Upload Ouro-2.6B_smoothquant_W8A8 with bundled source code

qouro/__init__.py

@@ -0,0 +1 @@
+__all__ = []

qouro/configuration_ouro.py

@@ -0,0 +1,222 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+"""Ouro model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig, layer_type_validation
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class OuroConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`OuroModel`]. It is used to instantiate a
+    Ouro model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Ouro-7B-beta [Qwen/Ouro-7B-beta](https://huggingface.co/Qwen/Ouro-7B-beta).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 151936):
+            Vocabulary size of the Ouro model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`OuroModel`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 22016):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 32):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details, check out [this
+            paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `32`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`list[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`list[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 28):
+            The number of layers using full attention. The first `max_window_layers` layers will use full attention, while any
+            additional layer afterwards will use SWA (Sliding Window Attention).
+        layer_types (`list`, *optional*):
+            Attention pattern for each layer.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import OuroModel, OuroConfig
+
+    >>> # Initializing a Ouro style configuration
+    >>> configuration = OuroConfig()
+
+    >>> # Initializing a model from the Ouro-7B style configuration
+    >>> model = OuroModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "ouro"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    # Default tensor parallel plan for base model `Ouro`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        intermediate_size=22016,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        layer_types=None,
+        attention_dropout=0.0,
+        total_ut_steps=4,
+        early_exit_threshold=1.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window if self.use_sliding_window else None
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_dropout = attention_dropout
+        self.total_ut_steps = total_ut_steps
+        self.early_exit_threshold = early_exit_threshold
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+
+        self.layer_types = layer_types
+        if self.layer_types is None:
+            self.layer_types = [
+                "sliding_attention"
+                if self.sliding_window is not None and i >= self.max_window_layers
+                else "full_attention"
+                for i in range(self.num_hidden_layers)
+            ]
+        layer_type_validation(self.layer_types)
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["OuroConfig"]

qouro/modeling_ouro.py

@@ -0,0 +1,701 @@
+from typing import Callable, Optional, Union
+
+import torch
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.generation import GenerationMixin
+from transformers.integrations import use_kernel_forward_from_hub
+from transformers.masking_utils import (
+    create_causal_mask,
+    create_sliding_window_causal_mask,
+)
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_layers import (
+    GenericForQuestionAnswering,
+    GenericForSequenceClassification,
+    GenericForTokenClassification,
+    GradientCheckpointingLayer,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple
+from transformers.utils.generic import check_model_inputs
+from .configuration_ouro import OuroConfig
+
+
+class OuroMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(
+        batch, num_key_value_heads, n_rep, slen, head_dim
+    )
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(
+        query.dtype
+    )
+    attn_weights = nn.functional.dropout(
+        attn_weights, p=dropout, training=module.training
+    )
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+class OuroAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: OuroConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(
+            config, "head_dim", config.hidden_size // config.num_attention_heads
+        )
+        self.num_key_value_groups = (
+            config.num_attention_heads // config.num_key_value_heads
+        )
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+        self.q_proj = nn.Linear(
+            config.hidden_size, config.num_attention_heads * self.head_dim, bias=False
+        )
+        self.k_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(
+            config.num_attention_heads * self.head_dim, config.hidden_size, bias=False
+        )
+        self.sliding_window = (
+            config.sliding_window
+            if config.layer_types[layer_idx] == "sliding_attention"
+            else None
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        current_ut: int = 0,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(
+            query_states, key_states, cos, sin
+        )
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(
+                key_states,
+                value_states,
+                current_ut * self.config.num_hidden_layers + self.layer_idx,
+                cache_kwargs,
+            )
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[
+                self.config._attn_implementation
+            ]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            sliding_window=self.sliding_window,  # main diff with Llama
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class OuroRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        OuroRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class OuroDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: OuroConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = OuroAttention(config=config, layer_idx=layer_idx)
+
+        self.mlp = OuroMLP(config)
+        self.input_layernorm = OuroRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.input_layernorm_2 = OuroRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.post_attention_layernorm = OuroRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.post_attention_layernorm_2 = OuroRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.attention_type = config.layer_types[layer_idx]
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[
+            tuple[torch.Tensor, torch.Tensor]
+        ] = None,  # necessary, but kept here for BC
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor]:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        # Self Attention
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = self.input_layernorm_2(hidden_states)
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = self.post_attention_layernorm_2(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+@auto_docstring
+class OuroPreTrainedModel(PreTrainedModel):
+    config: OuroConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["OuroDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+
+    _can_compile_fullgraph = True
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": OuroDecoderLayer,
+        "attentions": OuroAttention,
+    }
+
+
+class OuroRotaryEmbedding(nn.Module):
+    def __init__(self, config: OuroConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
+            self.rope_type = config.rope_scaling.get(
+                "rope_type", config.rope_scaling.get("type")
+            )
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = (
+            self.inv_freq[None, :, None]
+            .float()
+            .expand(position_ids.shape[0], -1, 1)
+            .to(x.device)
+        )
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = (
+            x.device.type
+            if isinstance(x.device.type, str) and x.device.type != "mps"
+            else "cpu"
+        )
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (
+                inv_freq_expanded.float() @ position_ids_expanded.float()
+            ).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+@auto_docstring
+class OuroModel(OuroPreTrainedModel):
+    def __init__(self, config: OuroConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(
+            config.vocab_size, config.hidden_size, self.padding_idx
+        )
+        self.layers = nn.ModuleList(
+            [
+                OuroDecoderLayer(config, layer_idx)
+                for layer_idx in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = OuroRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = OuroRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+        self.has_sliding_layers = "sliding_attention" in self.config.layer_types
+        self.total_ut_steps = getattr(self.config, "total_ut_steps", 4)
+        self.early_exit_gate = nn.Linear(config.hidden_size, 1)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @check_model_inputs
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> BaseModelOutputWithPast:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You must specify exactly one of input_ids or inputs_embeds"
+            )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = (
+                past_key_values.get_seq_length() if past_key_values is not None else 0
+            )
+            cache_position = torch.arange(
+                past_seen_tokens,
+                past_seen_tokens + inputs_embeds.shape[1],
+                device=inputs_embeds.device,
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        # It may already have been prepared by e.g. `generate`
+        if not isinstance(causal_mask_mapping := attention_mask, dict):
+            # Prepare mask arguments
+            mask_kwargs = {
+                "config": self.config,
+                "input_embeds": inputs_embeds,
+                "attention_mask": attention_mask,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "position_ids": position_ids,
+            }
+            # Create the masks
+            causal_mask_mapping = {
+                "full_attention": create_causal_mask(**mask_kwargs),
+            }
+            # The sliding window alternating layers are not always activated depending on the config
+            if self.has_sliding_layers:
+                causal_mask_mapping["sliding_attention"] = (
+                    create_sliding_window_causal_mask(**mask_kwargs)
+                )
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+        hidden_states_list = []
+        gate_list = []
+
+        for current_ut in range(self.total_ut_steps):
+            for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+                hidden_states = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask_mapping[decoder_layer.attention_type],
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                    current_ut=current_ut,
+                    **kwargs,
+                )
+
+            hidden_states = self.norm(hidden_states)
+            hidden_states_list.append(hidden_states)
+            gate_list.append(self.early_exit_gate(hidden_states))
+
+        return (
+            BaseModelOutputWithPast(
+                last_hidden_state=hidden_states,
+                past_key_values=past_key_values if use_cache else None,
+            ),
+            hidden_states_list,
+            gate_list,
+        )
+
+
+@auto_docstring
+class OuroForCausalLM(OuroPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = OuroModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # 分块大小配置
+        self.chunk_size = getattr(config, "chunk_size", 2)  # 默认分块大小为2
+        self.early_exit_step = getattr(config, "early_exit_step", None)
+        self.early_exit_threshold = getattr(config, "early_exit_threshold", None)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        use_weighted_exit: Optional[bool] = False,  # 控制是否使用加权 early exit
+        exit_at_step: Optional[int] = None,
+        exit_threshold: Optional[float] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> CausalLMOutputWithPast:
+        r"""
+        Args:
+            use_weighted_exit (`bool`, *optional*, defaults to `False`):
+                Whether to use weighted early exit. If `True`, the logits from all UT steps will be
+                averaged according to the exit probability distribution.
+            exit_at_step (`int`, *optional*):
+                Specifies which UT step to exit at. If set, the model will directly use the hidden states
+                from this step to generate logits, ignoring other exit strategies.
+            exit_threshold (`float`, *optional*):
+                The cumulative probability threshold for early exit. When the cumulative exit probability
+                reaches this threshold, the model will exit at that step.
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, OuroForCausalLM
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        exit_at_step = (
+            exit_at_step if exit_at_step is not None else self.early_exit_step
+        )
+        exit_threshold = (
+            exit_threshold if exit_threshold is not None else self.early_exit_threshold
+        )
+
+        outputs, hidden_states_list, gate_list = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            **kwargs,
+        )
+        slice_indices = (
+            slice(-logits_to_keep, None)
+            if isinstance(logits_to_keep, int)
+            else logits_to_keep
+        )
+
+        def _select_token_positions(tensor: torch.Tensor) -> torch.Tensor:
+            if isinstance(slice_indices, slice):
+                return tensor[:, slice_indices, ...]
+            if isinstance(slice_indices, torch.Tensor):
+                return tensor.index_select(1, slice_indices.to(tensor.device))
+            raise TypeError(
+                f"Unsupported index type for logits_to_keep: {type(slice_indices)}"
+            )
+
+        stacked_exit_pdf = None
+        if gate_list:
+            pdf_list = []
+            remaining_prob = torch.ones_like(gate_list[0].squeeze(-1))
+            for idx, gate_tensor in enumerate(gate_list):
+                lambda_i = torch.sigmoid(gate_tensor.squeeze(-1))
+                if idx < len(gate_list) - 1:
+                    p_i = lambda_i * remaining_prob
+                    remaining_prob = remaining_prob * (1.0 - lambda_i)
+                else:
+                    p_i = remaining_prob
+                pdf_list.append(p_i)
+            stacked_exit_pdf = torch.stack(pdf_list, dim=2)
+
+        expected_logits_cache: Optional[torch.Tensor] = None
+
+        def compute_expected_logits() -> Optional[torch.Tensor]:
+            nonlocal expected_logits_cache
+            if expected_logits_cache is not None:
+                return expected_logits_cache
+            if stacked_exit_pdf is None or not hidden_states_list:
+                return None
+            token_exit_pdf = _select_token_positions(stacked_exit_pdf)
+            expected_logits = None
+            for step_idx, hidden in enumerate(hidden_states_list):
+                step_hidden = _select_token_positions(hidden)
+                step_logits = self.lm_head(step_hidden)
+                weight = (
+                    token_exit_pdf[..., step_idx].unsqueeze(-1).to(step_logits.dtype)
+                )
+                expected_logits = (
+                    step_logits * weight
+                    if expected_logits is None
+                    else expected_logits + step_logits * weight
+                )
+            expected_logits_cache = expected_logits
+            return expected_logits_cache
+
+        logits: Optional[torch.Tensor] = None
+        loss: Optional[torch.Tensor] = None
+
+        if labels is not None:
+            logits = compute_expected_logits()
+            if logits is None:
+                hidden_states = outputs.last_hidden_state
+                logits = self.lm_head(_select_token_positions(hidden_states))
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+        else:
+            if stacked_exit_pdf is not None and hidden_states_list:
+                if exit_at_step is not None and 0 <= exit_at_step < len(
+                    hidden_states_list
+                ):
+                    selected_hidden = hidden_states_list[exit_at_step]
+                    logits = self.lm_head(_select_token_positions(selected_hidden))
+                elif exit_threshold is not None:
+                    cumulative_probs = torch.cumsum(stacked_exit_pdf, dim=2)
+                    threshold_value = exit_threshold
+                    if isinstance(threshold_value, torch.Tensor):
+                        threshold_value = threshold_value.to(cumulative_probs.device)
+                    threshold_mask = cumulative_probs >= threshold_value
+                    exit_steps = torch.argmax(threshold_mask.float(), dim=2)
+                    last_step_idx = stacked_exit_pdf.shape[2] - 1
+                    if last_step_idx >= 0:
+                        never_exceeded = ~threshold_mask.any(dim=2)
+                        exit_steps[never_exceeded] = last_step_idx
+                    stacked_hidden = torch.stack(hidden_states_list, dim=2)
+                    gather_index = (
+                        exit_steps.unsqueeze(-1)
+                        .unsqueeze(-1)
+                        .expand(-1, -1, 1, stacked_hidden.size(-1))
+                    )
+                    final_hidden_states = torch.gather(
+                        stacked_hidden, 2, gather_index
+                    ).squeeze(2)
+                    logits = self.lm_head(_select_token_positions(final_hidden_states))
+                elif use_weighted_exit:
+                    logits = compute_expected_logits()
+
+            if logits is None:
+                hidden_states = outputs.last_hidden_state
+                logits = self.lm_head(_select_token_positions(hidden_states))
+
+        result = CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+        return result
+
+
+class OuroForSequenceClassification(
+    GenericForSequenceClassification, OuroPreTrainedModel
+):
+    pass
+
+
+class OuroForTokenClassification(GenericForTokenClassification, OuroPreTrainedModel):
+    pass
+
+
+class OuroForQuestionAnswering(GenericForQuestionAnswering, OuroPreTrainedModel):
+    base_model_prefix = (
+        "transformer"  # For BC, where `transformer` was used instead of `model`
+    )
+
+
+__all__ = [
+    "OuroPreTrainedModel",
+    "OuroModel",
+    "OuroForCausalLM",
+    "OuroForSequenceClassification",
+    "OuroForTokenClassification",
+    "OuroForQuestionAnswering",
+]

qouro/modeling_qouro.py

@@ -0,0 +1,110 @@
+from __future__ import annotations
+
+from copy import deepcopy
+from typing import Dict, Iterable, Tuple
+
+import torch
+from torch import nn
+import fnmatch
+
+from .modeling_ouro import OuroForCausalLM
+from .quantization.modules import QuantizedLinear, SmoothQuantLinear
+
+
+def _resolve_parent_module(
+    root: nn.Module, qualified_name: str
+) -> Tuple[nn.Module | None, str]:
+    if not qualified_name:
+        return None, qualified_name
+    path = qualified_name.split(".")
+    parent = root
+    for item in path[:-1]:
+        parent = getattr(parent, item)
+    return parent, path[-1]
+
+
+def _collect_linear_modules(model: nn.Module) -> Iterable[Tuple[str, nn.Linear]]:
+    for name, module in model.named_modules():
+        if isinstance(module, nn.Linear):
+            yield name, module
+
+
+class OuroForCausalLMQuantized(OuroForCausalLM):
+    """
+    Quantized variant of `OuroForCausalLM` that relies on `QuantizedLinear` modules.
+    """
+
+    def __init__(self, config):
+        quant_config = getattr(config, "quantization", None)
+        if quant_config is None:
+            raise ValueError(
+                "The provided config does not contain 'quantization' settings required "
+                "to instantiate OuroForCausalLMQuantized."
+            )
+        self.quantization_config = deepcopy(quant_config)
+        super().__init__(config)
+        self._replace_linear_layers()
+
+    def _replace_linear_layers(self) -> None:
+        method = str(self.quantization_config.get("method", "awq")).lower()
+        weight_bits = int(self.quantization_config.get("weight_bits", 4))
+        activation_bits = int(self.quantization_config.get("activation_bits", 8))
+        group_size = int(self.quantization_config.get("group_size", 128))
+        include_patterns = list(
+            self.quantization_config.get("include_modules", []) or []
+        )
+        exclude_patterns = list(
+            self.quantization_config.get("exclude_modules", []) or []
+        )
+
+        replacements = list(_collect_linear_modules(self))
+
+        # apply include/exclude filters on qualified module names
+        def _is_included(name: str) -> bool:
+            if exclude_patterns:
+                if any(fnmatch.fnmatch(name, pat) for pat in exclude_patterns):
+                    return False
+            if include_patterns:
+                if not any(fnmatch.fnmatch(name, pat) for pat in include_patterns):
+                    return False
+            return True
+
+        replacements = [(n, m) for (n, m) in replacements if _is_included(n)]
+        for name, module in replacements:
+            parent, attribute = _resolve_parent_module(self, name)
+            if parent is None:
+                continue
+            if method == "smoothquant":
+                quantized = SmoothQuantLinear(
+                    module.in_features,
+                    module.out_features,
+                    weight_bits=weight_bits,
+                    activation_bits=activation_bits,
+                    bias=module.bias is not None,
+                )
+            else:
+                quantized = QuantizedLinear(
+                    module.in_features,
+                    module.out_features,
+                    weight_bits=weight_bits,
+                    group_size=group_size,
+                    bias=module.bias is not None,
+                )
+            if module.bias is not None:
+                quantized.bias.data.copy_(module.bias.detach().to(quantized.bias.dtype))
+            setattr(parent, attribute, quantized)
+
+    def apply_quantized_weights(
+        self, weights: Dict[str, Dict[str, torch.Tensor]]
+    ) -> None:
+        module_map = dict(self.named_modules())
+        for name, tensors in weights.items():
+            candidate = module_map.get(name)
+            if candidate is None or not hasattr(candidate, "load_quant_state"):
+                continue
+            try:
+                candidate.load_quant_state(**tensors)
+            except TypeError as exc:
+                raise ValueError(
+                    f"Failed to load quantized tensors for module '{name}': {exc}"
+                ) from exc

qouro/quantization/__init__.py

@@ -0,0 +1,4 @@
+from .config import CalibrationConfig, QuantizationConfig
+from .pipeline import run_quantization_pipeline
+
+__all__ = ["CalibrationConfig", "QuantizationConfig", "run_quantization_pipeline"]

qouro/quantization/awq_core.py

@@ -0,0 +1,102 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Dict
+
+import torch
+from torch import Tensor, nn
+
+from .config import QuantizationConfig
+from .observers import ObserverState
+
+
+@dataclass
+class QuantizedWeights:
+    """Holds quantized weight representations for a linear layer."""
+
+    weight: Tensor
+    weight_scales: Tensor
+
+
+def _prepare_importance_vector(
+    observer_state: ObserverState,
+    in_features: int,
+    device: torch.device,
+    epsilon: float,
+) -> Tensor:
+    stats = observer_state.max_abs_values.to(device=device, dtype=torch.float32)
+    if stats.numel() < in_features:
+        stats = torch.nn.functional.pad(stats, (0, in_features - stats.numel()), value=1.0)
+
+    stats = stats[:in_features]
+    stats = stats.clamp_min(epsilon)
+    return stats
+
+
+def quantize_linear(
+    module: nn.Linear,
+    observer_state: ObserverState,
+    config: QuantizationConfig,
+) -> QuantizedWeights:
+    """
+    Quantize a linear layer's weights using group-wise symmetric quantization.
+    """
+
+    weight = module.weight.detach().to(torch.float32).clone()
+    device = weight.device
+    in_features = module.in_features
+    out_features = module.out_features
+    group_size = config.group_size
+    bits = config.weight_bits
+    qmin = -(2 ** (bits - 1))
+    qmax = (2 ** (bits - 1)) - 1
+    num_groups = (in_features + group_size - 1) // group_size
+
+    importance = _prepare_importance_vector(
+        observer_state, in_features, device, config.epsilon
+    )
+    if config.activation_clip is not None:
+        importance = importance.clamp(max=config.activation_clip)
+
+    quant_dtype = torch.int8 if config.weight_bits <= 8 else torch.int16
+    quantized = torch.zeros_like(weight, dtype=quant_dtype, device=device)
+    weight_scales = torch.zeros(
+        (out_features, num_groups), dtype=torch.float32, device=device
+    )
+
+    for group_idx in range(num_groups):
+        start = group_idx * group_size
+        end = min((group_idx + 1) * group_size, in_features)
+        weight_block = weight[:, start:end]
+        importance_block = importance[start:end]
+
+        weighted_block = weight_block * importance_block.unsqueeze(0)
+        max_abs = weighted_block.abs().amax(dim=1)
+        scale = (max_abs / qmax).clamp_min(config.epsilon)
+        dequant_scale = scale.unsqueeze(1)
+
+        quant_block = torch.round(weight_block / dequant_scale)
+        quant_block = quant_block.clamp(qmin, qmax)
+
+        quantized[:, start:end] = quant_block.to(quant_dtype)
+        weight_scales[:, group_idx] = scale
+
+    return QuantizedWeights(weight=quantized.cpu(), weight_scales=weight_scales.cpu())
+
+
+def summarize_quantization(
+    stats: Dict[str, QuantizedWeights]
+) -> Dict[str, Dict[str, float]]:
+    """
+    Generate simple per-layer statistics to debug quantization quality.
+    """
+
+    summary: Dict[str, Dict[str, float]] = {}
+    for name, record in stats.items():
+        scale = record.weight_scales
+        summary[name] = {
+            "scale_min": float(scale.min()),
+            "scale_max": float(scale.max()),
+            "scale_mean": float(scale.mean()),
+        }
+    return summary

qouro/quantization/calibration.py

@@ -0,0 +1,111 @@
+from __future__ import annotations
+
+from typing import Iterable, Iterator, List
+
+from datasets import IterableDataset, load_dataset
+from loguru import logger
+from torch import device as TorchDevice
+from transformers import PreTrainedTokenizerBase
+
+from .config import CalibrationConfig
+
+
+def collect_calibration_texts(config: CalibrationConfig) -> List[str]:
+    """
+    Fetch calibration samples from a Hugging Face dataset.
+
+    Returns a list of raw text prompts that will be consumed by the calibration loop.
+    """
+
+    if config.sample_count <= 0:
+        logger.warning("Calibration requested with zero samples; skipping collection.")
+        return []
+
+    logger.info(
+        f"Loading calibration dataset '{config.dataset_name}' "
+        f"(split={config.dataset_split}, streaming={config.streaming})..."
+    )
+    try:
+        dataset = load_dataset(
+            config.dataset_name,
+            split=config.dataset_split,
+            streaming=config.streaming,
+        )
+    except Exception as exc:  # noqa: BLE001 - surface dataset issues to the caller
+        logger.warning(
+            f"Unable to load dataset '{config.dataset_name}': {exc}"
+        )
+        return []
+
+    if isinstance(dataset, IterableDataset):
+        iterator: Iterable[dict] = dataset
+        samples: List[str] = []
+        for example in iterator:
+            text = example.get(config.text_column)
+            if not text:
+                continue
+            samples.append(str(text))
+            if len(samples) >= config.sample_count:
+                break
+        return samples
+
+    if len(dataset) == 0:
+        logger.warning(f"Dataset '{config.dataset_name}' returned no rows.")
+        return []
+
+    upper = min(config.sample_count, len(dataset))
+    if config.shuffle:
+        dataset = dataset.shuffle(seed=config.seed)
+    selected = dataset.select(range(upper))
+    texts = []
+    for entry in selected:
+        text = entry.get(config.text_column)
+        if not isinstance(text, str):
+            continue
+        texts.append(text)
+    if not texts:
+        logger.warning(
+            f"Failed to collect calibration texts from column '{config.text_column}'."
+        )
+    return texts
+
+
+def iter_tokenized_batches(
+    tokenizer: PreTrainedTokenizerBase,
+    texts: Iterable[str],
+    device: TorchDevice,
+    batch_size: int,
+    max_length: int,
+) -> Iterator[dict]:
+    """
+    Yield tokenized calibration inputs suitable for feeding into the model.
+    """
+
+    if batch_size <= 0:
+        raise ValueError("Batch size must be at least 1 for calibration.")
+
+    buffer: List[str] = []
+    for text in texts:
+        buffer.append(text)
+        if len(buffer) < batch_size:
+            continue
+
+        batch_inputs = tokenizer(
+            buffer,
+            padding=True,
+            truncation=True,
+            max_length=max_length,
+            return_tensors="pt",
+        ).to(device)
+        yield batch_inputs
+        buffer.clear()
+
+    if buffer:
+        batch_inputs = tokenizer(
+            buffer,
+            padding=True,
+            truncation=True,
+            max_length=max_length,
+            return_tensors="pt",
+        ).to(device)
+        yield batch_inputs

qouro/quantization/config.py

@@ -0,0 +1,57 @@
+from __future__ import annotations
+
+from dataclasses import dataclass, field, asdict
+from typing import Dict, List
+
+
+@dataclass
+class CalibrationConfig:
+    """Settings that govern how calibration samples are collected."""
+
+    dataset_name: str = "mit-han-lab/pile-val-backup"
+    dataset_split: str = "validation"
+    text_column: str = "text"
+    sample_count: int = 128
+    max_sequence_length: int = 512
+    batch_size: int = 8
+    shuffle: bool = False
+    streaming: bool = False
+    seed: int = 0
+
+    def to_dict(self) -> Dict[str, int | str | bool]:
+        return asdict(self)
+
+
+@dataclass
+class QuantizationConfig:
+    """Configuration for quantization pipelines (AWQ, SmoothQuant, etc.)."""
+
+    enabled: bool = True
+    method: str = "awq"
+    weight_bits: int = 4
+    activation_bits: int = 8
+    group_size: int = 128
+    per_channel: bool = True
+    calibration: CalibrationConfig = field(default_factory=CalibrationConfig)
+    activation_clip: float | None = None
+    epsilon: float = 1e-5
+    alpha: float = 0.5
+    # module name filters (glob patterns allowed), applied on qualified names from named_modules()
+    include_modules: List[str] = field(default_factory=list)
+    exclude_modules: List[str] = field(default_factory=list)
+
+    def to_dict(self) -> Dict[str, int | float | bool | Dict[str, int | str | bool]]:
+        return {
+            "enabled": self.enabled,
+            "method": self.method,
+            "weight_bits": self.weight_bits,
+            "activation_bits": self.activation_bits,
+            "group_size": self.group_size,
+            "per_channel": self.per_channel,
+            "calibration": self.calibration.to_dict(),
+            "activation_clip": self.activation_clip,
+            "epsilon": self.epsilon,
+            "alpha": self.alpha,
+            "include_modules": list(self.include_modules),
+            "exclude_modules": list(self.exclude_modules),
+        }

qouro/quantization/modules.py

@@ -0,0 +1,221 @@
+from __future__ import annotations
+
+import math
+from typing import Optional
+
+import torch
+from torch import Tensor, nn
+
+
+def _select_quant_dtype(bits: int) -> torch.dtype:
+    if bits <= 0:
+        raise ValueError("Quantization bits must be positive.")
+    if bits <= 8:
+        return torch.int8
+    if bits <= 16:
+        return torch.int16
+    raise ValueError("Quantization bits above 16 are not supported.")
+
+
+class QuantizedLinear(nn.Module):
+    """Weight-only linear layer with per-group scales."""
+
+    def __init__(
+        self,
+        in_features: int,
+        out_features: int,
+        *,
+        weight_bits: int = 4,
+        group_size: int = 128,
+        bias: bool = True,
+    ) -> None:
+        super().__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.weight_bits = weight_bits
+        self.group_size = group_size
+        self.qmin = -(2 ** (weight_bits - 1))
+        self.qmax = (2 ** (weight_bits - 1)) - 1
+        self.num_groups = math.ceil(in_features / group_size)
+        self.quant_dtype = _select_quant_dtype(weight_bits)
+
+        weight_shape = (out_features, in_features)
+        scale_shape = (out_features, self.num_groups)
+        self.register_buffer("weight", torch.zeros(weight_shape, dtype=self.quant_dtype))
+        self.register_buffer(
+            "weight_scales", torch.ones(scale_shape, dtype=torch.float32)
+        )
+
+        self.bias = nn.Parameter(torch.zeros(out_features)) if bias else None
+        self._weight_cache: Optional[Tensor] = None
+
+    def _invalidate_cache(self) -> None:
+        self._weight_cache = None
+
+    def refresh_weight_cache(self) -> None:
+        self._weight_cache = self._dequantize_weight()
+
+    def _dequantize_weight(self) -> Tensor:
+        group_tensors = []
+        for group_idx in range(self.num_groups):
+            start = group_idx * self.group_size
+            end = min((group_idx + 1) * self.group_size, self.in_features)
+            block = self.weight[:, start:end].float()
+            scale = self.weight_scales[:, group_idx].unsqueeze(1)
+            group_tensors.append(block * scale)
+        return torch.cat(group_tensors, dim=1)
+
+    def forward(self, input: Tensor) -> Tensor:
+        if self._weight_cache is None or self._weight_cache.device != input.device:
+            self.refresh_weight_cache()
+            self._weight_cache = self._weight_cache.to(input.device)
+
+        weight = self._weight_cache
+        if weight.dtype != input.dtype:
+            weight = weight.to(input.dtype)
+
+        bias = self.bias
+        if bias is not None and bias.device != input.device:
+            bias = bias.to(input.device)
+        if bias is not None and bias.dtype != input.dtype:
+            bias = bias.to(input.dtype)
+
+        return nn.functional.linear(input, weight, bias)
+
+    def load_quant_state(self, weight: Tensor, weight_scales: Tensor) -> None:
+        if weight.shape != self.weight.shape:
+            raise ValueError(
+                f"Quantized weight shape mismatch: expected {tuple(self.weight.shape)}, "
+                f"got {tuple(weight.shape)}"
+            )
+        if weight_scales.shape != self.weight_scales.shape:
+            raise ValueError(
+                f"Scale tensor shape mismatch: expected {tuple(self.weight_scales.shape)}, "
+                f"got {tuple(weight_scales.shape)}"
+            )
+        self.weight.copy_(weight.to(dtype=self.quant_dtype))
+        self.weight_scales.copy_(weight_scales.to(dtype=torch.float32))
+        self._invalidate_cache()
+
+    def extra_repr(self) -> str:
+        return (
+            f"in_features={self.in_features}, out_features={self.out_features}, "
+            f"group_size={self.group_size}, bits={self.weight_bits}, bias={self.bias is not None}"
+        )
+
+
+class SmoothQuantLinear(nn.Module):
+    """Linear layer with SmoothQuant W8A8 (or configurable) quantization."""
+
+    def __init__(
+        self,
+        in_features: int,
+        out_features: int,
+        *,
+        weight_bits: int = 8,
+        activation_bits: int = 8,
+        bias: bool = True,
+    ) -> None:
+        super().__init__()
+        if weight_bits <= 0 or weight_bits > 16:
+            raise ValueError("Weight bits must be in range [1, 16].")
+        if activation_bits <= 0 or activation_bits > 16:
+            raise ValueError("Activation bits must be in range [1, 16].")
+        self.in_features = in_features
+        self.out_features = out_features
+        self.weight_bits = weight_bits
+        self.activation_bits = activation_bits
+        self.weight_qmin = -(2 ** (weight_bits - 1))
+        self.weight_qmax = (2 ** (weight_bits - 1)) - 1
+        self.activation_qmin = -(2 ** (activation_bits - 1))
+        self.activation_qmax = (2 ** (activation_bits - 1)) - 1
+        self.quant_dtype = _select_quant_dtype(weight_bits)
+
+        weight_shape = (out_features, in_features)
+        self.register_buffer("weight", torch.zeros(weight_shape, dtype=self.quant_dtype))
+        self.register_buffer(
+            "weight_scales", torch.ones(out_features, 1, dtype=torch.float32)
+        )
+        self.register_buffer(
+            "input_scale", torch.ones(in_features, dtype=torch.float32)
+        )
+        self.register_buffer(
+            "activation_scale", torch.ones(in_features, dtype=torch.float32)
+        )
+
+        self.bias = nn.Parameter(torch.zeros(out_features)) if bias else None
+        self._weight_cache: Optional[Tensor] = None
+
+    def _invalidate_cache(self) -> None:
+        self._weight_cache = None
+
+    def refresh_weight_cache(self) -> None:
+        weight = self.weight.float() * self.weight_scales
+        self._weight_cache = weight
+
+    def forward(self, input: Tensor) -> Tensor:
+        if self._weight_cache is None or self._weight_cache.device != input.device:
+            self.refresh_weight_cache()
+            self._weight_cache = self._weight_cache.to(input.device)
+
+        activation_scale = self.activation_scale.to(input.device)
+        input_scale = self.input_scale.to(input.device)
+
+        scaled_input = input * input_scale
+        quantized = torch.round(scaled_input / activation_scale).clamp(
+            self.activation_qmin, self.activation_qmax
+        )
+        dequant_input = quantized * activation_scale
+
+        weight = self._weight_cache
+        if weight.dtype != dequant_input.dtype:
+            weight = weight.to(dequant_input.dtype)
+
+        bias = self.bias
+        if bias is not None and bias.device != input.device:
+            bias = bias.to(input.device)
+        if bias is not None and bias.dtype != dequant_input.dtype:
+            bias = bias.to(dequant_input.dtype)
+
+        return nn.functional.linear(dequant_input, weight, bias)
+
+    def load_quant_state(
+        self,
+        weight: Tensor,
+        weight_scales: Tensor,
+        input_scale: Tensor,
+        activation_scale: Tensor,
+    ) -> None:
+        if weight.shape != self.weight.shape:
+            raise ValueError(
+                f"Quantized weight shape mismatch: expected {tuple(self.weight.shape)}, "
+                f"got {tuple(weight.shape)}"
+            )
+        if weight_scales.shape != self.weight_scales.shape:
+            raise ValueError(
+                f"Weight scale shape mismatch: expected {tuple(self.weight_scales.shape)}, "
+                f"got {tuple(weight_scales.shape)}"
+            )
+        if input_scale.shape != self.input_scale.shape:
+            raise ValueError(
+                f"Input scale shape mismatch: expected {tuple(self.input_scale.shape)}, "
+                f"got {tuple(input_scale.shape)}"
+            )
+        if activation_scale.shape != self.activation_scale.shape:
+            raise ValueError(
+                f"Activation scale shape mismatch: expected {tuple(self.activation_scale.shape)}, "
+                f"got {tuple(activation_scale.shape)}"
+            )
+
+        self.weight.copy_(weight.to(dtype=self.quant_dtype))
+        self.weight_scales.copy_(weight_scales.to(dtype=torch.float32))
+        self.input_scale.copy_(input_scale.to(dtype=torch.float32))
+        self.activation_scale.copy_(activation_scale.to(dtype=torch.float32))
+        self._invalidate_cache()
+
+    def extra_repr(self) -> str:
+        return (
+            f"in_features={self.in_features}, out_features={self.out_features}, "
+            f"weight_bits={self.weight_bits}, activation_bits={self.activation_bits}, "
+            f"bias={self.bias is not None}"
+        )

qouro/quantization/observers.py

@@ -0,0 +1,71 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+from torch import Tensor, nn
+
+
+@dataclass
+class ObserverState:
+    """Activation statistics collected for a linear layer."""
+
+    max_abs_values: Tensor
+
+
+class LinearInputObserver:
+    """
+    Collects per-feature activation maxima for a `nn.Linear` module.
+
+    The statistics are later used to derive quantization scales that take input
+    distribution into account, mimicking the behaviour of AWQ.
+    """
+
+    def __init__(self, module_name: str):
+        self.module_name = module_name
+        self._max_abs: Optional[Tensor] = None
+
+    def __call__(self, module: nn.Module, inputs: tuple[Tensor, ...]) -> None:
+        if not inputs:
+            return
+
+        data = inputs[0]
+        if data is None:
+            return
+
+        if data.dim() > 2:
+            data = data.reshape(-1, data.size(-1))
+        elif data.dim() < 2:
+            data = data.unsqueeze(0)
+
+        data = data.detach()
+        if data.dtype in (torch.float16, torch.bfloat16):
+            data = data.to(torch.float32)
+
+        max_vals = data.abs().amax(dim=0)
+        if self._max_abs is None:
+            self._max_abs = max_vals
+        else:
+            # Pad in case dimensionality changes due to mixed inputs.
+            if max_vals.size(0) != self._max_abs.size(0):
+                target = max(self._max_abs.size(0), max_vals.size(0))
+                self._max_abs = torch.nn.functional.pad(
+                    self._max_abs,
+                    (0, target - self._max_abs.size(0)),
+                    value=0.0,
+                )
+                max_vals = torch.nn.functional.pad(
+                    max_vals,
+                    (0, target - max_vals.size(0)),
+                    value=0.0,
+                )
+            self._max_abs = torch.maximum(self._max_abs, max_vals)
+
+    def to_state(self) -> ObserverState:
+        if self._max_abs is None:
+            raise RuntimeError(
+                f"No activation statistics recorded for module '{self.module_name}'."
+            )
+        return ObserverState(max_abs_values=self._max_abs)
+

qouro/quantization/pipeline.py

@@ -0,0 +1,162 @@
+from __future__ import annotations
+
+import copy
+from typing import Dict, Tuple
+
+import torch
+from loguru import logger
+from torch import nn
+from transformers import PreTrainedModel, PreTrainedTokenizerBase
+
+from .awq_core import QuantizedWeights, quantize_linear, summarize_quantization
+from .calibration import collect_calibration_texts, iter_tokenized_batches
+from .config import QuantizationConfig
+from .observers import LinearInputObserver
+from .smoothquant import (
+    SmoothQuantWeights,
+    quantize_linear_smooth,
+    summarize_smoothquant,
+)
+from ..modeling_qouro import OuroForCausalLMQuantized
+
+
+def _gather_linear_modules(model: nn.Module) -> Dict[str, nn.Linear]:
+    return {
+        name: module
+        for name, module in model.named_modules()
+        if isinstance(module, nn.Linear)
+    }
+
+
+def _attach_observers(
+    modules: Dict[str, nn.Linear],
+) -> Tuple[Dict[str, LinearInputObserver], list[torch.utils.hooks.RemovableHandle]]:
+    observers: Dict[str, LinearInputObserver] = {}
+    handles: list[torch.utils.hooks.RemovableHandle] = []
+    for name, module in modules.items():
+        observer = LinearInputObserver(name)
+        handle = module.register_forward_pre_hook(observer)
+        observers[name] = observer
+        handles.append(handle)
+    return observers, handles
+
+
+def _detach_handles(handles: list[torch.utils.hooks.RemovableHandle]) -> None:
+    for handle in handles:
+        handle.remove()
+
+
+def run_quantization_pipeline(
+    base_model: PreTrainedModel,
+    tokenizer: PreTrainedTokenizerBase,
+    device: torch.device,
+    config: QuantizationConfig,
+) -> Tuple[OuroForCausalLMQuantized, Dict[str, Dict[str, float]]]:
+
+    if not config.enabled:
+        raise ValueError("Quantization pipeline requested while disabled in config.")
+
+    calibration_texts = collect_calibration_texts(config.calibration)
+    if not calibration_texts:
+        raise RuntimeError(
+            "Unable to collect calibration texts; aborting quantization."
+        )
+
+    linear_modules = _gather_linear_modules(base_model)
+    if not linear_modules:
+        raise RuntimeError("No linear modules found in model; cannot quantize.")
+
+    observers, handles = _attach_observers(linear_modules)
+    base_model.eval()
+
+    logger.info(f"Running calibration with {len(calibration_texts)} texts...")
+    with torch.no_grad():
+        for batch_inputs in iter_tokenized_batches(
+            tokenizer=tokenizer,
+            texts=calibration_texts,
+            device=device,
+            batch_size=config.calibration.batch_size,
+            max_length=config.calibration.max_sequence_length,
+        ):
+            base_model(**batch_inputs)
+
+    _detach_handles(handles)
+
+    method = config.method.lower()
+
+    observer_states: Dict[str, LinearInputObserver] = {}
+    for name, observer in observers.items():
+        observer_states[name] = observer
+
+    quantized_payload: Dict[str, Dict[str, torch.Tensor]] = {}
+    summary: Dict[str, Dict[str, float]] = {}
+
+    if method == "awq":
+        awq_weights: Dict[str, QuantizedWeights] = {}
+        for name, module in linear_modules.items():
+            observer = observer_states.get(name)
+            if observer is None:
+                continue
+            try:
+                state = observer.to_state()
+            except RuntimeError as exc:
+                logger.warning(f"Skipping module '{name}': {exc}")
+                continue
+            awq_weights[name] = quantize_linear(module, state, config)
+
+        logger.info(f"Quantized {len(awq_weights)} linear modules with AWQ.")
+
+        for name, record in awq_weights.items():
+            quantized_payload[name] = {
+                "weight": record.weight,
+                "weight_scales": record.weight_scales,
+            }
+        summary = summarize_quantization(awq_weights)
+    elif method == "smoothquant":
+        smooth_weights: Dict[str, SmoothQuantWeights] = {}
+        for name, module in linear_modules.items():
+            observer = observer_states.get(name)
+            if observer is None:
+                continue
+            try:
+                state = observer.to_state()
+            except RuntimeError as exc:
+                logger.warning(f"Skipping module '{name}': {exc}")
+                continue
+            smooth_weights[name] = quantize_linear_smooth(module, state, config)
+
+        logger.info(f"Quantized {len(smooth_weights)} linear modules with SmoothQuant.")
+
+        for name, record in smooth_weights.items():
+            quantized_payload[name] = {
+                "weight": record.weight,
+                "weight_scales": record.weight_scales,
+                "input_scale": record.input_scale,
+                "activation_scale": record.activation_scale,
+            }
+        summary = summarize_smoothquant(smooth_weights)
+    else:
+        raise ValueError(f"Unsupported quantization method '{config.method}'.")
+
+    quant_config_dict = config.to_dict()
+    quantized_config = copy.deepcopy(base_model.config)
+    quantized_config.quantization = quant_config_dict
+    quantized_config.architectures = ["OuroForCausalLMQuantized"]
+    quantized_config.auto_map = {
+        "AutoModelForCausalLM": "qouro.modeling_qouro::OuroForCausalLMQuantized"
+    }
+
+    quantized_model = OuroForCausalLMQuantized(quantized_config)
+
+    float_state = base_model.state_dict()
+    missing, unexpected = quantized_model.load_state_dict(float_state, strict=False)
+    if unexpected:
+        logger.debug(f"Unexpected keys during transfer: {unexpected}")
+    if missing:
+        logger.debug(f"Missing keys during transfer: {missing}")
+
+    quantized_model.apply_quantized_weights(quantized_payload)
+    quantized_model.to(device)
+    quantized_model.eval()
+
+    return quantized_model, summary

qouro/quantization/smoothquant.py

@@ -0,0 +1,95 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Dict
+
+import torch
+from torch import Tensor, nn
+
+from .config import QuantizationConfig
+from .observers import ObserverState
+
+
+@dataclass
+class SmoothQuantWeights:
+    """Quantized representation for SmoothQuant linear layers."""
+
+    weight: Tensor
+    weight_scales: Tensor
+    input_scale: Tensor
+    activation_scale: Tensor
+
+
+def _prepare_stats(
+    observer_state: ObserverState,
+    weight: Tensor,
+    epsilon: float,
+) -> tuple[Tensor, Tensor]:
+    activation_stats = observer_state.max_abs_values.to(dtype=torch.float32)
+    if activation_stats.numel() < weight.size(1):
+        activation_stats = torch.nn.functional.pad(
+            activation_stats,
+            (0, weight.size(1) - activation_stats.numel()),
+            value=1.0,
+        )
+    activation_stats = activation_stats[: weight.size(1)]
+    activation_stats = activation_stats.clamp_min(epsilon)
+
+    weight_stats = weight.abs().amax(dim=0).clamp_min(epsilon)
+    return activation_stats, weight_stats
+
+
+def quantize_linear_smooth(
+    module: nn.Linear,
+    observer_state: ObserverState,
+    config: QuantizationConfig,
+) -> SmoothQuantWeights:
+    """
+    Apply SmoothQuant to a linear layer, producing int quantized weights and activation scales.
+    """
+
+    weight_bits = config.weight_bits
+    activation_bits = config.activation_bits
+    epsilon = config.epsilon
+    alpha = config.alpha
+    quant_dtype = torch.int8 if weight_bits <= 8 else torch.int16
+
+    weight = module.weight.detach().to(torch.float32).clone()
+    activation_stats, weight_stats = _prepare_stats(observer_state, weight, epsilon)
+
+    ratio = activation_stats / weight_stats
+    smoothing_factor = torch.pow(ratio, alpha).clamp_min(epsilon)
+
+    input_scale = (1.0 / smoothing_factor).to(torch.float32)
+    scaled_weight = weight * smoothing_factor.unsqueeze(0)
+
+    act_max_scaled = activation_stats * input_scale
+    act_qmax = (2 ** (activation_bits - 1)) - 1
+    activation_scale = (act_max_scaled / act_qmax).clamp_min(epsilon)
+
+    weight_qmax = (2 ** (weight_bits - 1)) - 1
+    weight_max = scaled_weight.abs().amax(dim=1).clamp_min(epsilon)
+    weight_scales = (weight_max / weight_qmax).unsqueeze(1)
+
+    quantized_weight = torch.round(scaled_weight / weight_scales).clamp(
+        -(2 ** (weight_bits - 1)), weight_qmax
+    ).to(quant_dtype)
+
+    return SmoothQuantWeights(
+        weight=quantized_weight.cpu(),
+        weight_scales=weight_scales.to(torch.float32).cpu(),
+        input_scale=input_scale.cpu(),
+        activation_scale=activation_scale.cpu(),
+    )
+
+
+def summarize_smoothquant(
+    stats: Dict[str, SmoothQuantWeights]
+) -> Dict[str, Dict[str, float]]:
+    summary: Dict[str, Dict[str, float]] = {}
+    for name, record in stats.items():
+        summary[name] = {
+            "weight_scale_mean": float(record.weight_scales.mean()),
+            "activation_scale_mean": float(record.activation_scale.mean()),
+        }
+    return summary