transformers 5.0.0rc0__py3-none-any.whl → 5.0.0rc2__py3-none-any.whl

transformers/integrations/fp_quant.py

@@ -13,6 +13,10 @@
 # limitations under the License.
 "FP-Quant integration file"
 
+from typing import Optional
+
+import torch
+
 from ..utils import (
     is_fp_quant_available,
 )
@@ -24,6 +28,94 @@ if is_fp_quant_available():
 
 from transformers.utils.quantization_config import FPQuantConfig
 
+from ..core_model_loading import ConversionOps
+from ..quantizers.quantizers_utils import get_module_from_name
+
+
+class FpQuantQuantize(ConversionOps):
+    def __init__(self, hf_quantizer):
+        self.hf_quantizer = hf_quantizer
+
+    def convert(
+        self,
+        input_dict: torch.Tensor,
+        model: Optional[torch.nn.Module] = None,
+        missing_keys: Optional[list[str]] = None,
+        **kwargs,
+    ) -> dict[str, torch.Tensor]:
+        target_key, value = tuple(input_dict.items())[0]
+        value = value[0]
+        # Loading master weights or an unquantized checkpoint
+        weight = torch.nn.Parameter(value)
+        module, _ = get_module_from_name(model, target_key)
+        module.weight = weight
+
+        # Let pre-forward handle the quantization and set None where necessary
+        # This operation will quantize the weights internally
+        with torch.cuda.device(value.device):
+            module.pre_forward()
+
+        prefix_target_key = target_key.rsplit(".", 1)[0]
+
+        # keys are set inside the module.pre_forward() method, we don't need remove them from the missing keys list
+        missing_keys.discard(target_key)
+        missing_keys.discard(f"{prefix_target_key}.backward_hadamard_matrix")
+        missing_keys.discard(f"{prefix_target_key}.forward_hadamard_matrix")
+        missing_keys.discard(f"{prefix_target_key}.act_global_scale")
+        missing_keys.discard(f"{prefix_target_key}.weight_global_scale")
+        missing_keys.discard(f"{prefix_target_key}.qweight")
+        missing_keys.discard(f"{prefix_target_key}.scales")
+        missing_keys.discard(f"{prefix_target_key}.dqweight")
+        return {}
+
+
+class FpQuantDeserialize(ConversionOps):
+    def __init__(self, hf_quantizer):
+        self.hf_quantizer = hf_quantizer
+
+    def convert(
+        self,
+        input_dict: torch.Tensor,
+        model: Optional[torch.nn.Module] = None,
+        full_layer_name: str | None = None,
+        missing_keys: Optional[list[str]] = None,
+        **kwargs,
+    ) -> dict[str, torch.Tensor]:
+        target_key, value = tuple(input_dict.items())[0]
+        value = value[0] if isinstance(value, list) else value
+        module, _ = get_module_from_name(model, target_key)
+        # The module holds either:
+        #  * `weight` when `store_master_weights=True`
+        #  * `qweight` and `scales` when `store_master_weights=False` and `pseudoquantization=False`
+        #  * `dqweight` when `store_master_weights=False` and `pseudoquantization=True`
+        if target_key == ".qweight":
+            # Loading a real quantized checkpoint without master weights
+            qweight = torch.nn.Parameter(
+                value,
+                requires_grad=False,
+            )
+
+            return {
+                ".qweight": qweight,
+                # the way the FPQuantLinear module is designed, these parameters are expected in the model
+                # even though they are not used so we need to set them to zeros
+                ".weight": torch.nn.Parameter(torch.zeros(0)),
+                ".dqweight": torch.nn.Parameter(torch.zeros(0)),
+            }
+
+        if target_key == ".dqweight":
+            # Loading a pseudo-quantized checkpoint without master weights
+            dqweight = torch.nn.Parameter(value)
+
+            return {
+                ".dqweight": dqweight,
+                # the way the FPQuantLinear module ips designed, these parameters are expected in the model
+                # even though they are not used so we need to set them to zeros
+                ".weight": torch.nn.Parameter(torch.zeros(0)),
+                ".qweight": torch.nn.Parameter(torch.zeros(0)),
+                ".scales": torch.nn.Parameter(torch.zeros(0)),
+            }
+
 
 def adapt_fp_quant_config(config: FPQuantConfig):
     if config.forward_dtype == "mxfp4":

transformers/integrations/ggml.py

@@ -76,7 +76,7 @@ GGUF_CONFIG_MAPPING = {
         "attention.layer_norm_rms_epsilon": "rms_norm_eps",
         "vocab_size": "vocab_size",
     },
-    "qwen2moe": {
+    "qwen2_moe": {
         "context_length": "max_position_embeddings",
         "block_count": "num_hidden_layers",
         "feed_forward_length": "intermediate_size",
@@ -313,6 +313,16 @@ GGUF_TOKENIZER_MAPPING = {
     },
 }
 
+# We only need to set here the parameters that default to different values between transformers and llamacpp.
+GGUF_CONFIG_DEFAULTS_MAPPING = {
+    "qwen3_moe": {
+        # NOTE: Qwen3MoeConfig defaults to false but llama.cpp needs this to be true.
+        # See: https://github.com/ggml-org/llama.cpp/blob/17f7f4baad8b3a716ee139da7bb56ae984e8c0fa/src/models/qwen3moe.cpp#L85-L96
+        #      (the parameter right after LLM_FFN_SILU corresponds to norm_topk_prob)
+        "norm_topk_prob": True,
+    },
+}
+
 
 def _gguf_parse_value(_value, data_type):
     if not isinstance(data_type, list):

transformers/integrations/higgs.py

@@ -15,17 +15,13 @@
 
 from math import sqrt
 
-from ..utils import (
-    is_flute_available,
-    is_hadamard_available,
-    is_torch_available,
-)
+from ..quantizers.quantizers_utils import should_convert_module
+from ..utils import is_flute_available, is_hadamard_available, is_torch_available, logging
 
 
 if is_torch_available():
     import torch
-    from torch import nn
-
+    import torch.nn as nn
 
 if is_flute_available():
     from flute.integrations.higgs import prepare_data_transposed
@@ -34,6 +30,8 @@ if is_flute_available():
 if is_hadamard_available():
     from fast_hadamard_transform import hadamard_transform
 
+logger = logging.get_logger(__name__)
+
 
 def pad_to_block(tensor, dims, had_block_size, value=0):
     pad_dims = [0 for _ in range(2 * len(tensor.shape))]
@@ -549,70 +547,47 @@ class HiggsLinear(torch.nn.Module):
         )
 
 
-def replace_with_higgs_linear(
-    model,
-    quantization_config=None,
-    current_key_name=None,
-    has_been_replaced=False,
-    modules_to_not_convert=None,
-):
+def replace_with_higgs_linear(model, modules_to_not_convert: list[str] | None = None, quantization_config=None):
     """
-    Public method that recursively replaces the Linear layers of the given model with HIGGS quantized layers.
-    `accelerate` is needed to use this method. Returns the converted model and a boolean that indicates if the
-    conversion has been successful or not.
+    Public method that replaces the Linear layers of the given model with HIGGS quantized layers.
 
     Args:
         model (`torch.nn.Module`):
             The model to convert, can be any `torch.nn.Module` instance.
+        modules_to_not_convert (`list[str]`, *optional*, defaults to `None`):
+            A list of nn.Linear weights to not convert. If a parameter path is in the list (e.g. `lm_head.weight`), the corresponding module will not be
+            converted.
         quantization_config (`HiggsConfig`):
             The quantization config object that contains the quantization parameters.
-        current_key_name (`list`, *optional*):
-            A list that contains the current key name. This is used for recursion and should not be passed by the user.
-        has_been_replaced (`bool`, *optional*):
-            A boolean that indicates if the conversion has been successful or not. This is used for recursion and
-            should not be passed by the user.
     """
 
-    from accelerate import init_empty_weights
-
-    for name, module in model.named_children():
-        if current_key_name is None:
-            current_key_name = []
-        current_key_name.append(name)
-
-        if isinstance(module, nn.Linear):
-            # Check if the current key is not in the `quantization_config.modules_to_not_convert`
-            current_key_name_str = ".".join(current_key_name)
-            if not any(current_key_name_str.endswith(key) for key in modules_to_not_convert):
-                with init_empty_weights():
-                    in_features = module.in_features
-                    out_features = module.out_features
-
-                    model._modules[name] = HiggsLinear(
-                        in_features,
-                        out_features,
-                        bias=module.bias is not None,
-                        num_bits=quantization_config.bits,
-                        hadamard_size=quantization_config.hadamard_size,
-                        group_size=quantization_config.group_size,
-                    )
-                    has_been_replaced = True
-
-                    # Store the module class in case we need to transpose the weight later
-                    model._modules[name].source_cls = type(module)
-                    # Force requires grad to False to avoid unexpected errors
-                    model._modules[name].requires_grad_(False)
-        if len(list(module.children())) > 0:
-            _, has_been_replaced = replace_with_higgs_linear(
-                module,
-                quantization_config=quantization_config,
-                current_key_name=current_key_name,
-                has_been_replaced=has_been_replaced,
-                modules_to_not_convert=modules_to_not_convert,
-            )
-        # Remove the last key for recursion
-        current_key_name.pop(-1)
-    return model, has_been_replaced
+    has_been_replaced = False
+    # we need this to correctly materialize the weights during quantization
+    for module_name, module in model.named_modules():
+        if not should_convert_module(module_name, modules_to_not_convert):
+            continue
+        with torch.device("meta"):
+            if isinstance(module, nn.Linear):
+                new_module = HiggsLinear(
+                    module.in_features,
+                    module.out_features,
+                    bias=module.bias is not None,
+                    num_bits=quantization_config.bits,
+                    hadamard_size=quantization_config.hadamard_size,
+                    group_size=quantization_config.group_size,
+                )
+                new_module.source_cls = type(module)
+                new_module.requires_grad_(False)
+                model.set_submodule(module_name, new_module)
+                has_been_replaced = True
+
+    if not has_been_replaced:
+        logger.warning(
+            "You are loading your model using eetq but no linear modules were found in your model."
+            " Please double check your model architecture, or submit an issue on github if you think this is"
+            " a bug."
+        )
+    return model
 
 
 def dequantize_higgs(model, current_key_name=None):

transformers/integrations/hub_kernels.py

@@ -11,11 +11,14 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import importlib.metadata
 import os
 import re
 from collections.abc import Callable
 from types import ModuleType
 
+from packaging import version as pkg_version
+
 from ..utils import ENV_VARS_TRUE_VALUES, logging
 from ..utils.import_utils import is_kernels_available
 from .flash_attention import flash_attention_forward
@@ -28,10 +31,12 @@ try:
         Device,
         LayerRepository,
         Mode,
-        get_kernel,
         register_kernel_mapping,
         replace_kernel_forward_from_hub,
     )
+    from kernels import (
+        get_kernel as get_kernel_hub,
+    )
     from kernels import (
         use_kernel_forward_from_hub as _kernels_use_kernel_forward_from_hub,
     )
@@ -78,7 +83,7 @@ try:
                 )
             return lambda func: func
 
-    _KERNEL_MAPPING: dict[str, dict[Device | str, LayerRepository]] = {
+    _KERNEL_MAPPING: dict[str, dict[Device | str, LayerRepository | dict[Mode, LayerRepository]]] = {
         "MultiScaleDeformableAttention": {
             "cuda": LayerRepository(
                 repo_id="kernels-community/deformable-detr",
@@ -111,6 +116,12 @@ try:
                     layer_name="RMSNorm",
                 )
             },
+            "mps": {
+                Mode.INFERENCE: LayerRepository(
+                    repo_id="kernels-community/mlx_rmsnorm",
+                    layer_name="RMSNorm",
+                )
+            },
             "npu": {
                 Mode.INFERENCE: LayerRepository(
                     repo_id="kernels-community/liger_kernels",
@@ -253,6 +264,8 @@ except ImportError:
 
 _HUB_KERNEL_MAPPING: dict[str, dict[str, str]] = {
     "causal-conv1d": {"repo_id": "kernels-community/causal-conv1d"},
+    "mamba-ssm": {"repo_id": "kernels-community/mamba-ssm", "revision": "v0.0.4"},
+    "falcon_mamba-ssm": {"repo_id": "kernels-community/mamba-ssm", "revision": "v0.0.4"},
 }
 
 _KERNEL_MODULE_MAPPING: dict[str, ModuleType | None] = {}
@@ -328,19 +341,21 @@ def lazy_load_kernel(kernel_name: str, mapping: dict[str, ModuleType | None] = _
     if kernel_name in mapping and isinstance(mapping[kernel_name], ModuleType):
         return mapping[kernel_name]
     if kernel_name not in _HUB_KERNEL_MAPPING:
-        logger.warning(f"Kernel {kernel_name} not found in _HUB_KERNEL_MAPPING")
+        logger.warning_once(f"Kernel {kernel_name} not found in _HUB_KERNEL_MAPPING")
         mapping[kernel_name] = None
         return None
     if _kernels_available:
-        from kernels import get_kernel
-
         try:
             repo_id = _HUB_KERNEL_MAPPING[kernel_name]["repo_id"]
+            revision = _HUB_KERNEL_MAPPING[kernel_name].get("revision", None)
             version = _HUB_KERNEL_MAPPING[kernel_name].get("version", None)
-            kernel = get_kernel(repo_id, version=version)
+            kernel = get_kernel(repo_id, revision=revision, version=version)
             mapping[kernel_name] = kernel
         except FileNotFoundError:
             mapping[kernel_name] = None
+        except AssertionError:
+            # Happens when torch is built without an accelerator backend; fall back to slow path.
+            mapping[kernel_name] = None
 
     else:
         # Try to import is_{kernel_name}_available from ..utils
@@ -358,7 +373,7 @@ def lazy_load_kernel(kernel_name: str, mapping: dict[str, ModuleType | None] = _
         if callable(is_kernel_available) and is_kernel_available():
             # Try to import the module "{kernel_name}" from parent package level
             try:
-                module = importlib.import_module(f"{kernel_name}")
+                module = importlib.import_module(f"{new_kernel_name}")
                 mapping[kernel_name] = module
                 return module
             except Exception:
@@ -369,6 +384,46 @@ def lazy_load_kernel(kernel_name: str, mapping: dict[str, ModuleType | None] = _
     return mapping[kernel_name]
 
 
+def get_kernel(kernel_name: str, revision: str | None = None, version: str | None = None) -> ModuleType:
+    from .. import __version__
+
+    user_agent = {"framework": "transformers", "version": __version__, "repo_id": kernel_name}
+    if _kernels_available:
+        kernels_version = importlib.metadata.version("kernels")
+        if pkg_version.parse(kernels_version) >= pkg_version.parse("0.10.4"):
+            return get_kernel_hub(kernel_name, revision=revision, version=version, user_agent=user_agent)
+        else:
+            return get_kernel_hub(kernel_name, revision=revision)
+    else:
+        raise ImportError("kernels is not installed, please install it with `pip install kernels`")
+
+
+def use_kernelized_func(module_names: list[Callable] | Callable):
+    """
+    This decorator attaches the target function as an attribute of the module.
+    The function must already be decorated with @use_kernel_func_from_hub
+    this decorator then wraps it as an nn.Module internally.
+    When kernelize is later applied to the full model, the function can be accessed as a regular module attribute and kernelized just like any other layer.
+    The kernelization is performed in place, modifying the module directly.
+    """
+    if isinstance(module_names, Callable):
+        module_names = [module_names]
+
+    def decorator(cls):
+        orig_init = cls.__init__
+
+        def new_init(self, *args, **kwargs):
+            orig_init(self, *args, **kwargs)
+            for fn in module_names:
+                # we hardcode the name of the function to "rotary_fn" for now
+                setattr(self, "rotary_fn", fn)
+
+        cls.__init__ = new_init
+        return cls
+
+    return decorator
+
+
 __all__ = [
     "LayerRepository",
     "use_kernel_forward_from_hub",
@@ -377,4 +432,6 @@ __all__ = [
     "register_kernel_mapping_transformers",
     "replace_kernel_forward_from_hub",
     "lazy_load_kernel",
-]
+    "get_kernel",
+    "use_kernelized_func",
+]  # type: ignore

transformers/integrations/integration_utils.py

@@ -26,6 +26,7 @@ import re
 import shutil
 import sys
 import tempfile
+import warnings
 from dataclasses import fields
 from enum import Enum
 from pathlib import Path
@@ -939,6 +940,8 @@ class TrackioCallback(TrainerCallback):
     ```
     """
 
+    SPACE_URL = "https://huggingface.co/spaces/{space_id}"
+
     def __init__(self):
         has_trackio = is_trackio_available()
         if not has_trackio:
@@ -1057,6 +1060,39 @@ class TrackioCallback(TrainerCallback):
             metrics = rewrite_logs(metrics)
             self._trackio.log(metrics)
 
+    def on_push_begin(self, args, state, control, model, **kwargs):
+        if not state.is_world_process_zero or self._trackio is None:
+            return
+        if (current_project := self._trackio.context_vars.current_project.get()) is None:
+            return
+        trackio_version = packaging.version.parse(self._trackio.__version__)
+        if trackio_version < packaging.version.parse("0.13.0"):
+            warnings.warn(
+                "The version of `trackio` that is installed is <=0.13.0, so "
+                "the local Trackio project will not be pushed to Hugging Face. Run "
+                "`pip install --upgrade trackio` to fix this."
+            )
+            return
+
+        space_id = self._trackio.context_vars.current_space_id.get()
+        if space_id is None:
+            space_id = self._trackio.sync(current_project, force=True)
+        space_url = self.SPACE_URL.format(space_id=space_id)
+
+        badge_markdown = (
+            f'<a href="{space_url}" target="_blank"><img src="https://raw.githubusercontent.com/gradio-app/trackio/refs/heads/main/trackio/assets/badge.png" alt="Visualize in Trackio"'
+            ' title="Visualize in Trackio" style="height: 40px;"/></a>'
+        )
+        if badge_markdown not in modelcard.AUTOGENERATED_TRAINER_COMMENT:
+            modelcard.AUTOGENERATED_TRAINER_COMMENT += f"\n{badge_markdown}"
+
+        trackio_tags = ["trackio", f"trackio:{space_url}"]
+        if getattr(model, "model_tags", None) is not None:
+            if "trackio" not in model.model_tags:
+                model.model_tags.extend(trackio_tags)
+        else:
+            model.model_tags = trackio_tags
+
 
 class CometCallback(TrainerCallback):
     """
@@ -1455,6 +1491,10 @@ class NeptuneMissingConfiguration(Exception):
 class NeptuneCallback(TrainerCallback):
     """TrainerCallback that sends the logs to [Neptune](https://app.neptune.ai).
 
+    > [!WARNING]
+    > Neptune integration is deprecated and will be removed in a future version of Transformers. We recommend using
+    > other supported experiment tracking integrations.
+
     Args:
         api_token (`str`, *optional*): Neptune API token obtained upon registration.
             You can leave this argument out if you have saved your token to the `NEPTUNE_API_TOKEN` environment
@@ -1500,6 +1540,11 @@ class NeptuneCallback(TrainerCallback):
         log_checkpoints: str | None = None,
         **neptune_run_kwargs,
     ):
+        warnings.warn(
+            "The NeptuneCallback is deprecated and will be removed in a future version of Transformers. We recommend "
+            "using other supported experiment tracking integrations.",
+            FutureWarning,
+        )
         if not is_neptune_available():
             raise ValueError(
                 "NeptuneCallback requires the Neptune client library to be installed. "

transformers/integrations/mistral.py

@@ -77,6 +77,7 @@ def convert_tekken_tokenizer(tokenizer_file: str):
     """Convert a "tekken" tokenizer to a fast Tokenizer."""
     # Tekken format -- need to use the Converter
 
+    from mistral_common.tokens.tokenizers.base import SpecialTokens
     from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
 
     # Load directly using their lib
@@ -106,4 +107,15 @@ def convert_tekken_tokenizer(tokenizer_file: str):
     # Post-process
     tokenizer.add_special_tokens({"additional_special_tokens": all_special})
 
+    MAP_SPECAL = {
+        "bos_token": SpecialTokens.bos.value,
+        "eos_token": SpecialTokens.eos.value,
+        "pad_token": SpecialTokens.pad.value,
+        "unk_token": SpecialTokens.unk.value,
+    }
+
+    for special_key, special_token in MAP_SPECAL.items():
+        if special_token in all_special:
+            tokenizer.add_special_tokens({special_key: special_token})
+
     return tokenizer