image-match/python/py/Lib/site-packages/transformers/models/glmasr/modeling_glmasr.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/glmasr/modular_glmasr.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_glmasr.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2025 the HuggingFace 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.

from collections.abc import Callable
from typing import Optional

from ...activations import ACT2FN
from ...cache_utils import Cache
from ...generation import GenerationMixin
from ...integrations import use_kernelized_func
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutputWithPooling, CausalLMOutputWithPast
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, is_torch_available
from ...utils.generic import can_return_tuple, maybe_autocast, merge_with_config_defaults
from ...utils.output_capturing import capture_outputs
from ..auto import AutoModel, AutoModelForCausalLM
from .configuration_glmasr import GlmAsrConfig, GlmAsrEncoderConfig


if is_torch_available():
    import torch
    from torch import nn


class GlmAsrRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: GlmAsrConfig, device=None):
        super().__init__()
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config

        self.rope_type = self.config.rope_parameters["rope_type"]
        rope_init_fn: Callable = self.compute_default_rope_parameters
        if self.rope_type != "default":
            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)

        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)

    @staticmethod
    def compute_default_rope_parameters(
        config: GlmAsrConfig | None = None,
        device: Optional["torch.device"] = None,
        seq_len: int | None = None,
    ) -> tuple["torch.Tensor", float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            device (`torch.device`):
                The device to use for initialization of the inverse frequencies.
            seq_len (`int`, *optional*):
                The current sequence length. Unused for this type of RoPE.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        """
        base = config.rope_parameters["rope_theta"]
        partial_rotary_factor = config.rope_parameters.get("partial_rotary_factor", 1.0)
        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
        dim = int(head_dim * partial_rotary_factor)

        attention_factor = 1.0  # Unused in this type of RoPE

        # Compute the inverse frequencies
        inv_freq = 1.0 / (
            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
        )
        return inv_freq, attention_factor

    @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 maybe_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)


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 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: torch.Tensor | None,
    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:
        attn_weights = attn_weights + attention_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


def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)

    rotary_dim = cos.shape[-1]
    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]

    # Apply rotary embeddings on the first half or full tensor
    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)

    # Concatenate back to full shape
    q_embed = torch.cat([q_embed, q_pass], dim=-1)
    k_embed = torch.cat([k_embed, k_pass], dim=-1)
    return q_embed, k_embed


@use_kernelized_func(apply_rotary_pos_emb)
class GlmAsrAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: GlmAsrConfig, 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 = False
        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=True)
        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=True)
        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=True)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, 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)

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask=None,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class GlmAsrMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, hidden_states: torch.Tensor):
        hidden_states = self.fc1(hidden_states)
        hidden_states = self.act_fn(hidden_states)
        hidden_states = self.fc2(hidden_states)
        return hidden_states


class GlmAsrEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: GlmAsrConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        self.self_attn = GlmAsrAttention(config=config, layer_idx=layer_idx)

        self.mlp = GlmAsrMLP(config)
        self.input_layernorm = nn.LayerNorm(config.hidden_size)
        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        # Self Attention
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        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 = residual + hidden_states
        return hidden_states


@auto_docstring
class GlmAsrPreTrainedModel(PreTrainedModel):
    config: GlmAsrConfig
    base_model_prefix = "model"
    input_modalities = ("audio", "text")
    supports_gradient_checkpointing = True
    _no_split_modules = ["GlmAsrAttention"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn = True
    _supports_sdpa = True


# TODO: @eustlb, this is what WhisperEncoder should look like
class GlmAsrEncoder(GlmAsrPreTrainedModel):
    config: GlmAsrEncoderConfig
    main_input_name = "input_features"
    input_modalities = "audio"
    _no_split_modules = ["GlmAsrEncoderLayer"]
    _can_record_outputs = {
        "hidden_states": GlmAsrEncoderLayer,
        "attentions": GlmAsrAttention,
    }

    def __init__(self, config: GlmAsrEncoderConfig):
        super().__init__(config)
        self.conv1 = nn.Conv1d(config.num_mel_bins, config.hidden_size, kernel_size=3, padding=1)
        self.conv2 = nn.Conv1d(config.hidden_size, config.hidden_size, kernel_size=3, stride=2, padding=1)

        self.layers = nn.ModuleList(
            [GlmAsrEncoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = nn.LayerNorm(config.hidden_size)
        self.rotary_emb = GlmAsrRotaryEmbedding(config=config)
        self.gradient_checkpointing = False
        self.post_init()

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(self, input_features, **kwargs: Unpack[TransformersKwargs]):
        inputs_embeds = nn.functional.gelu(self.conv1(input_features))
        inputs_embeds = nn.functional.gelu(self.conv2(inputs_embeds))
        inputs_embeds = inputs_embeds.transpose(1, 2)

        hidden_states = inputs_embeds
        position_embeddings = self.rotary_emb(
            hidden_states, position_ids=torch.arange(hidden_states.shape[1], device=hidden_states.device)[None, :]
        )

        for encoder_layer in self.layers:
            hidden_states = encoder_layer(hidden_states, position_embeddings=position_embeddings, **kwargs)

        hidden_states = self.norm(hidden_states)
        return BaseModelOutputWithPooling(last_hidden_state=hidden_states)


class GlmAsrMultiModalProjector(nn.Module):
    """
    Audio adaptor (small MLP) that projects GlmAsrEncoder features
    to the LLM embedding space so they can replace `<sound>` tokens.
    """

    def __init__(self, config: GlmAsrConfig):
        super().__init__()
        self.linear_1 = nn.Linear(config.audio_config.intermediate_size, config.text_config.hidden_size * 2)
        self.act = ACT2FN[config.projector_hidden_act]
        self.linear_2 = nn.Linear(config.text_config.hidden_size * 2, config.text_config.hidden_size)

    def forward(self, audio_features):
        hidden_states = self.linear_1(audio_features)
        hidden_states = self.act(hidden_states)
        hidden_states = self.linear_2(hidden_states)
        return hidden_states


@auto_docstring(
    custom_intro="""
    The GlmAsr model which consists of a fine-tuned Whisper encoder, a multi-modal projector and a Llama language model.
    """
)
class GlmAsrForConditionalGeneration(GlmAsrPreTrainedModel, GenerationMixin):
    _keep_in_fp32_modules_strict = None
    _tp_plan = None
    _pp_plan = None

    def __init__(self, config):
        super().__init__(config)
        self.vocab_size = config.text_config.vocab_size
        self.audio_tower = AutoModel.from_config(config.audio_config)
        self.language_model = AutoModelForCausalLM.from_config(config.text_config)
        self.multi_modal_projector = GlmAsrMultiModalProjector(config)

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

    def get_output_embeddings(self):
        return self.language_model.get_output_embeddings()

    def set_output_embeddings(self, new_embeddings):
        self.language_model.set_output_embeddings(new_embeddings)

    def set_decoder(self, decoder):
        self.language_model.set_decoder(decoder)

    def get_decoder(self):
        return self.language_model.get_decoder()

    @can_return_tuple
    @auto_docstring(
        custom_intro="Compute audio embeddings from log-mel input features using the audio encoder and multi-modal projector."
    )
    def get_audio_features(
        self,
        input_features: torch.FloatTensor,
        input_features_mask: torch.Tensor,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        input_features (`torch.FloatTensor`):
            Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be
            obtained by loading a `.flac` or `.wav` audio file into an array of type `list[float]` or a
            `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
            `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding
            and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
        input_features_mask (`torch.Tensor` of shape `(batch_size, feature_sequence_length)`):
            Mask to avoid performing attention on padded feature indices.
        """
        audio_outputs = self.audio_tower(input_features, return_dict=True, **kwargs)
        audio_hidden_states = audio_outputs.last_hidden_state
        audio_hidden_states = audio_hidden_states.reshape(
            input_features.shape[0], -1, self.config.audio_config.intermediate_size
        )
        audio_embeds = self.multi_modal_projector(audio_hidden_states)

        audio_lengths = input_features_mask.sum(-1)
        for padding, kernel_size, stride in [(1, 3, 1), (1, 3, 2)]:
            audio_lengths = (audio_lengths + 2 * padding - (kernel_size - 1) - 1) // stride + 1
        merge_factor = 4
        post_lengths = (audio_lengths - merge_factor) // merge_factor + 1

        valid_mask = torch.arange(audio_embeds.shape[1], device=post_lengths.device)[None, :] < post_lengths[:, None]
        audio_outputs.pooler_output = audio_embeds[valid_mask.to(audio_embeds.device)]

        return audio_outputs

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        input_features: torch.FloatTensor | None = None,
        input_features_mask: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> CausalLMOutputWithPast:
        r"""
        input_features_mask (`torch.Tensor` of shape `(batch_size, feature_sequence_length)`):
            Mask to avoid performing attention on padding feature indices. Mask values selected in `[0, 1]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

        Example:

        ```python
        >>> from transformers import GlmAsrForConditionalGeneration, AutoProcessor

        >>> model_id = "zai-org/GLM-ASR-Nano-2512"
        >>> processor = AutoProcessor.from_pretrained(model_id)
        >>> model = GlmAsrForConditionalGeneration.from_pretrained(model_id, dtype="auto", device_map="auto")
        >>> inputs = processor.apply_transcription_request("https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/bcn_weather.mp3")

        >>> inputs = inputs.to(model.device, dtype=model.dtype)

        >>> outputs = model.generate(**inputs, do_sample=False, max_new_tokens=500)

        >>> decoded_outputs = processor.batch_decode(outputs[:, inputs.input_ids.shape[1] :], skip_special_tokens=True)
        >>> print(decoded_outputs)
        ```"""

        if inputs_embeds is None:
            inputs_embeds = self.get_input_embeddings()(input_ids)

        if input_features is not None and input_ids is not None:
            audio_embeds = self.get_audio_features(input_features, input_features_mask, return_dict=True).pooler_output

            # replace text-audio token placeholders with audio embeddings
            audio_token_mask = (input_ids == self.config.audio_token_id).unsqueeze(-1)
            inputs_embeds = inputs_embeds.masked_scatter(
                audio_token_mask.to(inputs_embeds.device), audio_embeds.to(inputs_embeds.device)
            )

        outputs: CausalLMOutputWithPast = self.language_model(
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            labels=labels,
            use_cache=use_cache,
            logits_to_keep=logits_to_keep,
            **kwargs,
        )
        return outputs

    def prepare_inputs_for_generation(self, *args, is_first_iteration: bool = False, **kwargs):
        input_features = kwargs.pop("input_features", None)
        input_features_mask = kwargs.pop("input_features_mask", None)

        model_inputs = super().prepare_inputs_for_generation(*args, **kwargs)

        if is_first_iteration or not model_inputs.get("use_cache", False):
            if input_features is not None:
                model_inputs["input_features"] = input_features
            if input_features_mask is not None:
                model_inputs["input_features_mask"] = input_features_mask

        return model_inputs


__all__ = ["GlmAsrEncoder", "GlmAsrForConditionalGeneration", "GlmAsrPreTrainedModel"]