from collections.abc import Callable from typing import Optional, Union, Any import math import copy import torch import torch.nn.functional as F from torch import nn from transformers.activations import ACT2FN from transformers.cache_utils import Cache, DynamicCache from transformers.generation import GenerationMixin 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 transformers.utils.import_utils import is_causal_conv1d_available from .configuration_gear import GearConfig if is_causal_conv1d_available(): from causal_conv1d import causal_conv1d_fn, causal_conv1d_update else: causal_conv1d_fn, causal_conv1d_update = None, None kernel_modules = (causal_conv1d_fn, causal_conv1d_update) is_fast_path_available = all(kernel_modules) class GearHybridKVConvCache: """ Attention and conv cache for Gear. It stores the Key and Value states as a list of tensors, one for each layer. Attention layer cache shape: `[batch_size, num_heads, seq_len, head_dim]`. Key-Value Conv layer cache shape: `[batch_size, hidden_size, L_cache-1]`. """ # Override @property existing in Cache max_batch_size = None is_compileable = False key_cache = None value_cache = None def __init__( self, config: GearConfig, max_batch_size: int, dtype: torch.dtype = torch.float32, device: Union[torch.device, str, None] = None, ): self.key_cache = [] self.value_cache = [] self.max_batch_size = max_batch_size self.layer_types = config.layer_types self.first_attention_layer = self.layer_types.index("full_attention") self.conv_L_cache = config.conv_L_cache self._dtype = dtype self.key_conv_cache: list[torch.Tensor] = [] self.value_conv_cache: list[torch.Tensor] = [] device = torch.device(device) if device is not None else None for _ in range(config.num_hidden_layers): key_conv_state = torch.zeros( self.max_batch_size, config.num_key_value_heads * config.head_dim, self.conv_L_cache, dtype=self._dtype, device=device, ) value_conv_state = torch.zeros_like(key_conv_state) torch._dynamo.mark_static_address(key_conv_state) torch._dynamo.mark_static_address(value_conv_state) self.key_conv_cache.append(key_conv_state) self.value_conv_cache.append(value_conv_state) def update( self, key_states: torch.Tensor, value_states: torch.Tensor, layer_idx: int, cache_kwargs: Optional[dict[str, Any]] = None, ) -> tuple[torch.Tensor, torch.Tensor]: """ Updates the cache with the new `key_states` and `value_states` for the layer `layer_idx`. Parameters: key_states (`torch.Tensor`): The new key states to cache. value_states (`torch.Tensor`): The new value states to cache. layer_idx (`int`): The index of the layer to cache the states for. cache_kwargs (`Dict[str, Any]`, `optional`): Additional arguments for the cache subclass. No additional arguments are used in `DynamicCache`. Return: A tuple containing the updated key and value states. """ # Update the cache if key_states is not None: if len(self.key_cache) <= layer_idx: # There may be skipped layers, fill them with empty lists for _ in range(len(self.key_cache), layer_idx): self.key_cache.append(torch.tensor([])) self.value_cache.append(torch.tensor([])) self.key_cache.append(key_states) self.value_cache.append(value_states) elif ( not self.key_cache[layer_idx].numel() # prefers not t.numel() to len(t) == 0 to export the model ): # fills previously skipped layers; checking for tensor causes errors self.key_cache[layer_idx] = key_states self.value_cache[layer_idx] = value_states else: self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=-2) self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=-2) return self.key_cache[layer_idx], self.value_cache[layer_idx] def reorder_cache(self, beam_idx: torch.LongTensor): """Reorders the cache for beam search, given the selected beam indices.""" for layer_idx in range(len(self.key_cache)): device = self.key_cache[layer_idx].device self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx.to(device)) device = self.value_cache[layer_idx].device self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx.to(device)) device = self.key_conv_cache[layer_idx].device self.key_conv_cache[layer_idx] = self.key_conv_cache[layer_idx].index_select(0, beam_idx.to(device)) device = self.value_conv_cache[layer_idx].device self.value_conv_cache[layer_idx] = self.value_conv_cache[layer_idx].index_select(0, beam_idx.to(device)) def get_seq_length(self, layer_idx: Optional[int] = 0) -> int: """Returns the sequence length of the cached states. A layer index can be optionally passed.""" # take any layer that contains cache and not empty tensor layer_idx = self.first_attention_layer if self.layer_types[layer_idx] != "full_attention" else layer_idx if len(self.key_cache) <= layer_idx or self.key_cache[layer_idx].numel() == 0: return 0 return self.key_cache[layer_idx].shape[-2] def get_mask_sizes(self, cache_position: torch.Tensor, layer_idx: int) -> tuple[int, int]: """ Return a tuple (kv_length, kv_offset) corresponding to the length and offset that will be returned for the given layer at `layer_idx`. The masks are then prepared according to the given lengths (kv_length, kv_offset) and patterns (i.e. sliding_window, chunk_size), for each layer. """ full_mask_kv_offset = 0 query_length = cache_position.shape[0] past_seen_tokens = self.get_seq_length() kv_length = query_length + past_seen_tokens return kv_length, full_mask_kv_offset def crop(self, max_length: int): """Crop the cache to the given length""" if max_length < 0: max_length = self.get_seq_length() - abs(max_length) if self.get_seq_length() <= max_length: return for idx in range(len(self.key_cache)): if self.key_cache[idx].numel(): self.key_cache[idx] = self.key_cache[idx][..., :max_length, :] self.value_cache[idx] = self.value_cache[idx][..., :max_length, :] def __len__(self) -> int: return len(self.key_cache) def __getitem__(self, layer_idx: int) -> tuple[torch.Tensor, torch.Tensor]: return self.key_cache[layer_idx], self.value_cache[layer_idx] def reset(self): for layer_idx in range(len(self.key_conv_cache)): self.key_conv_cache[layer_idx].zero_() self.value_conv_cache[layer_idx].zero_() class GearTextScaledWordEmbedding(nn.Embedding): """ This module overrides nn.Embeddings' forward by multiplying with embeddings scale. """ def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int, embed_scale: float = 1.0): super().__init__(num_embeddings, embedding_dim, padding_idx) self.register_buffer("embed_scale", torch.tensor(embed_scale), persistent=False) def forward(self, input_ids: torch.Tensor): return super().forward(input_ids) * self.embed_scale.to(self.weight.dtype) class GearRMSNorm(nn.Module): def __init__(self, dim: int, eps: float = 1e-6): super().__init__() self.eps = eps self.weight = nn.Parameter(torch.zeros(dim)) def _norm(self, x): return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) def forward(self, x): output = self._norm(x.float()) output = output * (1.0 + self.weight.float()) return output.type_as(x) def extra_repr(self): return f"{tuple(self.weight.shape)}, eps={self.eps}" class GearRMSNormGated(nn.Module): def __init__(self, config, hidden_size, eps=1e-6, **kwargs): super().__init__() self.weight = nn.Parameter(torch.ones(hidden_size)) self.eps = eps self.act_fn = ACT2FN[config.hidden_activation] def _norm(self, x): return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) def forward(self, x, gate=None): # Norm before gate output = self._norm(x.float()) output = output * (1.0 + self.weight.float()) output = output * self.act_fn(gate.float()) return output.type_as(x) class GearMLP(nn.Module): def __init__(self, config: GearConfig): 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_activation] def forward(self, x): down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) return down_proj class GearRotaryEmbedding(nn.Module): inv_freq: torch.Tensor # fix linting for `register_buffer` def __init__(self, config: GearConfig, 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) 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, 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. 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], dropout: float = 0.0, scaling: Optional[float] = None, softcap: Optional[float] = None, **kwargs, ) -> tuple[torch.Tensor, torch.Tensor]: if scaling is None: scaling = module.head_dim**-0.5 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 softcap is not None: attn_weights = attn_weights / softcap attn_weights = torch.tanh(attn_weights) attn_weights = attn_weights * softcap if attention_mask is not None: # no matter the length, we just slice it causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] attn_weights = attn_weights + causal_mask # upcast attention to fp32 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 GearAttention(nn.Module): """Multi-headed attention from 'Attention Is All You Need' paper""" def __init__(self, config: GearConfig, layer_idx: int): super().__init__() self.is_sliding = config.layer_types[layer_idx] == "sliding_attention" 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 = config.query_pre_attn_scalar**-0.5 self.attention_dropout = self.config.attention_dropout self.is_causal = not self.config.use_bidirectional_attention self.q_proj = nn.Linear( config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias ) self.k_proj = nn.Linear( config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias ) self.v_proj = nn.Linear( config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias ) self.o_proj = nn.Linear( config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias ) self.attn_logit_softcapping = self.config.attn_logit_softcapping self.sliding_window = config.sliding_window if self.is_sliding else None self.q_norm = GearRMSNorm(dim=config.head_dim, eps=config.rms_norm_eps) self.k_norm = GearRMSNorm(dim=config.head_dim, eps=config.rms_norm_eps) def forward( self, hidden_states: torch.Tensor, position_embeddings: torch.Tensor, attention_mask: Optional[torch.Tensor], past_key_values: Optional[Cache] = None, cache_position: Optional[torch.LongTensor] = None, **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) query_states = self.q_norm(query_states) key_states = self.k_norm(key_states) cos, sin = position_embeddings query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) if past_key_values 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_values.update(key_states, value_states, 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=self.attention_dropout if self.training else 0.0, scaling=self.scaling, sliding_window=self.sliding_window, **kwargs, ) attn_output = attn_output.reshape(*input_shape, -1).contiguous() attn_output = self.o_proj(attn_output) return attn_output, attn_weights class GearConvKVGatedMixer(nn.Module): """Convolutional Key-Value Gated Mixer is using Key-Value Convolution State and Sigmoid Gating""" def __init__( self, config: GearConfig, layer_idx: int, ): super().__init__() self.config = config self.layer_idx = layer_idx self.hidden_size = config.hidden_size 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.L_cache = config.conv_L_cache self.bias = config.attention_bias # self.adaptive_scaling = nn.Parameter(torch.tensor(1.0)) self.key_conv = nn.Conv1d( in_channels=config.num_key_value_heads * self.head_dim, out_channels=config.num_key_value_heads * self.head_dim, kernel_size=self.L_cache, groups=config.num_key_value_heads * self.head_dim, bias=self.bias, padding=self.L_cache - 1, ) self.value_conv = nn.Conv1d( in_channels=config.num_key_value_heads * self.head_dim, out_channels=config.num_key_value_heads * self.head_dim, kernel_size=self.L_cache, groups=config.num_key_value_heads * self.head_dim, bias=self.bias, padding=self.L_cache - 1, ) self.q_proj = nn.Linear( config.hidden_size, config.num_attention_heads * self.head_dim, bias=self.bias ) self.k_proj = nn.Linear( config.hidden_size, config.num_key_value_heads * self.head_dim, bias=self.bias ) self.v_proj = nn.Linear( config.hidden_size, config.num_key_value_heads * self.head_dim, bias=self.bias ) self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=self.bias) self.q_norm = GearRMSNorm(dim=self.head_dim, eps=config.rms_norm_eps) self.k_norm = GearRMSNorm(dim=self.head_dim, eps=config.rms_norm_eps) def apply_mask_to_padding_states(self, hidden_states, attention_mask): """ Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66 """ # NOTE: attention mask is a 2D boolean tensor if hidden_states.shape[1] == 1: return hidden_states if attention_mask is not None and attention_mask.dim() == 2 and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1: dtype = hidden_states.dtype hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype) return hidden_states def expand_kv(self, kv_states): # kv_states: [B, S, kv_heads * head_dim] B, S, _ = kv_states.shape kv_states = kv_states.view(B, S, self.config.num_key_value_heads, self.head_dim) kv_states = kv_states[:, :, :, None, :].expand( B, S, self.config.num_key_value_heads, self.num_key_value_groups, self.head_dim, ) return kv_states.reshape(B, S, self.config.num_attention_heads * self.head_dim) def cuda_kernels_forward( self, hidden_states: torch.Tensor, past_key_values: GearHybridKVConvCache | None = None, cache_position: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, ): input_shape = hidden_states.shape[:-1] hidden_shape = (*input_shape, -1, self.head_dim) hidden_states = self.apply_mask_to_padding_states(hidden_states, attention_mask) query = self.q_proj(hidden_states).view(hidden_shape) key = self.k_proj(hidden_states).view(hidden_shape) value = self.v_proj(hidden_states).transpose(-1, -2) query = self.q_norm(query).reshape(*input_shape, -1) key = self.k_norm(key).reshape(*input_shape, -1).transpose(-1, -2) key_conv_weights = self.key_conv.weight.view(self.key_conv.weight.size(0), self.key_conv.weight.size(2)) value_conv_weights = self.value_conv.weight.view(self.value_conv.weight.size(0), self.value_conv.weight.size(2)) if past_key_values is not None and cache_position[0] > 0: key = causal_conv1d_update( key.squeeze(-1), past_key_values.key_conv_cache[self.layer_idx], key_conv_weights, self.key_conv.bias, None, ) value = causal_conv1d_update( value.squeeze(-1), past_key_values.value_conv_cache[self.layer_idx], value_conv_weights, self.value_conv.bias, None, ) key, value = key.unsqueeze(-1), value.unsqueeze(-1) else: if past_key_values is not None: key_conv_state = nn.functional.pad(key, (self.L_cache - key.shape[-1], 0)) past_key_values.key_conv_cache[self.layer_idx].copy_(key_conv_state) value_conv_state = nn.functional.pad(value, (self.L_cache - value.shape[-1], 0)) past_key_values.value_conv_cache[self.layer_idx].copy_(value_conv_state) key = causal_conv1d_fn(key, key_conv_weights, self.key_conv.bias, activation=None) value = causal_conv1d_fn(value, value_conv_weights, self.value_conv.bias, activation=None) key = self.expand_kv(key.transpose(-1, -2)) value = self.expand_kv(value.transpose(-1, -2)) mixer_weights = torch.sigmoid(query * key) core_mixer_out = mixer_weights * value core_mixer_out = self.o_proj(core_mixer_out) return core_mixer_out def torch_native_forward( self, hidden_states: torch.Tensor, past_key_values: GearHybridKVConvCache | None = None, cache_position: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, ): seqlen = hidden_states.shape[1] input_shape = hidden_states.shape[:-1] hidden_shape = (*input_shape, -1, self.head_dim) hidden_states = self.apply_mask_to_padding_states(hidden_states, attention_mask) query = self.q_proj(hidden_states).view(hidden_shape) key = self.k_proj(hidden_states).view(hidden_shape) value = self.v_proj(hidden_states).transpose(-1, -2) query = self.q_norm(query).reshape(*input_shape, -1) key = self.k_norm(key).reshape(*input_shape, -1).transpose(-1, -2) # NOTE: This native path is kept numerically close to the # causal_conv1d fast path: # - the new token always goes to the *last* cache slot (matching # causal_conv1d_update semantics) instead of the previous # clamped-cache_position behavior that misplaced the first # decode steps. # - depthwise reductions are accumulated in fp32 then cast back, # matching the CUDA kernel which also accumulates in fp32. if past_key_values is not None and cache_position[0] > 0: key_conv_state = past_key_values.key_conv_cache[self.layer_idx] key_conv_state = key_conv_state.roll(shifts=-1, dims=-1) key_conv_state[:, :, -1] = key.squeeze(-1).to( device=key_conv_state.device, dtype=key_conv_state.dtype ) past_key_values.key_conv_cache[self.layer_idx].copy_(key_conv_state) key_fp32 = ( key_conv_state.to(key.device, dtype=torch.float32) * self.key_conv.weight[:, 0, :].to(torch.float32) ).sum(dim=-1) key = key_fp32.to(self.k_proj.weight.dtype) value_conv_state = past_key_values.value_conv_cache[self.layer_idx] value_conv_state = value_conv_state.roll(shifts=-1, dims=-1) value_conv_state[:, :, -1] = value.squeeze(-1).to( device=value_conv_state.device, dtype=value_conv_state.dtype ) past_key_values.value_conv_cache[self.layer_idx].copy_(value_conv_state) value_fp32 = ( value_conv_state.to(value.device, dtype=torch.float32) * self.value_conv.weight[:, 0, :].to(torch.float32) ).sum(dim=-1) value = value_fp32.to(self.v_proj.weight.dtype) if self.bias and self.key_conv.bias is not None: key = key + self.key_conv.bias if self.bias and self.value_conv.bias is not None: value = value + self.value_conv.bias key, value = key.unsqueeze(-1), value.unsqueeze(-1) else: if past_key_values is not None: key_conv_state = nn.functional.pad(key, (self.L_cache - key.shape[-1], 0)) past_key_values.key_conv_cache[self.layer_idx].copy_(key_conv_state) value_conv_state = nn.functional.pad(value, (self.L_cache - value.shape[-1], 0)) past_key_values.value_conv_cache[self.layer_idx].copy_(value_conv_state) # Run conv in fp32 to match the CUDA kernel's accumulation # precision; cast result back to the parameter dtype. orig_dtype = self.k_proj.weight.dtype key_fp = nn.functional.conv1d( key.to(torch.float32), self.key_conv.weight.to(torch.float32), bias=(self.key_conv.bias.to(torch.float32) if self.key_conv.bias is not None else None), stride=self.key_conv.stride, padding=self.key_conv.padding, dilation=self.key_conv.dilation, groups=self.key_conv.groups, ) value_fp = nn.functional.conv1d( value.to(torch.float32), self.value_conv.weight.to(torch.float32), bias=(self.value_conv.bias.to(torch.float32) if self.value_conv.bias is not None else None), stride=self.value_conv.stride, padding=self.value_conv.padding, dilation=self.value_conv.dilation, groups=self.value_conv.groups, ) key = key_fp[..., :seqlen].to(orig_dtype) value = value_fp[..., :seqlen].to(orig_dtype) key = self.expand_kv(key.transpose(-1, -2)) value = self.expand_kv(value.transpose(-1, -2)) mixer_weights = torch.sigmoid(query * key) core_mixer_out = mixer_weights * value core_mixer_out = self.o_proj(core_mixer_out) return core_mixer_out def forward( self, hidden_states: torch.Tensor, past_key_values: Cache | None = None, cache_position: torch.LongTensor = None, attention_mask: torch.Tensor | None = None, ): if is_fast_path_available: return self.cuda_kernels_forward(hidden_states, past_key_values, cache_position, attention_mask) return self.torch_native_forward(hidden_states, past_key_values, cache_position, attention_mask) class GearDecoderLayer(GradientCheckpointingLayer): def __init__(self, config: GearConfig, layer_idx: int): super().__init__() self.config = config self.hidden_size = config.hidden_size self.layer_idx = layer_idx self.attention_type = config.layer_types[layer_idx] if self.attention_type == "conv_mixer": self.local_mixer = GearConvKVGatedMixer(config=config, layer_idx=layer_idx) else: self.self_attn = GearAttention(config=config, layer_idx=layer_idx) self.mlp = GearMLP(config) self.input_layernorm = GearRMSNorm(self.hidden_size, eps=config.rms_norm_eps) self.post_attention_layernorm = GearRMSNorm(self.hidden_size, eps=config.rms_norm_eps) self.pre_feedforward_layernorm = GearRMSNorm(self.hidden_size, eps=config.rms_norm_eps) self.post_feedforward_layernorm = GearRMSNorm(self.hidden_size, eps=config.rms_norm_eps) def forward( self, hidden_states: torch.Tensor, position_embeddings_global: torch.Tensor, position_embeddings_local: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Cache] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, cache_position: Optional[torch.LongTensor] = None, **kwargs, ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]: residual = hidden_states hidden_states = self.input_layernorm(hidden_states) if self.attention_type == "conv_mixer": hidden_states = self.local_mixer( hidden_states=hidden_states, past_key_values=past_key_values, cache_position=cache_position, attention_mask=attention_mask, ) else: # apply global RoPE to non-sliding layer only if self.self_attn.is_sliding: position_embeddings = position_embeddings_local else: position_embeddings = position_embeddings_global hidden_states, self_attn_weights = self.self_attn( hidden_states=hidden_states, position_embeddings=position_embeddings, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, **kwargs, ) hidden_states = self.post_attention_layernorm(hidden_states) hidden_states = residual + hidden_states residual = hidden_states hidden_states = self.pre_feedforward_layernorm(hidden_states) hidden_states = self.mlp(hidden_states) hidden_states = self.post_feedforward_layernorm(hidden_states) hidden_states = residual + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) return outputs @auto_docstring class GearPreTrainedModel(PreTrainedModel): config: GearConfig base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = ["GearDecoderLayer"] _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": GearDecoderLayer, "attentions": GearAttention, "conv_mixer": GearConvKVGatedMixer, } @auto_docstring class GearModel(GearPreTrainedModel): config: GearConfig def __init__(self, config: GearConfig): super().__init__(config) self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.embed_tokens = GearTextScaledWordEmbedding( config.vocab_size, config.hidden_size, self.padding_idx, embed_scale=self.config.hidden_size**0.5 ) self.layers = nn.ModuleList( [GearDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] ) self.norm = GearRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.rotary_emb = GearRotaryEmbedding(config=config) self.gradient_checkpointing = False config = copy.deepcopy(config) config.rope_theta = config.rope_local_base_freq config.rope_scaling = {"rope_type": "default"} self.rotary_emb_local = GearRotaryEmbedding(config=config) # 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, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, **kwargs: Unpack[TransformersKwargs], ) -> BaseModelOutputWithPast: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) use_cache = use_cache if use_cache is not None else self.config.use_cache if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") if self.gradient_checkpointing and self.training and use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`." ) use_cache = False if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) # if use_cache and past_key_values is None and not self.training: # batch_size = inputs_embeds.shape[0] # past_key_values = GearHybridKVConvCache(config=self.config, max_batch_size=batch_size, dtype=self.dtype, device=self.device) if use_cache and not self.training: if past_key_values is None or type(past_key_values).__name__ == "DynamicCache": batch_size = inputs_embeds.shape[0] past_key_values = GearHybridKVConvCache( config=self.config, max_batch_size=batch_size, dtype=self.dtype, device=self.device ) 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, } sliding_mask_kwargs = mask_kwargs.copy() if self.config.use_bidirectional_attention: mask_kwargs["or_mask_function"] = lambda *args: torch.tensor(True, dtype=torch.bool) sliding_mask_kwargs["or_mask_function"] = _bidirectional_window_overlay(self.config.sliding_window) # Create the masks causal_mask_mapping = { "full_attention": create_causal_mask(**mask_kwargs), "sliding_attention": create_sliding_window_causal_mask(**sliding_mask_kwargs), "conv_mixer": attention_mask if inputs_embeds.shape[-1] != 1 else None, } # embed positions hidden_states = inputs_embeds # create position embeddings to be shared across the decoder layers position_embeddings_global = self.rotary_emb(hidden_states, position_ids) position_embeddings_local = self.rotary_emb_local(hidden_states, position_ids) # decoder layers all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None for decoder_layer in self.layers[: self.config.num_hidden_layers]: if output_hidden_states: all_hidden_states += (hidden_states,) layer_outputs = decoder_layer( hidden_states, position_embeddings_global=position_embeddings_global, position_embeddings_local=position_embeddings_local, attention_mask=causal_mask_mapping[decoder_layer.attention_type], position_ids=position_ids, past_key_values=past_key_values, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, **kwargs, ) hidden_states = layer_outputs[0] if output_attentions: all_self_attns += (layer_outputs[1],) hidden_states = self.norm(hidden_states) if output_hidden_states: all_hidden_states += (hidden_states,) return BaseModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=past_key_values, hidden_states=all_hidden_states, attentions=all_self_attns, ) @auto_docstring class GearForCausalLM(GearPreTrainedModel, GenerationMixin): _tied_weights_keys = ["lm_head.weight"] _tp_plan = {"lm_head": "colwise_rep"} _pp_plan = {"lm_head": (["hidden_states"], ["logits"])} config: GearConfig def __init__(self, config: GearConfig): super().__init__(config) self.model = GearModel(config) self.vocab_size = config.vocab_size self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) # Initialize weights and apply final processing self.post_init() @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, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, logits_to_keep: Union[int, torch.Tensor] = 0, **kwargs, ) -> CausalLMOutputWithPast: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) outputs: BaseModelOutputWithPast = 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, output_attentions=output_attentions, output_hidden_states=output_hidden_states, cache_position=cache_position, **kwargs, ) hidden_states = outputs.last_hidden_state # Only compute necessary logits, and do not upcast them to float if we are not computing the loss slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep logits = self.lm_head(hidden_states[:, slice_indices, :]) # if self.config.final_logit_softcapping is not None: # logits = logits / self.config.final_logit_softcapping # logits = torch.tanh(logits) # logits = logits * self.config.final_logit_softcapping loss = None if labels is not None: loss = self.loss_function(logits, labels, self.vocab_size, **kwargs) return CausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) class GearForSequenceClassification(GenericForSequenceClassification, GearPreTrainedModel): pass class GearForTokenClassification(GenericForTokenClassification, GearPreTrainedModel): pass class GearForQuestionAnswering(GenericForQuestionAnswering, GearPreTrainedModel): base_model_prefix = "transformer" # For BC, where `transformer` was used instead of `model` __all__ = [ "GearForCausalLM", "GearForQuestionAnswering", "GearPreTrainedModel", "GearModel", "GearForSequenceClassification", "GearForTokenClassification", ]