n_head:  int
-    n_layer: int
+    n_vocab:   int
+    n_embd:    int
+    n_mult:    int
+    n_head:    int
+    n_layer:   int
+    n_kv_head: Optional[int]  # This parameter is only used for Llama 2
 
     @staticmethod
     def guessed(model: 'LazyModel') -> 'Params':
@@ -167,11 +168,12 @@ def guessed(model: 'LazyModel') -> 'Params':
         n_head=n_embd // 128 # guessed
 
         return Params(
-            n_vocab = n_vocab,
-            n_embd  = n_embd,
-            n_mult  = 256,
-            n_head  = n_head,
-            n_layer = n_layer,
+            n_vocab   = n_vocab,
+            n_embd    = n_embd,
+            n_mult    = 256,
+            n_head    = n_head,
+            n_layer   = n_layer,
+            n_kv_head = None,
         )
 
     @staticmethod
@@ -183,15 +185,17 @@ def loadHFTransformerJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
         n_head  = config["num_attention_heads"];
         n_layer = config["num_hidden_layers"];
         n_ff    = config["intermediate_size"];
+        n_kv_head = config.get("num_key_value_heads")
 
         n_mult = find_n_mult(n_ff, n_embd);
 
         return Params(
-            n_vocab = n_vocab,
-            n_embd  = n_embd,
-            n_mult  = n_mult,
-            n_head  = n_head,
-            n_layer = n_layer,
+            n_vocab   = n_vocab,
+            n_embd    = n_embd,
+            n_mult    = n_mult,
+            n_head    = n_head,
+            n_layer   = n_layer,
+            n_kv_head = n_kv_head,
         )
 
     # LLaMA v2 70B params.json
@@ -200,21 +204,22 @@ def loadHFTransformerJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
     def loadOriginalParamsJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
         config = json.load(open(config_path))
 
-        n_vocab = config["vocab_size"];
-        n_embd  = config["dim"];
-        n_head  = config["n_heads"];
-        n_layer = config["n_layers"];
-        n_mult  = config["multiple_of"];
+        n_vocab   = config["vocab_size"];
+        n_embd    = config["dim"];
+        n_head    = config["n_heads"];
+        n_layer   = config["n_layers"];
+        n_mult    = config["multiple_of"];
 
         if n_vocab == -1:
             n_vocab = model["tok_embeddings.weight"].shape[0]
 
         return Params(
-            n_vocab = n_vocab,
-            n_embd  = n_embd,
-            n_mult  = n_mult,
-            n_head  = n_head,
-            n_layer = n_layer,
+            n_vocab   = n_vocab,
+            n_embd    = n_embd,
+            n_mult    = n_mult,
+            n_head    = n_head,
+            n_layer   = n_layer,
+            n_kv_head = None,
         )
 
     @staticmethod
@@ -317,10 +322,12 @@ def __repr__(self) -> str:
 Vocab = Union[SentencePieceVocab, GGMLVocab]
 
 
-def permute(weights: NDArray, n_head: int) -> NDArray:
+def permute(weights: NDArray, n_head: int, n_kv_head: Optional[int] = None) -> NDArray:
+    if n_kv_head is not None and n_head != n_kv_head:
+        n_head //= n_kv_head
     return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
-                   .swapaxes(1, 2)
-                   .reshape(weights.shape))
+                .swapaxes(1, 2)
+                .reshape(weights.shape))
 
 
 def dequantize_q4(qvalues_pack32: NDArray, scales: NDArray, addends: Optional[NDArray], g_idx: Optional[NDArray]) -> NDArray:
@@ -368,7 +375,7 @@ class Tensor(metaclass=ABCMeta):
     @abstractmethod
     def astype(self, data_type: DataType) -> 'Tensor': ...
     @abstractmethod
-    def permute(self, n_head: int) -> 'Tensor': ...
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> 'Tensor': ...
     @abstractmethod
     def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor': ...
     @abstractmethod
@@ -406,8 +413,8 @@ def part(self, n_part: int) -> 'UnquantizedTensor':
         r = self.ndarray.shape[0] // 3
         return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...])
 
-    def permute(self, n_head: int) -> 'UnquantizedTensor':
-        return UnquantizedTensor(permute(self.ndarray, n_head))
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> 'UnquantizedTensor':
+        return UnquantizedTensor(permute(self.ndarray, n_head, n_kv_head))
 
 
 def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, convert: bool = False) -> NDArray:
@@ -455,26 +462,27 @@ def astype(self, data_type: DataType) -> Tensor:
     def to_ggml(self) -> 'GGMLQuantizedTensor':
         return self
 
-    def permute(self, n_head: int) -> 'GGMLQuantizedTensor':
-        return GGMLQuantizedTensor(permute(self.ndarray, n_head), self.shape, self.data_type)
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> 'GGMLQuantizedTensor':
+        return GGMLQuantizedTensor(permute(self.ndarray, n_head, n_kv_head), self.shape, self.data_type)
 
 
 GGMLCompatibleTensor = Union[UnquantizedTensor, GGMLQuantizedTensor]
 
 
 class DeferredPermutedTensor(Tensor):
-    def __init__(self, base: Tensor, n_head: int) -> None:
+    def __init__(self, base: Tensor, n_head: int, n_kv_head: Optional[int] = None) -> None:
         self.base = base
         self.n_head = n_head
+        self.n_kv_head = n_kv_head
         self.data_type = self.base.data_type
 
     def astype(self, data_type: DataType) -> Tensor:
-        return self.base.astype(data_type).permute(self.n_head)
+        return self.base.astype(data_type).permute(self.n_head, self.n_kv_head)
 
     def to_ggml(self) -> GGMLCompatibleTensor:
-        return self.base.to_ggml().permute(self.n_head)
+        return self.base.to_ggml().permute(self.n_head, self.n_kv_head)
 
-    def permute(self, n_head: int) -> Tensor:
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> Tensor:
         raise Exception("shouldn't permute twice")
 
 
@@ -566,8 +574,8 @@ def regroup(self, new_groupsize: int = 32) -> 'GPTQForLLaMaQuantizedTensor':
         ret.data_type = QuantizedDataType(groupsize=new_groupsize, have_addends=True, have_g_idx=False)
         return ret
 
-    def permute(self, n_head: int) -> Tensor:
-        return DeferredPermutedTensor(self, n_head)
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> Tensor:
+        return DeferredPermutedTensor(self, n_head, n_kv_head)
 
     def to_ggml(self) -> GGMLQuantizedTensor:
         # The output format looks like this:
@@ -698,10 +706,10 @@ def merge_multifile_models(models_plus: List[ModelPlus]) -> ModelPlus:
     return ModelPlus(model, paths, format, vocab)
 
 
-def permute_lazy(lazy_tensor: LazyTensor, n_head: int) -> LazyTensor:
+def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_kv_head: Optional[int] = None) -> LazyTensor:
     def load() -> Tensor:
-        return lazy_tensor.load().permute(n_head)
-    return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description)
+        return lazy_tensor.load().permute(n_head, n_kv_head)
+    return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_kv_head}) ' + lazy_tensor.description)
 
 def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int) -> LazyTensor:
     def load() -> Tensor:
@@ -726,7 +734,7 @@ def convert_transformers_to_orig(model: LazyModel, params: Params) -> LazyModel:
     for i in itertools.count():
         if f"model.layers.{i}.self_attn.q_proj.weight" in model:
             out[f"layers.{i}.attention.wq.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head)
-            out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head)
+            out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head, params.n_kv_head)
             out[f"layers.{i}.attention.wv.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"]
         elif f"model.layers.{i}.self_attn.W_pack.weight" in model:
             out[f"layers.{i}.attention.wq.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head)