remove convert-lora-to-ggml.py (#7204 )

metal : fix warnings (skipme) (#0 )
sync : ggml
2026-07-14 08:25:55 +02:00 · 2024-05-12 02:29:33 +02:00 · 2024-05-11 21:38:13 +03:00 · 2024-05-11 21:35:05 +03:00 · 2024-05-11 21:34:21 +03:00 · 2024-05-11 21:33:08 +03:00
24 changed files with 801 additions and 477 deletions
@@ -1281,17 +1281,6 @@ install(
        WORLD_READ
        WORLD_EXECUTE
    DESTINATION ${CMAKE_INSTALL_BINDIR})
-install(
-    FILES convert-lora-to-ggml.py
-    PERMISSIONS
-        OWNER_READ
-        OWNER_WRITE
-        OWNER_EXECUTE
-        GROUP_READ
-        GROUP_EXECUTE
-        WORLD_READ
-        WORLD_EXECUTE
-    DESTINATION ${CMAKE_INSTALL_BINDIR})
 if (LLAMA_METAL)
    install(
        FILES ggml-metal.metal
@@ -365,47 +365,6 @@ function gg_run_open_llama_3b_v2 {

    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log

-    # lora
-    function compare_ppl {
-        qnt="$1"
-        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
-            return 20
-        fi
-
-        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
-        return 0
-    }
-
-    path_lora="../models-mnt/open-llama/3B-v2/lora"
-    path_shakespeare="../models-mnt/shakespeare"
-
-    shakespeare="${path_shakespeare}/shakespeare.txt"
-    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
-
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_config.json
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_model.bin
-    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/shakespeare.txt
-
-    python3 ../convert-lora-to-ggml.py ${path_lora}
-
-    # f16
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
-    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # q8_0
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
-    compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # q8_0 + f16 lora-base
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
-    compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
    set +e
 }

@@ -416,7 +375,6 @@ function gg_sum_open_llama_3b_v2 {
    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -429,11 +387,6 @@ function gg_sum_open_llama_3b_v2 {
    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
-    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
-    gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
-    gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
-    gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }

 # open_llama_7b_v2
@@ -549,48 +502,6 @@ function gg_run_open_llama_7b_v2 {

    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log

-    # lora
-    function compare_ppl {
-        qnt="$1"
-        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
-            return 20
-        fi
-
-        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
-        return 0
-    }
-
-    path_lora="../models-mnt/open-llama/7B-v2/lora"
-    path_shakespeare="../models-mnt/shakespeare"
-
-    shakespeare="${path_shakespeare}/shakespeare.txt"
-    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
-
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_config.json
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_model.bin
-    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/shakespeare.txt
-
-    python3 ../convert-lora-to-ggml.py ${path_lora}
-
-    # f16
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
-    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # currently not supported by the CUDA backend
-    # q8_0
-    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
-    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
-    #compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # q8_0 + f16 lora-base
-    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
-    #compare_ppl "q8_0 / f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
    set +e
 }

@@ -601,7 +512,6 @@ function gg_sum_open_llama_7b_v2 {
    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -614,11 +524,6 @@ function gg_sum_open_llama_7b_v2 {
    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
-    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
-    #gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
-    #gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
-    #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }

 # bge-small
@@ -145,8 +145,17 @@ for model in models:
    if tokt == TOKENIZER_TYPE.SPM:
        continue

+    # Skip if the tokenizer folder does not exist or there are other download issues previously
+    if not os.path.exists(f"models/tokenizers/{name}"):
+        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
+        continue
+
    # create the tokenizer
-    tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+    try:
+        tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+    except OSError as e:
+        logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
+        continue  # Skip to the next model if the tokenizer can't be loaded

    chktok = tokenizer.encode(chktxt)
    chkhsh = sha256(str(chktok).encode()).hexdigest()
@@ -287,8 +296,17 @@ for model in models:
    name = model["name"]
    tokt = model["tokt"]

+    # Skip if the tokenizer folder does not exist or there are other download issues previously
+    if not os.path.exists(f"models/tokenizers/{name}"):
+        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
+        continue
+
    # create the tokenizer
-    tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+    try:
+        tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+    except OSError as e:
+        logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
+        continue  # Skip this model and continue with the next one in the loop

    with open(f"models/ggml-vocab-{name}.gguf.inp", "w", encoding="utf-8") as f:
        for text in tests:
@@ -12,7 +12,7 @@ import sys
 from enum import IntEnum
 from pathlib import Path
 from hashlib import sha256
-from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Sequence, TypeVar, cast, overload
+from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Sequence, TypeVar, cast

 import numpy as np
 import torch
@@ -48,7 +48,6 @@ class Model:

    dir_model: Path
    ftype: int
-    fname_out: Path
    is_big_endian: bool
    endianess: gguf.GGUFEndian
    use_temp_file: bool
@@ -56,20 +55,20 @@ class Model:
    part_names: list[str]
    is_safetensors: bool
    hparams: dict[str, Any]
-    gguf_writer: gguf.GGUFWriter
    block_count: int
    tensor_map: gguf.TensorNameMap
    tensor_names: set[str] | None
+    fname_out: Path
+    gguf_writer: gguf.GGUFWriter

    # subclasses should define this!
    model_arch: gguf.MODEL_ARCH

-    def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool, use_temp_file: bool, eager: bool):
-        if self.__class__ == Model:
-            raise TypeError(f"{self.__class__.__name__!r} should not be directly instantiated")
+    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool, use_temp_file: bool, eager: bool):
+        if type(self) is Model:
+            raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
        self.dir_model = dir_model
        self.ftype = ftype
-        self.fname_out = fname_out
        self.is_big_endian = is_big_endian
        self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
        self.use_temp_file = use_temp_file
@@ -79,10 +78,23 @@ class Model:
        if not self.is_safetensors:
            self.part_names = Model.get_model_part_names(self.dir_model, ".bin")
        self.hparams = Model.load_hparams(self.dir_model)
-        self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file)
        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
        self.tensor_names = None
+        if self.ftype == gguf.LlamaFileType.GUESSED:
+            # NOTE: can't use field "torch_dtype" in config.json, because some finetunes lie.
+            _, first_tensor = next(self.get_tensors())
+            if first_tensor.dtype == torch.float16:
+                logger.info(f"choosing --outtype f16 from first tensor type ({first_tensor.dtype})")
+                self.ftype = gguf.LlamaFileType.MOSTLY_F16
+            else:
+                logger.info(f"choosing --outtype bf16 from first tensor type ({first_tensor.dtype})")
+                self.ftype = gguf.LlamaFileType.MOSTLY_BF16
+        ftype_up: str = self.ftype.name.partition("_")[2].upper()
+        ftype_lw: str = ftype_up.lower()
+        # allow templating the file name with the output ftype, useful with the "auto" ftype
+        self.fname_out = fname_out.parent / fname_out.name.format(ftype_lw, outtype=ftype_lw, ftype=ftype_lw, OUTTYPE=ftype_up, FTYPE=ftype_up)
+        self.gguf_writer = gguf.GGUFWriter(self.fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file)

    @classmethod
    def __init_subclass__(cls):
@@ -142,14 +154,27 @@ class Model:
            raise ValueError(f"Mismatch between weight map and model parts for tensor names: {sym_diff}")

    def format_tensor_name(self, key: gguf.MODEL_TENSOR, bid: int | None = None, suffix: str = ".weight") -> str:
-        name: str = gguf.TENSOR_NAMES[key]
        if key not in gguf.MODEL_TENSORS[self.model_arch]:
            raise ValueError(f"Missing {key!r} for MODEL_TENSORS of {self.model_arch!r}")
+        name: str = gguf.TENSOR_NAMES[key]
        if "{bid}" in name:
            assert bid is not None
            name = name.format(bid=bid)
        return name + suffix

+    def match_model_tensor_name(self, name: str, key: gguf.MODEL_TENSOR, bid: int | None, suffix: str = ".weight") -> bool:
+        if key not in gguf.MODEL_TENSORS[self.model_arch]:
+            return False
+        key_name: str = gguf.TENSOR_NAMES[key]
+        if "{bid}" in key_name:
+            if bid is None:
+                return False
+            key_name = key_name.format(bid=bid)
+        else:
+            if bid is not None:
+                return False
+        return name == (key_name + suffix)
+
    def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", ".bias")) -> str:
        new_name = self.tensor_map.get_name(key=name, try_suffixes=try_suffixes)
        if new_name is None:
@@ -215,6 +240,23 @@ class Model:
        return False

    def write_tensors(self):
+        # same as ggml_compute_fp32_to_bf16 in ggml-impl.h
+        def np_fp32_to_bf16(n: np.ndarray):
+            # force nan to quiet
+            n = np.where((n & 0x7fffffff) > 0x7f800000, (n & 0xffff0000) | (64 << 16), n)
+            # flush subnormals to zero
+            n = np.where((n & 0x7f800000) == 0, n & 0x80000000, n)
+            # round to nearest even
+            n = (n + (0x7fff + ((n >> 16) & 1))) >> 16
+            return n.astype(np.int16)
+
+        # Doing this row-wise is much, much faster than element-wise, hence the signature
+        v_fp32_to_bf16 = np.vectorize(np_fp32_to_bf16, otypes=[np.int16], signature="(n)->(n)")
+        if self.lazy:
+            # TODO: find a way to implicitly wrap np.vectorize functions
+            # NOTE: the type is changed to reflect otypes passed to np.vectorize above
+            v_fp32_to_bf16 = gguf.LazyNumpyTensor._wrap_fn(v_fp32_to_bf16, meta_noop=np.int16)
+
        max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")

        for name, data_torch in self.get_tensors():
@@ -239,35 +281,60 @@ class Model:
                data: np.ndarray = data  # type hint
                n_dims = len(data.shape)
                data_dtype = data.dtype
-
-                # if f32 desired, convert any float16 to float32
-                if self.ftype == 0 and data_dtype == np.float16:
-                    data = data.astype(np.float32)
+                data_qtype: gguf.GGMLQuantizationType | None = None

                # when both are True, f32 should win
                extra_f32 = self.extra_f32_tensors(name, new_name, bid, n_dims)
                extra_f16 = self.extra_f16_tensors(name, new_name, bid, n_dims)

                # Most of the codebase that takes in 1D tensors or norms only handles F32 tensors
-                extra_f32 = extra_f32 or n_dims == 1 or new_name.endswith("_norm.weight")
+                # Conditions should closely match those in llama_model_quantize_internal in llama.cpp
+                extra_f32 = any(cond for cond in (
+                    extra_f32,
+                    n_dims == 1,
+                    new_name.endswith("_norm.weight"),
+                ))
+
+                # Some tensor types are always in float32
+                extra_f32 = extra_f32 or any(self.match_model_tensor_name(new_name, key, bid) for key in (
+                    gguf.MODEL_TENSOR.FFN_GATE_INP,
+                    gguf.MODEL_TENSOR.POS_EMBD,
+                    gguf.MODEL_TENSOR.TOKEN_TYPES,
+                ))

                # if f16 desired, convert any float32 2-dim weight tensors to float16
-                extra_f16 = extra_f16 or (name.endswith(".weight") and n_dims >= 2)
+                extra_f16 = any(cond for cond in (
+                    extra_f16,
+                    (name.endswith(".weight") and n_dims >= 2),
+                ))

-                # when both extra_f32 and extra_f16 are False, convert to float32 by default
-                if self.ftype == 1 and data_dtype == np.float16 and (extra_f32 or not extra_f16):
-                    data = data.astype(np.float32)
+                if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
+                    if self.ftype == gguf.LlamaFileType.MOSTLY_F16:
+                        if data_dtype != np.float16:
+                            data = data.astype(np.float16)
+                        data_qtype = gguf.GGMLQuantizationType.F16

-                if self.ftype == 1 and data_dtype == np.float32 and extra_f16 and not extra_f32:
-                    data = data.astype(np.float16)
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
+                        if data_dtype != np.float32:
+                            data = data.astype(np.float32)
+                        data = v_fp32_to_bf16(data.view(np.int32))
+                        assert data.dtype == np.int16
+                        data_qtype = gguf.GGMLQuantizationType.BF16
+
+                else:  # by default, convert to float32
+                    if data_dtype != np.float32:
+                        data = data.astype(np.float32)
+                    data_qtype = gguf.GGMLQuantizationType.F32
+
+                assert data_qtype is not None

                # reverse shape to make it similar to the internal ggml dimension order
                shape_str = f"{{{', '.join(str(n) for n in reversed(data.shape))}}}"

                # n_dims is implicit in the shape
-                logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data.dtype}, shape = {shape_str}")
+                logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")

-                self.gguf_writer.add_tensor(new_name, data)
+                self.gguf_writer.add_tensor(new_name, data, raw_dtype=data_qtype)

    def write(self):
        self.write_tensors()
@@ -2044,12 +2111,6 @@ class BertModel(Model):

        return [(self.map_tensor_name(name), data_torch)]

-    def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
-        del new_name, bid, n_dims  # unused
-
-        # not used with get_rows, must be F32
-        return name == "embeddings.token_type_embeddings.weight"
-

@Model.register("NomicBertModel")
 class NomicBertModel(BertModel):
@@ -2339,92 +2400,40 @@ class JinaBertV2Model(BertModel):


 # tree of lazy tensors
-class LazyTorchTensor:
-    _meta: Tensor
-    _data: Tensor | None
-    _args: tuple
-    _func: Callable[[tuple], Tensor] | None
-
-    def __init__(self, *, meta: Tensor, data: Tensor | None = None, args: tuple = (), func: Callable[[tuple], Tensor] | None = None):
-        self._meta = meta
-        self._data = data
-        self._args = args
-        self._func = func
-
-    @staticmethod
-    def _recurse_apply(o: Any, fn: Callable[[Any], Any]) -> Any:
-        # TODO: dict and set
-        if isinstance(o, (list, tuple)):
-            L = []
-            for item in o:
-                L.append(LazyTorchTensor._recurse_apply(item, fn))
-            if isinstance(o, tuple):
-                L = tuple(L)
-            return L
-        elif isinstance(o, LazyTorchTensor):
-            return fn(o)
-        else:
-            return o
-
-    def _wrap_fn(self, fn: Callable, use_self: bool = False) -> Callable[[Any], LazyTorchTensor]:
-        def wrapped_fn(*args, **kwargs):
-            if kwargs is None:
-                kwargs = {}
-            args = ((self,) if use_self else ()) + args
-
-            meta_args = LazyTorchTensor._recurse_apply(args, lambda t: t._meta)
-
-            return LazyTorchTensor(meta=fn(*meta_args, **kwargs), args=args, func=lambda a: fn(*a, **kwargs))
-        return wrapped_fn
-
-    def __getattr__(self, __name: str) -> Any:
-        meta_attr = getattr(self._meta, __name)
-        if callable(meta_attr):
-            return self._wrap_fn(getattr(torch.Tensor, __name), use_self=True)
-        elif isinstance(meta_attr, torch.Tensor):
-            # for things like self.T
-            return self._wrap_fn(lambda s: getattr(s, __name))(self)
-        else:
-            return meta_attr
+class LazyTorchTensor(gguf.LazyBase):
+    _tensor_type = torch.Tensor
+    # to keep the type-checker happy
+    dtype: torch.dtype
+    shape: torch.Size

+    # only used when converting a torch.Tensor to a np.ndarray
    _dtype_map: dict[torch.dtype, type] = {
        torch.float16: np.float16,
        torch.float32: np.float32,
    }

-    def numpy(self) -> gguf.LazyTensor:
+    def numpy(self) -> gguf.LazyNumpyTensor:
        dtype = self._dtype_map[self.dtype]
-        return gguf.LazyTensor(lambda: LazyTorchTensor.to_eager(self).numpy(), dtype=dtype, shape=self.shape)
+        return gguf.LazyNumpyTensor(
+            meta=np.lib.stride_tricks.as_strided(np.zeros(1, dtype), self.shape, (0 for _ in self.shape)),
+            lazy=self._lazy,
+            args=(self,),
+            func=(lambda s: s[0].numpy())
+        )

-    @overload
-    @staticmethod
-    def to_eager(t: Tensor | LazyTorchTensor) -> Tensor: ...
-
-    @overload
-    @staticmethod
-    def to_eager(t: tuple) -> tuple: ...
-
-    @staticmethod
-    def to_eager(t: Any) -> Any:
-        def simple_to_eager(_t: LazyTorchTensor) -> Tensor:
-            # wake up the lazy tensor
-            if _t._data is None and _t._func is not None:
-                # recurse into its arguments
-                _t._args = LazyTorchTensor.to_eager(_t._args)
-                _t._data = _t._func(_t._args)
-            if _t._data is not None:
-                return _t._data
-            else:
-                raise ValueError(f"Could not compute lazy tensor {_t!r} with args {_t._args!r}")
-
-        # recurse into lists and/or tuples, keeping their structure
-        return LazyTorchTensor._recurse_apply(t, simple_to_eager)
-
-    @staticmethod
-    def from_eager(t: Tensor) -> Tensor:
-        if (t.__class__ == LazyTorchTensor):
+    @classmethod
+    def eager_to_meta(cls, t: Tensor) -> Tensor:
+        if t.is_meta:
            return t
-        return LazyTorchTensor(meta=t.detach().to("meta"), data=t)  # type: ignore
+        return t.detach().to("meta")
+
+    @classmethod
+    def meta_with_dtype(cls, m: Tensor, dtype: torch.dtype) -> Tensor:
+        m = m.detach()
+        if not m.is_meta:
+            m = m.to("meta")
+        m.dtype = dtype
+        return m

    @classmethod
    def __torch_function__(cls, func, types, args=(), kwargs=None):
@@ -2435,28 +2444,8 @@ class LazyTorchTensor:

        if func is torch.Tensor.numpy:
            return args[0].numpy()
-        if func is torch.equal:
-            eager_args = LazyTorchTensor.to_eager(args)
-            return func(*eager_args, **kwargs)

-        return LazyTorchTensor._wrap_fn(args[0], func)(*args, **kwargs)
-
-    # special methods bypass __getattr__, so they need to be added manually
-    # ref: https://docs.python.org/3/reference/datamodel.html#special-lookup
-    # NOTE: LazyTorchTensor can't be a subclass of Tensor (and then be used
-    #       as self._meta is currently used), because then the following
-    #       operations would by default not be wrapped, and so not propagated
-    #       when the tensor is made eager.
-    #       It's better to get non-silent errors for not-yet-supported operators.
-    # TODO: add more when needed to avoid clutter, or find a more concise way
-    def __neg__(self, *args):  # mamba
-        return self._wrap_fn(torch.Tensor.__neg__)(self, *args)
-
-    def __add__(self, *args):  # gemma
-        return self._wrap_fn(torch.Tensor.__add__)(self, *args)
-
-    def __getitem__(self, *args):  # bloom falcon refact internlm2
-        return self._wrap_fn(torch.Tensor.__getitem__)(self, *args)
+        return LazyTorchTensor._wrap_fn(func)(*args, **kwargs)


 def parse_args() -> argparse.Namespace:
@@ -2472,11 +2461,11 @@ def parse_args() -> argparse.Namespace:
    )
    parser.add_argument(
        "--outfile", type=Path,
-        help="path to write to; default: based on input",
+        help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
    )
    parser.add_argument(
-        "--outtype", type=str, choices=["f32", "f16"], default="f16",
-        help="output format - use f32 for float32, f16 for float16",
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "auto"], default="f16",
+        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
    )
    parser.add_argument(
        "--bigendian", action="store_true",
@@ -2530,16 +2519,18 @@ def main() -> None:
        logger.error(f'Error: {args.model} is not a directory')
        sys.exit(1)

-    ftype_map = {
-        "f32": gguf.GGMLQuantizationType.F32,
-        "f16": gguf.GGMLQuantizationType.F16,
+    ftype_map: dict[str, gguf.LlamaFileType] = {
+        "f32": gguf.LlamaFileType.ALL_F32,
+        "f16": gguf.LlamaFileType.MOSTLY_F16,
+        "bf16": gguf.LlamaFileType.MOSTLY_BF16,
+        "auto": gguf.LlamaFileType.GUESSED,
    }

    if args.outfile is not None:
        fname_out = args.outfile
    else:
        # output in the same directory as the model by default
-        fname_out = dir_model / f'ggml-model-{args.outtype}.gguf'
+        fname_out = dir_model / 'ggml-model-{ftype}.gguf'

    logger.info(f"Loading model: {dir_model.name}")

@@ -2555,14 +2546,16 @@ def main() -> None:
        logger.info("Set model tokenizer")
        model_instance.set_vocab()

+        model_instance.gguf_writer.add_quantization_version(gguf.GGML_QUANT_VERSION)
+
        if args.vocab_only:
-            logger.info(f"Exporting model vocab to '{fname_out}'")
+            logger.info(f"Exporting model vocab to '{model_instance.fname_out}'")
            model_instance.write_vocab()
        else:
-            logger.info(f"Exporting model to '{fname_out}'")
+            logger.info(f"Exporting model to '{model_instance.fname_out}'")
            model_instance.write()

-        logger.info(f"Model successfully exported to '{fname_out}'")
+        logger.info(f"Model successfully exported to '{model_instance.fname_out}'")


 if __name__ == '__main__':
@@ -1,150 +0,0 @@
-#!/usr/bin/env python3
-from __future__ import annotations
-
-import logging
-import json
-import os
-import struct
-import sys
-from pathlib import Path
-from typing import Any, BinaryIO, Sequence
-
-import numpy as np
-import torch
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-logging.basicConfig(level=logging.DEBUG)
-logger = logging.getLogger("lora-to-gguf")
-
-NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
-
-
-def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
-    fout.write(b"ggla"[::-1])  # magic (ggml lora)
-    fout.write(struct.pack("i", 1))  # file version
-    fout.write(struct.pack("i", params["r"]))
-    # https://opendelta.readthedocs.io/en/latest/modules/deltas.html says that `lora_alpha` is an int
-    # but some models ship a float value instead
-    # let's convert to int, but fail if lossless conversion is not possible
-    assert (
-        int(params["lora_alpha"]) == params["lora_alpha"]
-    ), "cannot convert float to int losslessly"
-    fout.write(struct.pack("i", int(params["lora_alpha"])))
-
-
-def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_type: np.dtype[Any]) -> None:
-    sname = name.encode("utf-8")
-    fout.write(
-        struct.pack(
-            "iii",
-            len(shape),
-            len(sname),
-            NUMPY_TYPE_TO_FTYPE[data_type.name],
-        )
-    )
-    fout.write(struct.pack("i" * len(shape), *shape[::-1]))
-    fout.write(sname)
-    fout.seek((fout.tell() + 31) & -32)
-
-
-if __name__ == '__main__':
-    if len(sys.argv) < 2:
-        logger.info(f"Usage: python {sys.argv[0]} <path> [arch]")
-        logger.info("Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'")
-        logger.info(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
-        sys.exit(1)
-
-    input_json = os.path.join(sys.argv[1], "adapter_config.json")
-    input_model = os.path.join(sys.argv[1], "adapter_model.bin")
-    output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
-
-    if os.path.exists(input_model):
-        model = torch.load(input_model, map_location="cpu")
-    else:
-        input_model = os.path.join(sys.argv[1], "adapter_model.safetensors")
-        # lazy import load_file only if lora is in safetensors format.
-        from safetensors.torch import load_file
-        model = load_file(input_model, device="cpu")
-
-    arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
-
-    if arch_name not in gguf.MODEL_ARCH_NAMES.values():
-        logger.error(f"Error: unsupported architecture {arch_name}")
-        sys.exit(1)
-
-    arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
-    name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone
-
-    with open(input_json, "r") as f:
-        params = json.load(f)
-
-    if params["peft_type"] != "LORA":
-        logger.error(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
-        sys.exit(1)
-
-    if params["fan_in_fan_out"] is True:
-        logger.error("Error: param fan_in_fan_out is not supported")
-        sys.exit(1)
-
-    if params["bias"] is not None and params["bias"] != "none":
-        logger.error("Error: param bias is not supported")
-        sys.exit(1)
-
-    # TODO: these seem to be layers that have been trained but without lora.
-    # doesn't seem widely used but eventually should be supported
-    if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0:
-        logger.error("Error: param modules_to_save is not supported")
-        sys.exit(1)
-
-    with open(output_path, "wb") as fout:
-        fout.truncate()
-
-        write_file_header(fout, params)
-        for k, v in model.items():
-            orig_k = k
-            if k.endswith(".default.weight"):
-                k = k.replace(".default.weight", ".weight")
-            if k in ["llama_proj.weight", "llama_proj.bias"]:
-                continue
-            if k.endswith("lora_A.weight"):
-                if v.dtype != torch.float16 and v.dtype != torch.float32:
-                    v = v.float()
-                v = v.T
-            else:
-                v = v.float()
-
-            t = v.detach().numpy()
-
-            prefix = "base_model.model."
-            if k.startswith(prefix):
-                k = k[len(prefix) :]
-
-            lora_suffixes = (".lora_A.weight", ".lora_B.weight")
-            if k.endswith(lora_suffixes):
-                suffix = k[-len(lora_suffixes[0]):]
-                k = k[: -len(lora_suffixes[0])]
-            else:
-                logger.error(f"Error: unrecognized tensor name {orig_k}")
-                sys.exit(1)
-
-            tname = name_map.get_name(k)
-            if tname is None:
-                logger.error(f"Error: could not map tensor name {orig_k}")
-                logger.error(" Note: the arch parameter must be specified if the model is not llama")
-                sys.exit(1)
-
-            if suffix == ".lora_A.weight":
-                tname += ".weight.loraA"
-            elif suffix == ".lora_B.weight":
-                tname += ".weight.loraB"
-            else:
-                assert False
-
-            logger.info(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
-            write_tensor_header(fout, tname, t.shape, t.dtype)
-            t.tofile(fout)
-
-    logger.info(f"Converted {input_json} and {input_model} to {output_path}")
@@ -0,0 +1,88 @@
+# Debugging Tests Tips
+
+## How to run & debug a specific test without anything else to keep the feedback loop short?
+
+There is a script called debug-test.sh in the scripts folder whose parameter takes a REGEX and an optional test number.
+
+For example, running the following command will output an interactive list from which you can select a test. It takes this form:
+
+`debug-test.sh [OPTION]... <test_regex> <test_number>`
+
+It will then build & run in the debugger for you.
+
+```bash
+./scripts/debug-test.sh test-tokenizer
+
+# Once in the debugger, i.e. at the chevrons prompt, setting a breakpoint could be as follows:
+>>> b main
+```
+
+For further reference use `debug-test.sh -h` to print help.
+
+&nbsp;
+
+### How does the script work?
+If you want to be able to use the concepts contained in the script separately, the important ones are briefly outlined below.
+
+#### Step 1: Reset and Setup folder context
+
+From base of this repository, let's create `build-ci-debug` as our build context.
+
+```bash
+rm -rf build-ci-debug && mkdir build-ci-debug && cd build-ci-debug
+```
+
+#### Step 2: Setup Build Environment and Compile Test Binaries
+
+Setup and trigger a build under debug mode. You may adapt the arguments as needed, but in this case these are sane defaults.
+
+```bash
+cmake -DCMAKE_BUILD_TYPE=Debug -DLLAMA_CUDA=1 -DLLAMA_FATAL_WARNINGS=ON ..
+make -j
+```
+
+#### Step 3.1: Identify Test Command for Debugging
+
+The output of this command will give you the command & arguments needed to run GDB.
+
+* `-R test-tokenizer` : looks for all the test files named `test-tokenizer*` (R=Regex)
+* `-N` : "show-only" disables test execution & shows test commands that you can feed to GDB.
+* `-V` : Verbose Mode
+
+```bash
+ctest -R "test-tokenizer" -V -N
+```
+
+This may return output similar to below (focusing on key lines to pay attention to):
+
+```bash
+...
+1: Test command: ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf"
+1: Working Directory: .
+Labels: main
+  Test  #1: test-tokenizer-0-llama-spm
+...
+4: Test command: ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-falcon.gguf"
+4: Working Directory: .
+Labels: main
+  Test  #4: test-tokenizer-0-falcon
+...
+```
+
+So for test #1 we can tell these two pieces of relevant information:
+* Test Binary: `~/llama.cpp/build-ci-debug/bin/test-tokenizer-0`
+* Test GGUF Model: `~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf`
+
+#### Step 3.2: Run GDB on test command
+
+Based on the ctest 'test command' report above we can then run a gdb session via this command below:
+
+```bash
+gdb --args ${Test Binary} ${Test GGUF Model}
+```
+
+Example:
+
+```bash
+gdb --args ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf"
+```
@@ -651,9 +651,6 @@ struct server_context {
    std::string              system_prompt;
    std::vector<llama_token> system_tokens;

-    std::string name_user;      // this should be the antiprompt
-    std::string name_assistant;
-
    // slots / clients
    std::vector<server_slot> slots;
    json default_generation_settings_for_props;
@@ -1100,15 +1097,11 @@ struct server_context {
        system_need_update = false;
    }

-    void system_prompt_set(const json & sys_props) {
-        system_prompt  = sys_props.value("prompt", "");
-        name_user      = sys_props.value("anti_prompt", "");
-        name_assistant = sys_props.value("assistant_name", "");
+    bool system_prompt_set(const std::string & sys_prompt) {
+        system_prompt = sys_prompt;

        LOG_VERBOSE("system prompt process", {
            {"system_prompt",  system_prompt},
-            {"name_user",      name_user},
-            {"name_assistant", name_assistant},
        });

        // release all slots
@@ -1117,6 +1110,7 @@ struct server_context {
        }

        system_need_update = true;
+        return true;
    }

    bool process_token(completion_token_output & result, server_slot & slot) {
@@ -1536,7 +1530,8 @@ struct server_context {
                    }

                    if (task.data.contains("system_prompt")) {
-                        system_prompt_set(task.data.at("system_prompt"));
+                        std::string sys_prompt = json_value(task.data, "system_prompt", std::string());
+                        system_prompt_set(sys_prompt);

                        for (server_slot & slot : slots) {
                            slot.n_past    = 0;
@@ -2920,7 +2915,7 @@ int main(int argc, char ** argv) {
    server_params_parse(argc, argv, sparams, params);

    if (!sparams.system_prompt.empty()) {
-        ctx_server.system_prompt_set(json::parse(sparams.system_prompt));
+        ctx_server.system_prompt_set(sparams.system_prompt);
    }

    if (params.model_alias == "unknown") {
@@ -3409,8 +3404,7 @@ int main(int argc, char ** argv) {
    const auto handle_props = [&ctx_server](const httplib::Request & req, httplib::Response & res) {
        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
        json data = {
-            { "user_name",                   ctx_server.name_user.c_str() },
-            { "assistant_name",              ctx_server.name_assistant.c_str() },
+            { "system_prompt",               ctx_server.system_prompt.c_str() },
            { "default_generation_settings", ctx_server.default_generation_settings_for_props },
            { "total_slots",                 ctx_server.params.n_parallel }
        };
@@ -1182,9 +1182,9 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
 static char * fmt_size(size_t size) {
    static char buffer[128];
    if (size >= 1024*1024) {
-        sprintf(buffer, "%zuM", size/1024/1024);
+        snprintf(buffer, sizeof(buffer), "%zuM", size/1024/1024);
    } else {
-        sprintf(buffer, "%zuK", size/1024);
+        snprintf(buffer, sizeof(buffer), "%zuK", size/1024);
    }
    return buffer;
 }
@@ -2204,6 +2204,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                case GGML_UNARY_OP_RELU:
                    ggml_cuda_op_relu(ctx, dst);
                    break;
+                case GGML_UNARY_OP_SIGMOID:
+                    ggml_cuda_op_sigmoid(ctx, dst);
+                    break;
                case GGML_UNARY_OP_HARDSIGMOID:
                    ggml_cuda_op_hardsigmoid(ctx, dst);
                    break;
@@ -2716,6 +2719,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                case GGML_UNARY_OP_GELU:
                case GGML_UNARY_OP_SILU:
                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_SIGMOID:
                case GGML_UNARY_OP_HARDSIGMOID:
                case GGML_UNARY_OP_HARDSWISH:
                case GGML_UNARY_OP_GELU_QUICK:
@@ -48,6 +48,15 @@ static __global__ void relu_f32(const float * x, float * dst, const int k) {
    dst[i] = fmaxf(x[i], 0);
 }

+static __global__ void sigmoid_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = 1.0f / (1.0f + expf(-x[i]));
+}
+
 static __global__ void hardsigmoid_f32(const float * x, float * dst, const int k) {
    const int i = blockDim.x*blockIdx.x + threadIdx.x;

@@ -108,6 +117,11 @@ static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_
    relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }

+static void sigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SIGMOID_BLOCK_SIZE - 1) / CUDA_SIGMOID_BLOCK_SIZE;
+    sigmoid_f32<<<num_blocks, CUDA_SIGMOID_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
 static void hardsigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
    const int num_blocks = (k + CUDA_HARDSIGMOID_BLOCK_SIZE - 1) / CUDA_HARDSIGMOID_BLOCK_SIZE;
    hardsigmoid_f32<<<num_blocks, CUDA_HARDSIGMOID_BLOCK_SIZE, 0, stream>>>(x, dst, k);
@@ -188,6 +202,18 @@ void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
 }

+void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    sigmoid_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
 void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];
    const float * src0_d = (const float *)src0->data;
@@ -4,6 +4,7 @@
 #define CUDA_SILU_BLOCK_SIZE 256
 #define CUDA_TANH_BLOCK_SIZE 256
 #define CUDA_RELU_BLOCK_SIZE 256
+#define CUDA_SIGMOID_BLOCK_SIZE 256
 #define CUDA_HARDSIGMOID_BLOCK_SIZE 256
 #define CUDA_HARDSWISH_BLOCK_SIZE 256
 #define CUDA_SQR_BLOCK_SIZE 256
@@ -18,6 +19,8 @@ void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

 void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

+void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

 void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
@@ -40,6 +40,7 @@ enum ggml_metal_kernel_type {
    GGML_METAL_KERNEL_TYPE_CLAMP,
    GGML_METAL_KERNEL_TYPE_TANH,
    GGML_METAL_KERNEL_TYPE_RELU,
+    GGML_METAL_KERNEL_TYPE_SIGMOID,
    GGML_METAL_KERNEL_TYPE_GELU,
    GGML_METAL_KERNEL_TYPE_GELU_4,
    GGML_METAL_KERNEL_TYPE_GELU_QUICK,
@@ -493,6 +494,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP,                         clamp,                          true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH,                          tanh,                           true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU,                          relu,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIGMOID,                       sigmoid,                        true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU,                          gelu,                           true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4,                        gelu_4,                         true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK,                    gelu_quick,                     true);
@@ -730,6 +732,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
            switch (ggml_get_unary_op(op)) {
                case GGML_UNARY_OP_TANH:
                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_SIGMOID:
                case GGML_UNARY_OP_GELU:
                case GGML_UNARY_OP_GELU_QUICK:
                case GGML_UNARY_OP_SILU:
@@ -1192,24 +1195,24 @@ static enum ggml_status ggml_metal_graph_compute(
                        [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                    } break;
                case GGML_OP_CLAMP:
-                {
-                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CLAMP].pipeline;
+                    {
+                        id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CLAMP].pipeline;

-                    float min;
-                    float max;
-                    memcpy(&min, ((int32_t *) dst->op_params) + 0, sizeof(float));
-                    memcpy(&max, ((int32_t *) dst->op_params) + 1, sizeof(float));
+                        float min;
+                        float max;
+                        memcpy(&min, ((int32_t *) dst->op_params) + 0, sizeof(float));
+                        memcpy(&max, ((int32_t *) dst->op_params) + 1, sizeof(float));

-                    [encoder setComputePipelineState:pipeline];
-                    [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
-                    [encoder setBytes:&min length:sizeof(min) atIndex:2];
-                    [encoder setBytes:&max length:sizeof(max) atIndex:3];
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
+                        [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
+                        [encoder setBytes:&min length:sizeof(min) atIndex:2];
+                        [encoder setBytes:&max length:sizeof(max) atIndex:3];

-                    const int64_t n = ggml_nelements(dst);
+                        const int64_t n = ggml_nelements(dst);

-                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                } break;
+                        [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    } break;
                case GGML_OP_UNARY:
                    switch (ggml_get_unary_op(gf->nodes[i])) {
                        // we are not taking into account the strides, so for now require contiguous tensors
@@ -1237,6 +1240,18 @@ static enum ggml_status ggml_metal_graph_compute(

                                const int64_t n = ggml_nelements(dst);

+                                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            } break;
+                        case GGML_UNARY_OP_SIGMOID:
+                            {
+                                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SIGMOID].pipeline;
+
+                                [encoder setComputePipelineState:pipeline];
+                                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                                const int64_t n = ggml_nelements(dst);
+
                                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                            } break;
                        case GGML_UNARY_OP_GELU:
@@ -229,6 +229,13 @@ kernel void kernel_relu(
    dst[tpig] = max(0.0f, src0[tpig]);
 }

+kernel void kernel_sigmoid(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = 1.0f / (1.0f + exp(-src0[tpig]));
+}
+
 kernel void kernel_tanh(
        device const float * src0,
        device       float * dst,
@@ -2210,7 +2217,7 @@ kernel void kernel_flash_attn_ext_f16(

        // ALiBi
        if (max_bias > 0.0f) {
-            const short h = iq2;
+            const uint32_t h = iq2;

            const float base = h < n_head_log2 ? m0 : m1;
            const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
@@ -2466,7 +2473,7 @@ kernel void kernel_flash_attn_ext_vec_f16(

    // ALiBi
    if (max_bias > 0.0f) {
-        const short h = iq2;
+        const uint32_t h = iq2;

        const float base = h < n_head_log2 ? m0 : m1;
        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
@@ -14,6 +14,12 @@
 #include <stdlib.h> // for qsort
 #include <stdio.h>  // for GGML_ASSERT

+#if defined(_MSC_VER)
+// disable "possible loss of data" to avoid warnings for hundreds of casts
+// we should just be careful :)
+#pragma warning(disable: 4244 4267)
+#endif
+
 #define UNUSED GGML_UNUSED

 // some compilers don't provide _mm256_set_m128i, e.g. gcc 7
@@ -4,7 +4,6 @@
 #include "ggml-impl.h"
 #include "ggml-quants.h"
 #include "ggml.h"
-#include "sgemm.h"

 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW
@@ -37,6 +36,10 @@
 #undef GGML_USE_LLAMAFILE
 #endif

+#ifdef GGML_USE_LLAMAFILE
+#include "sgemm.h"
+#endif
+
 #if defined(_MSC_VER)
 // disable "possible loss of data" to avoid hundreds of casts
 // we should just be careful :)
@@ -1949,6 +1952,7 @@ inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) {
 inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; }
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
 inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
+inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); }
 // TODO: optimize performance
 inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
 inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
@@ -2329,6 +2333,7 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
    "TANH",
    "ELU",
    "RELU",
+    "SIGMOID",
    "GELU",
    "GELU_QUICK",
    "SILU",
@@ -2336,7 +2341,7 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
    "HARDSIGMOID",
 };

-static_assert(GGML_UNARY_OP_COUNT == 12, "GGML_UNARY_OP_COUNT != 12");
+static_assert(GGML_UNARY_OP_COUNT == 13, "GGML_UNARY_OP_COUNT != 13");


 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
@@ -4561,6 +4566,20 @@ struct ggml_tensor * ggml_leaky_relu(
    return result;
 }

+// ggml_sigmoid
+
+struct ggml_tensor * ggml_sigmoid(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary(ctx, a, GGML_UNARY_OP_SIGMOID);
+}
+
+struct ggml_tensor * ggml_sigmoid_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_SIGMOID);
+}
+
 // ggml_gelu

 struct ggml_tensor * ggml_gelu(
@@ -10852,6 +10871,52 @@ static void ggml_compute_forward_relu(
    }
 }

+// ggml_compute_forward_sigmoid
+
+static void ggml_compute_forward_sigmoid_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    assert(params->ith == 0);
+    assert(ggml_are_same_shape(src0, dst));
+
+    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
+        return;
+    }
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    assert(dst->nb[0]  == sizeof(float));
+    assert(src0->nb[0] == sizeof(float));
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_sigmoid_f32(nc,
+                (float *) ((char *) dst->data  + i*( dst->nb[1])),
+                (float *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
+static void ggml_compute_forward_sigmoid(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_sigmoid_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
 // ggml_compute_forward_gelu

 static void ggml_compute_forward_gelu_f32(
@@ -16617,6 +16682,10 @@ static void ggml_compute_forward_unary(
            {
                ggml_compute_forward_relu(params, dst);
            } break;
+        case GGML_UNARY_OP_SIGMOID:
+            {
+                ggml_compute_forward_sigmoid(params, dst);
+            } break;
        case GGML_UNARY_OP_GELU:
            {
                ggml_compute_forward_gelu(params, dst);
@@ -18601,6 +18670,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                        zero_table);
                            }
                        } break;
+                    case GGML_UNARY_OP_SIGMOID:
+                        {
+                            GGML_ASSERT(false); // TODO: not implemented
+                        } break;
                    case GGML_UNARY_OP_GELU:
                        {
                            GGML_ASSERT(false); // TODO: not implemented
@@ -19130,6 +19203,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads, int n_cur_
                case GGML_UNARY_OP_TANH:
                case GGML_UNARY_OP_ELU:
                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_SIGMOID:
                case GGML_UNARY_OP_HARDSWISH: // to opt for multiple threads
                case GGML_UNARY_OP_HARDSIGMOID: // to opt for multiple threads
                    {
@@ -519,6 +519,7 @@ extern "C" {
        GGML_UNARY_OP_TANH,
        GGML_UNARY_OP_ELU,
        GGML_UNARY_OP_RELU,
+        GGML_UNARY_OP_SIGMOID,
        GGML_UNARY_OP_GELU,
        GGML_UNARY_OP_GELU_QUICK,
        GGML_UNARY_OP_SILU,
@@ -1073,6 +1074,14 @@ extern "C" {
            struct ggml_context * ctx,
            struct ggml_tensor  * a);

+    GGML_API struct ggml_tensor * ggml_sigmoid(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_sigmoid_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
    GGML_API struct ggml_tensor * ggml_gelu(
            struct ggml_context * ctx,
            struct ggml_tensor  * a);
@@ -1,4 +1,5 @@
 from .constants import *
+from .lazy import *
 from .gguf_reader import *
 from .gguf_writer import *
 from .tensor_mapping import *
@@ -10,6 +10,7 @@ from typing import Any
 GGUF_MAGIC             = 0x46554747  # "GGUF"
 GGUF_VERSION           = 3
 GGUF_DEFAULT_ALIGNMENT = 32
+GGML_QUANT_VERSION     = 2  # GGML_QNT_VERSION from ggml.h

 #
 # metadata keys
@@ -838,6 +839,49 @@ class GGMLQuantizationType(IntEnum):
    BF16    = 30


+# TODO: add GGMLFileType from ggml_ftype in ggml.h
+
+
+# from llama_ftype in llama.h
+# ALL VALUES SHOULD BE THE SAME HERE AS THEY ARE OVER THERE.
+class LlamaFileType(IntEnum):
+    ALL_F32              = 0
+    MOSTLY_F16           = 1   # except 1d tensors
+    MOSTLY_Q4_0          = 2   # except 1d tensors
+    MOSTLY_Q4_1          = 3   # except 1d tensors
+    MOSTLY_Q4_1_SOME_F16 = 4   # tok_embeddings.weight and output.weight are F16
+    # MOSTLY_Q4_2        = 5   # support has been removed
+    # MOSTLY_Q4_3        = 6   # support has been removed
+    MOSTLY_Q8_0          = 7   # except 1d tensors
+    MOSTLY_Q5_0          = 8   # except 1d tensors
+    MOSTLY_Q5_1          = 9   # except 1d tensors
+    MOSTLY_Q2_K          = 10  # except 1d tensors
+    MOSTLY_Q3_K_S        = 11  # except 1d tensors
+    MOSTLY_Q3_K_M        = 12  # except 1d tensors
+    MOSTLY_Q3_K_L        = 13  # except 1d tensors
+    MOSTLY_Q4_K_S        = 14  # except 1d tensors
+    MOSTLY_Q4_K_M        = 15  # except 1d tensors
+    MOSTLY_Q5_K_S        = 16  # except 1d tensors
+    MOSTLY_Q5_K_M        = 17  # except 1d tensors
+    MOSTLY_Q6_K          = 18  # except 1d tensors
+    MOSTLY_IQ2_XXS       = 19  # except 1d tensors
+    MOSTLY_IQ2_XS        = 20  # except 1d tensors
+    MOSTLY_Q2_K_S        = 21  # except 1d tensors
+    MOSTLY_IQ3_XS        = 22  # except 1d tensors
+    MOSTLY_IQ3_XXS       = 23  # except 1d tensors
+    MOSTLY_IQ1_S         = 24  # except 1d tensors
+    MOSTLY_IQ4_NL        = 25  # except 1d tensors
+    MOSTLY_IQ3_S         = 26  # except 1d tensors
+    MOSTLY_IQ3_M         = 27  # except 1d tensors
+    MOSTLY_IQ2_S         = 28  # except 1d tensors
+    MOSTLY_IQ2_M         = 29  # except 1d tensors
+    MOSTLY_IQ4_XS        = 30  # except 1d tensors
+    MOSTLY_IQ1_M         = 31  # except 1d tensors
+    MOSTLY_BF16          = 32  # except 1d tensors
+
+    GUESSED              = 1024  # not specified in the model file
+
+
 class GGUFEndian(IntEnum):
    LITTLE = 0
    BIG = 1
@@ -7,7 +7,7 @@ import struct
 import tempfile
 from enum import Enum, auto
 from io import BufferedWriter
-from typing import IO, Any, Callable, Sequence, Mapping
+from typing import IO, Any, Sequence, Mapping
 from string import ascii_letters, digits

 import numpy as np
@@ -28,47 +28,6 @@ from .constants import (
 logger = logging.getLogger(__name__)


-class LazyTensor:
-    data: Callable[[], np.ndarray[Any, Any]]
-    # to avoid too deep recursion
-    functions: list[Callable[[np.ndarray[Any, Any]], np.ndarray[Any, Any]]]
-    dtype: np.dtype[Any]
-    shape: tuple[int, ...]
-
-    def __init__(self, data: Callable[[], np.ndarray[Any, Any]], *, dtype: type, shape: tuple[int, ...]):
-        self.data = data
-        self.functions = []
-        self.dtype = np.dtype(dtype)
-        self.shape = shape
-
-    def astype(self, dtype: type, **kwargs) -> LazyTensor:
-        self.functions.append(lambda n: n.astype(dtype, **kwargs))
-        self.dtype = np.dtype(dtype)
-        return self
-
-    @property
-    def nbytes(self) -> int:
-        size = 1
-        for n in self.shape:
-            size *= n
-        return size * self.dtype.itemsize
-
-    def tofile(self, *args, **kwargs) -> None:
-        data = self.data()
-        for f in self.functions:
-            data = f(data)
-        assert data.shape == self.shape
-        assert data.dtype == self.dtype
-        assert data.nbytes == self.nbytes
-        self.functions = []
-        self.data = lambda: data
-        data.tofile(*args, **kwargs)
-
-    def byteswap(self, *args, **kwargs) -> LazyTensor:
-        self.functions.append(lambda n: n.byteswap(*args, **kwargs))
-        return self
-
-
 class WriterState(Enum):
    EMPTY   = auto()
    HEADER  = auto()
@@ -79,7 +38,7 @@ class WriterState(Enum):
 class GGUFWriter:
    fout: BufferedWriter
    temp_file: tempfile.SpooledTemporaryFile[bytes] | None
-    tensors: list[np.ndarray[Any, Any] | LazyTensor]
+    tensors: list[np.ndarray[Any, Any]]
    _simple_value_packing = {
        GGUFValueType.UINT8:   "B",
        GGUFValueType.INT8:    "b",
@@ -278,7 +237,7 @@ class GGUFWriter:
        self.ti_data_count += 1

    def add_tensor(
-        self, name: str, tensor: np.ndarray[Any, Any] | LazyTensor, raw_shape: Sequence[int] | None = None,
+        self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None,
        raw_dtype: GGMLQuantizationType | None = None,
    ) -> None:
        if self.endianess == GGUFEndian.BIG:
@@ -303,7 +262,7 @@ class GGUFWriter:
        if pad != 0:
            fp.write(bytes([0] * pad))

-    def write_tensor_data(self, tensor: np.ndarray[Any, Any] | LazyTensor) -> None:
+    def write_tensor_data(self, tensor: np.ndarray[Any, Any]) -> None:
        if self.state is not WriterState.TI_DATA:
            raise ValueError(f'Expected output file to contain tensor info, got {self.state}')

@@ -391,7 +350,7 @@ class GGUFWriter:
    def add_name(self, name: str) -> None:
        self.add_string(Keys.General.NAME, name)

-    def add_quantization_version(self, quantization_version: GGMLQuantizationType) -> None:
+    def add_quantization_version(self, quantization_version: int) -> None:
        self.add_uint32(
            Keys.General.QUANTIZATION_VERSION, quantization_version)

@@ -0,0 +1,225 @@
+from __future__ import annotations
+from abc import ABC, ABCMeta, abstractmethod
+
+import logging
+from typing import Any, Callable
+from collections import deque
+
+import numpy as np
+from numpy.typing import DTypeLike
+
+
+logger = logging.getLogger(__name__)
+
+
+class LazyMeta(ABCMeta):
+
+    def __new__(cls, name: str, bases: tuple[type, ...], namespace: dict[str, Any], **kwargs):
+        def __getattr__(self, __name: str) -> Any:
+            meta_attr = getattr(self._meta, __name)
+            if callable(meta_attr):
+                return type(self)._wrap_fn(
+                    (lambda s, *args, **kwargs: getattr(s, __name)(*args, **kwargs)),
+                    use_self=self,
+                )
+            elif isinstance(meta_attr, self._tensor_type):
+                # e.g. self.T with torch.Tensor should still be wrapped
+                return type(self)._wrap_fn(lambda s: getattr(s, __name))(self)
+            else:
+                # no need to wrap non-tensor properties,
+                # and they likely don't depend on the actual contents of the tensor
+                return meta_attr
+
+        namespace["__getattr__"] = __getattr__
+
+        # need to make a builder for the wrapped wrapper to copy the name,
+        # or else it fails with very cryptic error messages,
+        # because somehow the same string would end up in every closures
+        def mk_wrap(op_name: str, *, meta_noop: bool = False):
+            # need to wrap the wrapper to get self
+            def wrapped_special_op(self, *args, **kwargs):
+                return type(self)._wrap_fn(
+                    getattr(type(self)._tensor_type, op_name),
+                    meta_noop=meta_noop,
+                )(self, *args, **kwargs)
+            return wrapped_special_op
+
+        # special methods bypass __getattr__, so they need to be added manually
+        # ref: https://docs.python.org/3/reference/datamodel.html#special-lookup
+        # NOTE: doing this from a metaclass is very convenient
+        # TODO: make this even more comprehensive
+        for binary_op in (
+            "lt", "le", "eq", "ne", "ge", "gt", "not"
+            "abs", "add", "and", "floordiv", "invert", "lshift", "mod", "mul", "matmul",
+            "neg", "or", "pos", "pow", "rshift", "sub", "truediv", "xor",
+            "iadd", "iand", "ifloordiv", "ilshift", "imod", "imul", "ior", "irshift", "isub", "ixor",
+            "radd", "rand", "rfloordiv", "rmul", "ror", "rpow", "rsub", "rtruediv", "rxor",
+        ):
+            attr_name = f"__{binary_op}__"
+            # the result of these operators usually has the same shape and dtype as the input,
+            # so evaluation on the meta tensor can be skipped.
+            namespace[attr_name] = mk_wrap(attr_name, meta_noop=True)
+
+        for special_op in (
+            "getitem", "setitem", "len",
+        ):
+            attr_name = f"__{special_op}__"
+            namespace[attr_name] = mk_wrap(attr_name, meta_noop=False)
+
+        return super().__new__(cls, name, bases, namespace, **kwargs)
+
+
+# Tree of lazy tensors
+class LazyBase(ABC, metaclass=LazyMeta):
+    _tensor_type: type
+    _meta: Any
+    _data: Any | None
+    _lazy: deque[LazyBase]  # shared within a graph, to avoid deep recursion when making eager
+    _args: tuple
+    _func: Callable[[tuple], Any] | None
+
+    def __init__(self, *, meta: Any, data: Any | None = None, lazy: deque[LazyBase] | None = None, args: tuple = (), func: Callable[[tuple], Any] | None = None):
+        super().__init__()
+        self._meta = meta
+        self._data = data
+        self._lazy = lazy if lazy is not None else deque()
+        self._args = args
+        self._func = func
+        assert self._func is not None or self._data is not None
+        if self._data is None:
+            self._lazy.append(self)
+
+    def __init_subclass__(cls) -> None:
+        if "_tensor_type" not in cls.__dict__:
+            raise TypeError(f"property '_tensor_type' must be defined for {cls!r}")
+        return super().__init_subclass__()
+
+    @staticmethod
+    def _recurse_apply(o: Any, fn: Callable[[Any], Any]) -> Any:
+        # TODO: dict and set
+        if isinstance(o, (list, tuple)):
+            L = []
+            for item in o:
+                L.append(LazyBase._recurse_apply(item, fn))
+            if isinstance(o, tuple):
+                L = tuple(L)
+            return L
+        elif isinstance(o, LazyBase):
+            return fn(o)
+        else:
+            return o
+
+    @classmethod
+    def _wrap_fn(cls, fn: Callable, *, use_self: LazyBase | None = None, meta_noop: bool | DTypeLike = False) -> Callable[[Any], Any]:
+        def wrapped_fn(*args, **kwargs):
+            if kwargs is None:
+                kwargs = {}
+            args = ((use_self,) if use_self is not None else ()) + args
+
+            meta_args = LazyBase._recurse_apply(args, lambda t: t._meta)
+
+            if isinstance(meta_noop, bool) and not meta_noop:
+                try:
+                    res = fn(*meta_args, **kwargs)
+                except NotImplementedError:
+                    # running some operations on PyTorch's Meta tensors can cause this exception
+                    res = None
+            else:
+                # some operators don't need to actually run on the meta tensors
+                assert len(args) > 0
+                res = args[0]
+                assert isinstance(res, cls)
+                res = res._meta
+                # allow operations to override the dtype
+                if meta_noop is not True:
+                    res = cls.meta_with_dtype(res, meta_noop)
+
+            if isinstance(res, cls._tensor_type):
+                def collect_replace(t: LazyBase):
+                    if collect_replace.shared_lazy is None:
+                        collect_replace.shared_lazy = t._lazy
+                    else:
+                        collect_replace.shared_lazy.extend(t._lazy)
+                        t._lazy = collect_replace.shared_lazy
+
+                # emulating a static variable
+                collect_replace.shared_lazy = None
+
+                LazyBase._recurse_apply(args, collect_replace)
+
+                shared_lazy = collect_replace.shared_lazy
+
+                return cls(meta=cls.eager_to_meta(res), lazy=shared_lazy, args=args, func=lambda a: fn(*a, **kwargs))
+            else:
+                del res  # not needed
+                # non-tensor return likely relies on the contents of the args
+                # (e.g. the result of torch.equal)
+                eager_args = cls.to_eager(args)
+                return fn(*eager_args, **kwargs)
+        return wrapped_fn
+
+    @classmethod
+    def to_eager(cls, t: Any) -> Any:
+        def simple_to_eager(_t: LazyBase) -> Any:
+            def already_eager_to_eager(_t: LazyBase) -> Any:
+                assert _t._data is not None
+                return _t._data
+
+            while _t._data is None:
+                lt = _t._lazy.popleft()
+                if lt._data is not None:
+                    raise ValueError(f"{lt} did not belong in the lazy queue")
+                assert lt._func is not None
+                lt._args = cls._recurse_apply(lt._args, already_eager_to_eager)
+                lt._data = lt._func(lt._args)
+                # sanity check
+                assert lt._data.dtype == lt._meta.dtype
+                assert lt._data.shape == lt._meta.shape
+
+            return _t._data
+
+        # recurse into lists and/or tuples, keeping their structure
+        return cls._recurse_apply(t, simple_to_eager)
+
+    @classmethod
+    def eager_to_meta(cls, t: Any) -> Any:
+        return cls.meta_with_dtype(t, t.dtype)
+
+    # must be overridden, meta tensor init is backend-specific
+    @classmethod
+    @abstractmethod
+    def meta_with_dtype(cls, m: Any, dtype: Any) -> Any: pass
+
+    @classmethod
+    def from_eager(cls, t: Any) -> Any:
+        if type(t) is cls:
+            # already eager
+            return t
+        elif isinstance(t, cls._tensor_type):
+            return cls(meta=cls.eager_to_meta(t), data=t)
+        else:
+            return TypeError(f"{type(t)!r} is not compatible with {cls._tensor_type!r}")
+
+
+class LazyNumpyTensor(LazyBase):
+    _tensor_type = np.ndarray
+
+    @classmethod
+    def meta_with_dtype(cls, m: np.ndarray[Any, Any], dtype: DTypeLike) -> np.ndarray[Any, Any]:
+        # The initial idea was to use np.nan as the fill value,
+        # but non-float types like np.int16 can't use that.
+        # So zero it is.
+        cheat = np.zeros(1, dtype)
+        return np.lib.stride_tricks.as_strided(cheat, m.shape, (0 for _ in m.shape))
+
+    def astype(self, dtype, *args, **kwargs):
+        meta = type(self).meta_with_dtype(self._meta, dtype)
+        full_args = (self, dtype,) + args
+        # very important to pass the shared _lazy deque, or else there's an infinite loop somewhere.
+        return type(self)(meta=meta, args=full_args, lazy=self._lazy, func=(lambda a: a[0].astype(*a[1:], **kwargs)))
+
+    def tofile(self, *args, **kwargs):
+        eager = LazyNumpyTensor.to_eager(self)
+        return eager.tofile(*args, **kwargs)
+
+    # TODO: __array_function__
@@ -9,5 +9,4 @@
 -r ./requirements/requirements-convert-hf-to-gguf.txt
 -r ./requirements/requirements-convert-hf-to-gguf-update.txt
 -r ./requirements/requirements-convert-llama-ggml-to-gguf.txt
-r ./requirements/requirements-convert-lora-to-ggml.txt
 -r ./requirements/requirements-convert-persimmon-to-gguf.txt
@@ -1,2 +0,0 @@
-r ./requirements-convert.txt
-torch~=2.1.1
@@ -0,0 +1,117 @@
+#!/bin/bash
+test_suite=${1:-}
+test_number=${2:-}
+
+PROG=${0##*/}
+build_dir="build-ci-debug"
+
+if [ x"$1" = x"-h" ] || [ x"$1" = x"--help" ]; then
+    echo "Usage: $PROG [OPTION]... <test_regex> (test_number)"
+    echo "Debug specific ctest program."
+    echo
+    echo "Options:"
+    echo "  -h, --help       Display this help and exit"
+    echo
+    echo "Arguments:"
+    echo "  <test_regex>     (Mandatory) Supply one regex to the script to filter tests"
+    echo "  (test_number)    (Optional) Test number to run a specific test"
+    echo
+    echo "Example:"
+    echo "  $PROG test-tokenizer"
+    echo "  $PROG test-tokenizer 3"
+    echo
+    exit 0
+fi
+
+# Function to select and debug a test
+function select_test() {
+    test_suite=${1:-test}
+    test_number=${2:-}
+
+    # Sanity Check If Tests Is Detected
+    printf "\n\nGathering tests that fit REGEX: ${test_suite} ...\n"
+    tests=($(ctest -R ${test_suite} -V -N | grep -E " +Test +#[0-9]+*" | cut -d':' -f2 | awk '{$1=$1};1'))
+    if [ ${#tests[@]} -eq 0 ]
+    then
+        echo "No tests avaliable... check your compliation process..."
+        echo "Exiting."
+        exit 1
+    fi
+
+    if [ -z $test_number ]
+    then
+        # List out avaliable tests
+        printf "Which test would you like to debug?\n"
+        id=0
+        for s in "${tests[@]}"
+        do
+            echo "Test# ${id}"
+            echo "  $s"
+            ((id++))
+        done
+
+        # Prompt user which test they wanted to run
+        printf "\nRun test#? "
+        read test_number
+    else
+        printf "\nUser Already Requested #${test_number}"
+    fi
+
+    # Start GDB with the requested test binary and arguments
+    printf "Debugging(GDB) test: ${tests[test_number]}\n"
+    # Change IFS (Internal Field Separator)
+    sIFS=$IFS
+    IFS=$'\n'
+
+    # Get test args
+    gdb_args=($(ctest -R ${test_suite} -V -N | grep "Test command" | cut -d':' -f3 | awk '{$1=$1};1' ))
+    IFS=$sIFS
+    printf "Debug arguments: ${gdb_args[test_number]}\n\n"
+
+    # Expand paths if needed
+    args=()
+    for x in $(echo ${gdb_args[test_number]} | sed -e 's/"\/\<//' -e 's/\>"//')
+    do
+        args+=($(echo $x | sed -e 's/.*\/..\//..\//'))
+    done
+
+    # Execute debugger
+    echo "gdb args: ${args[@]}"
+    gdb --args ${args[@]}
+}
+
+# Step 0: Check the args
+if [ -z "$test_suite" ]
+then
+    echo "Usage: $PROG [OPTION]... <test_regex> (test_number)"
+    echo "Supply one regex to the script to filter tests,"
+    echo "and optionally a test number to run a specific test."
+    echo "Use --help flag for full instructions"
+    exit 1
+fi
+
+# Step 1: Reset and Setup folder context
+## Sanity check that we are actually in a git repo
+repo_root=$(git rev-parse --show-toplevel)
+if [ ! -d "$repo_root" ]; then
+    echo "Error: Not in a Git repository."
+    exit 1
+fi
+
+## Reset folder to root context of git repo
+pushd "$repo_root" || exit 1
+
+## Create and enter build directory
+rm -rf "$build_dir" && mkdir "$build_dir" || exit 1
+
+# Step 2: Setup Build Environment and Compile Test Binaries
+cmake -B "./$build_dir" -DCMAKE_BUILD_TYPE=Debug -DLLAMA_CUDA=1 -DLLAMA_FATAL_WARNINGS=ON || exit 1
+pushd "$build_dir" && make -j || exit 1
+
+# Step 3: Debug the Test
+select_test "$test_suite" "$test_number"
+
+# Step 4: Return to the directory from which the user ran the command.
+popd || exit 1
+popd || exit 1
+popd || exit 1
@@ -1 +1 @@
-98875cdb7e9ceeb726d1c196d2fecb3cbb59b93a
+30f54cbb3ada3e4c5bc6924de3e5918e5be4ff11
Author	SHA1	Message	Date
slaren	b228aba91a	remove convert-lora-to-ggml.py (#7204 )	2024-05-12 02:29:33 +02:00
Georgi Gerganov	7bd4ffb780	metal : fix warnings (skipme) (#0 )	2024-05-11 21:38:13 +03:00
Georgi Gerganov	1622ac023f	sync : ggml	2024-05-11 21:35:05 +03:00
Georgi Gerganov	6aeff24f8b	metal : fix indent (ggml/0)	2024-05-11 21:34:21 +03:00
Georgi Gerganov	325756d28d	ggml : resolve merge (ggml/0) ggml-ci	2024-05-11 21:33:08 +03:00
Josh Ramer	fed0108491	Scripting & documenting debugging one test without anything else in the loop. (#7096 ) * A little documentation that shares my quick tips for working in the repository. * Update startup-testing-debugging.md * script that shows a menu of tests to pick from & run the debugger on * debug-test.sh: Refactor CLI help message * debug-test.sh: documentation update * debug-test.sh: CLI Help output corrections * debug-test.sh: minor doc fix --------- authored-by: Josh Ramer <ubuntu@ip-172-31-32-53.ec2.internal> Assisted-by: brian khuu <mofosyne@gmail.com>	2024-05-12 03:26:35 +10:00
Xuan Son Nguyen	72c177c1f6	fix system prompt handling (#7153 )	2024-05-11 17:28:10 +02:00
compilade	5a419926b0	convert-hf : support bfloat16 conversion (#7158 ) * convert-hf : support bfloat16 conversion * gguf-py : flake8 fixes * convert-hf : add missing space after comma * convert-hf : get bit-exact same output as ./quantize The quantization version was missing. * convert-hf : don't round bf16 NANs * convert-hf : save some memory with np.int16 intermediate bf16 weights * convert-hf : more closely match llama.cpp with which weights to keep in f32 * convert-hf : add --outtype auto-f16 A reason for this to exist is for model quantizers who want an initial GGUF with the most fidelity to the original model while still using a 16-bit float type instead of 32-bit floats. * convert-hf : remove a semicolon because flake8 doesn't like it It's a reflex from when programming in C/C++, I guess. * convert-hf : support outtype templating in outfile name * convert-hf : rename --outtype auto-f16 to --outtype auto	2024-05-11 11:06:26 -04:00
Georgi Gerganov	fae9d234b6	sync : ggml ggml-ci	2024-05-11 15:38:34 +03:00
Justina Cho	f5ef34e428	feat: implemented sigmoid function (ggml/806) * added sigmoid function * implemented metal kernel for sigmoid * implemented cuda kernel for sigmoid * added sigmoid unary op and incremented count	2024-05-11 15:38:34 +03:00
Borislav Stanimirov	ef0d5e3ec9	build: fix and ignore msvc warnings (ggml/805)	2024-05-11 15:38:34 +03:00
CrispStrobe	3292733f95	convert : skip unaccessible HF repos (#7210 )	2024-05-11 11:18:35 +03:00