Merge branch 'master' into pr/7359

Author: ggerganov

.devops/nix/package.nix

@@ -17,19 +17,18 @@
   rocmPackages,
   vulkan-headers,
   vulkan-loader,
-  clblast,
+  curl,
   useBlas ? builtins.all (x: !x) [
     useCuda
     useMetalKit
-    useOpenCL
     useRocm
     useVulkan
   ] && blas.meta.available,
   useCuda ? config.cudaSupport,
-  useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
+  useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin,
   useMpi ? false, # Increases the runtime closure size by ~700M
-  useOpenCL ? false,
   useRocm ? config.rocmSupport,
+  enableCurl ? true,
   useVulkan ? false,
   llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
 
@@ -56,7 +55,6 @@ let
     ++ lib.optionals useCuda [ "CUDA" ]
     ++ lib.optionals useMetalKit [ "MetalKit" ]
     ++ lib.optionals useMpi [ "MPI" ]
-    ++ lib.optionals useOpenCL [ "OpenCL" ]
     ++ lib.optionals useRocm [ "ROCm" ]
     ++ lib.optionals useVulkan [ "Vulkan" ];
 
@@ -198,19 +196,19 @@ effectiveStdenv.mkDerivation (
       optionals effectiveStdenv.isDarwin darwinBuildInputs
       ++ optionals useCuda cudaBuildInputs
       ++ optionals useMpi [ mpi ]
-      ++ optionals useOpenCL [ clblast ]
       ++ optionals useRocm rocmBuildInputs
       ++ optionals useBlas [ blas ]
-      ++ optionals useVulkan vulkanBuildInputs;
+      ++ optionals useVulkan vulkanBuildInputs
+      ++ optionals enableCurl [ curl ];
 
     cmakeFlags =
       [
         (cmakeBool "LLAMA_BUILD_SERVER" true)
         (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
         (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
+        (cmakeBool "LLAMA_CURL" enableCurl)
         (cmakeBool "GGML_NATIVE" false)
         (cmakeBool "GGML_BLAS" useBlas)
-        (cmakeBool "GGML_CLBLAST" useOpenCL)
         (cmakeBool "GGML_CUDA" useCuda)
         (cmakeBool "GGML_HIPBLAS" useRocm)
         (cmakeBool "GGML_METAL" useMetalKit)
@@ -254,7 +252,6 @@ effectiveStdenv.mkDerivation (
         useCuda
         useMetalKit
         useMpi
-        useOpenCL
         useRocm
         useVulkan
         ;
@@ -281,7 +278,7 @@ effectiveStdenv.mkDerivation (
       # Configurations we don't want even the CI to evaluate. Results in the
       # "unsupported platform" messages. This is mostly a no-op, because
       # cudaPackages would've refused to evaluate anyway.
-      badPlatforms = optionals (useCuda || useOpenCL) lib.platforms.darwin;
+      badPlatforms = optionals useCuda lib.platforms.darwin;
 
       # Configurations that are known to result in build failures. Can be
       # overridden by importing Nixpkgs with `allowBroken = true`.

.github/ISSUE_TEMPLATE/config.yml

@@ -9,5 +9,3 @@ contact_links:
   - name: Want to contribute?
     url: https://github.com/ggerganov/llama.cpp/wiki/contribute
     about: Head to the contribution guide page of the wiki for areas you can help with
-
-

CMakeLists.txt

@@ -42,6 +42,10 @@ endif()
 
 option(BUILD_SHARED_LIBS "build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
 
+if (WIN32)
+    add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
+endif()
+
 #
 # option list
 #

CMakePresets.json

@@ -19,6 +19,7 @@
         "cacheVariables": {
             "CMAKE_EXPORT_COMPILE_COMMANDS": "ON",
             "CMAKE_CXX_COMPILER": "icx",
+            "CMAKE_C_COMPILER": "cl",
             "GGML_SYCL": "ON",
             "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
         }

Makefile

@@ -62,6 +62,11 @@ TEST_TARGETS = \
 	tests/test-tokenizer-1-bpe \
 	tests/test-tokenizer-1-spm
 
+# Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned
+LEGACY_TARGETS = main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
+	simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama \
+	retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm
+
 # Deprecation aliases
 ifdef LLAMA_CUBLAS
 $(error LLAMA_CUBLAS is removed. Use GGML_CUDA instead.)
@@ -1086,6 +1091,7 @@ clean:
 	rm -vrf ggml/src/ggml-cuda/template-instances/*.o
 	rm -rvf $(BUILD_TARGETS)
 	rm -rvf $(TEST_TARGETS)
+	rm -rvf $(LEGACY_TARGETS)
 	find examples pocs -type f -name "*.o" -delete
 
 #

README.md

@@ -108,6 +108,7 @@ Typically finetunes of the base models below are supported as well.
 - [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon)
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
+- [X] [BERT](https://github.com/ggerganov/llama.cpp/pull/5423)
 - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
 - [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft)
 - [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
@@ -217,6 +218,11 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 **Tools:**
 
 - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
+- [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
+
+**Infrastructure:**
+
+- [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
 
 ---
 

common/common.cpp

@@ -757,7 +757,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.cache_type_v = argv[++i];
         return true;
     }
-    if (arg == "--multiline-input") {
+    if (arg == "-mli" || arg == "--multiline-input") {
         params.multiline_input = true;
         return true;
     }

common/common.h

@@ -459,4 +459,3 @@ void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const cha
 void yaml_dump_non_result_info(
     FILE * stream, const gpt_params & params, const llama_context * lctx,
     const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
-

convert-hf-to-gguf-update.py

@@ -86,6 +86,9 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "poro-chat",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", },
     {"name": "jina-v2-code",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", },
     {"name": "viking",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B
+    {"name": "gemma",          "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", },
+    {"name": "gemma-2",        "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", },
+    {"name": "jais",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", },
 ]
 
 
@@ -272,7 +275,8 @@ def get_vocab_base_pre(self, tokenizer) -> str:
     "3333333",
     "33333333",
     "333333333",
-    # "Cửa Việt", # llama-bpe fails on this
+    "Cửa Việt", # llama-bpe fails on this
+    " discards",
     chktxt,
 ]
 

convert-hf-to-gguf.py

@@ -490,6 +490,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee":
             # ref: https://huggingface.co/LumiOpen/Viking-7B
             res = "viking"
+        if chkhsh == "b53802fb28e26d645c3a310b34bfe07da813026ec7c7716883404d5e0f8b1901":
+            # ref: https://huggingface.co/core42/jais-13b
+            res = "jais"
 
         if res is None:
             logger.warning("\n")
@@ -576,7 +579,19 @@ def _set_vocab_qwen(self):
         special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"])
         special_vocab.add_to_gguf(self.gguf_writer)
 
-    def _set_vocab_sentencepiece(self):
+    def _set_vocab_sentencepiece(self, add_to_gguf=True):
+        tokens, scores, toktypes = self._create_vocab_sentencepiece()
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def _create_vocab_sentencepiece(self):
         from sentencepiece import SentencePieceProcessor
 
         tokenizer_path = self.dir_model / 'tokenizer.model'
@@ -638,14 +653,7 @@ def _set_vocab_sentencepiece(self):
                 scores.append(-1000.0)
                 toktypes.append(SentencePieceTokenTypes.UNUSED)
 
-        self.gguf_writer.add_tokenizer_model("llama")
-        self.gguf_writer.add_tokenizer_pre("default")
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_scores(scores)
-        self.gguf_writer.add_token_types(toktypes)
-
-        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
-        special_vocab.add_to_gguf(self.gguf_writer)
+        return tokens, scores, toktypes
 
     def _set_vocab_llama_hf(self):
         vocab = gguf.LlamaHfVocab(self.dir_model)
@@ -1979,7 +1987,7 @@ def set_gguf_parameters(self):
         if len(rope_scaling_type) == 0:
             raise KeyError('Missing the required key rope_scaling.type')
 
-        if rope_scaling_type == 'su':
+        if rope_scaling_type == 'su' or rope_scaling_type == 'longrope':
             attn_factor = math.sqrt(1 + math.log(scale) / math.log(orig_max_pos_embds)) if scale > 1.0 else 1.0
         elif rope_scaling_type == 'yarn':
             attn_factor = 0.1 * math.log(scale) + 1.0 if scale > 1.0 else 1.0
@@ -2353,6 +2361,8 @@ def set_vocab(self):
         special_vocab._set_special_token("eot",    107)
         special_vocab.add_to_gguf(self.gguf_writer)
 
+        self.gguf_writer.add_add_space_prefix(False)
+
     def set_gguf_parameters(self):
         hparams = self.hparams
         block_count = hparams["num_hidden_layers"]
@@ -2390,7 +2400,20 @@ class Gemma2Model(Model):
     model_arch = gguf.MODEL_ARCH.GEMMA2
 
     def set_vocab(self):
-        self._set_vocab_llama_hf()
+        tokens, scores, toktypes = self._create_vocab_sentencepiece()
+        # hack: This is required so that we can properly use start/end-of-turn for chat template
+        for i in range(108):
+            # including <unusedX>, <start_of_turn>, <end_of_turn>
+            toktypes[i] = SentencePieceTokenTypes.CONTROL
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
         self.gguf_writer.add_add_space_prefix(False)
 
     def set_gguf_parameters(self):
@@ -2414,6 +2437,12 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_final_logit_softcapping(
             self.hparams["final_logit_softcapping"]
         )
+        self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
+
+        # sanity check
+        attn_scalar = self.hparams["query_pre_attn_scalar"]
+        if attn_scalar != hparams["hidden_size"] / hparams["num_attention_heads"]:
+            raise ValueError("query_pre_attn_scalar must be equal to n_embd / n_head")
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unusem
@@ -3031,6 +3060,96 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("JAISLMHeadModel")
+class JaisModel(Model):
+    model_arch = gguf.MODEL_ARCH.JAIS
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # SwigLU activation
+        assert self.hparams["activation_function"] == "swiglu"
+        # ALiBi position embedding
+        assert self.hparams["position_embedding_type"] == "alibi"
+
+        # Embeddings scale
+        self.embeddings_scale = 1.0
+        # note: For some JAIS flavors, output is tied to (same as) wte in original model
+        self.output_is_wte = False
+        if 'mup_embeddings_scale' in self.hparams:
+            self.output_is_wte = True   # Hack (?)
+            self.embeddings_scale = self.hparams['mup_embeddings_scale']
+        elif 'embeddings_scale' in self.hparams:
+            self.embeddings_scale = self.hparams['embeddings_scale']
+        else:
+            assert False
+
+        self.width_scale = 1.0
+        if 'mup_output_alpha' in self.hparams:
+            assert 'mup_width_scale' in self.hparams
+            self.width_scale = self.hparams['mup_output_alpha'] * self.hparams['mup_width_scale']
+        elif 'width_scale' in self.hparams:
+            self.width_scale = self.hparams['width_scale']
+        else:
+            assert False
+
+        self.max_alibi_bias = 8.0
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_context_length(self.hparams["n_positions"])
+        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
+        self.gguf_writer.add_feed_forward_length(self.hparams["n_inner"])
+        self.gguf_writer.add_head_count(self.hparams["n_head"])
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        tensors: list[tuple[str, Tensor]] = []
+
+        # we don't need these
+        if name.endswith((".attn.bias")):
+            return tensors
+
+        if name.endswith(("relative_pe.slopes")):
+            # Calculate max ALiBi bias (this is the inverse of the ALiBi calculation)
+            # Some other models has max_alibi_bias spelled out explicitly in the hyperparams,
+            # but Jais's PyTorch model simply precalculates the slope values and places them
+            # in relative_pes.slopes
+            n_head_closest_log2 = 2 ** math.floor(math.log2(self.hparams["n_head"]))
+            first_val = float(data_torch._data[0])
+            self.max_alibi_bias = -round(math.log2(first_val) * n_head_closest_log2)
+
+            return tensors
+
+        if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_fc2.weight")):
+            data_torch = data_torch.transpose(1, 0)
+
+        new_name = self.map_tensor_name(name)
+
+        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
+            tensors.append((new_name, data_torch * self.embeddings_scale))
+            if self.output_is_wte:
+                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch * self.width_scale))
+        elif new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT):
+            assert not self.output_is_wte
+            tensors.append((new_name, data_torch * self.width_scale))
+        else:
+            tensors.append((new_name, data_torch))
+
+        return tensors
+
+    def write_tensors(self):
+        super().write_tensors()
+        self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias)
+
+
 ###### CONVERSION LOGIC ######
 
 
@@ -3186,7 +3305,8 @@ def main() -> None:
         "auto": gguf.LlamaFileType.GUESSED,
     }
 
-    if args.use_temp_file and (args.split_max_tensors > 0 or args.split_max_size != "0"):
+    is_split = args.split_max_tensors > 0 or args.split_max_size != "0"
+    if args.use_temp_file and is_split:
         logger.error("Error: Cannot use temp file when splitting")
         sys.exit(1)
 
@@ -3223,11 +3343,12 @@ def main() -> None:
         if args.vocab_only:
             logger.info("Exporting model vocab...")
             model_instance.write_vocab()
-            logger.info("Model vocab successfully exported.")
+            logger.info(f"Model vocab successfully exported to {model_instance.fname_out}")
         else:
             logger.info("Exporting model...")
             model_instance.write()
-            logger.info("Model successfully exported.")
+            out_path = f"{model_instance.fname_out.parent}{os.sep}" if is_split else model_instance.fname_out
+            logger.info(f"Model successfully exported to {out_path}")
 
 
 if __name__ == '__main__':