import inspect import math import warnings from dataclasses import dataclass, field from typing import Any, Dict, List, Optional, Tuple, Union import numpy as np import torch import torch.nn.functional as F import torch.utils.checkpoint from torch import nn from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss from transformers.activations import ACT2FN from transformers.cache_utils import Cache, DynamicCache from transformers.modeling_attn_mask_utils import ( _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa, ) from transformers.modeling_outputs import ( MoeCausalLMOutputWithPast, MoeModelOutputWithPast, SequenceClassifierOutputWithPast, ) from transformers.modeling_utils import PreTrainedModel from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_13 from transformers.utils import ( add_start_docstrings, add_start_docstrings_to_model_forward, is_flash_attn_greater_or_equal_2_10, logging, replace_return_docstrings, ) from transformers.utils.import_utils import is_torch_fx_available from .configuration_hymba import HymbaConfig from torch.utils.checkpoint import checkpoint from flash_attn import flash_attn_func, flash_attn_varlen_func from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters) from einops import rearrange, repeat, reduce, pack, unpack from einops.layers.torch import Rearrange if is_torch_fx_available(): if not is_torch_greater_or_equal_than_1_13: import torch.fx _prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask) from mamba_ssm.ops.selective_scan_interface import mamba_inner_fn, selective_scan_fn from mamba_ssm.ops.triton.selective_state_update import selective_state_update from causal_conv1d import causal_conv1d_fn, causal_conv1d_update is_fast_path_available = all( (selective_state_update, selective_scan_fn, causal_conv1d_fn, causal_conv1d_update, mamba_inner_fn) ) logger = logging.get_logger(__name__) _CONFIG_FOR_DOC = "HymbaConfig" def pad_at_dim(t, pad: Tuple[int, int], dim = -1, value = 0.): if pad == (0, 0): return t dims_from_right = (- dim - 1) if dim < 0 else (t.ndim - dim - 1) zeros = ((0, 0) * dims_from_right) return F.pad(t, (*zeros, *pad), value = value) # Adapted from transformers.models.mixtral.modeling_mixtral.load_balancing_loss_func def load_balancing_loss_func( gate_logits: torch.Tensor, num_experts: torch.Tensor = None, top_k=2, attention_mask: Optional[torch.Tensor] = None ) -> float: r""" Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch. See Switch Transformer (https://arxiv.org/abs/2101.03961) for more details. This function implements the loss function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between experts is too unbalanced. Args: gate_logits (Union[`torch.Tensor`, Tuple[torch.Tensor]): Logits from the `router`, should be a tuple of model.config.num_hidden_layers tensors of shape [batch_size X sequence_length, num_experts]. attention_mask (`torch.Tensor`, None): The attention_mask used in forward function shape [batch_size X sequence_length] if not None. num_experts (`int`, *optional*): Number of experts Returns: The auxiliary loss. """ if gate_logits is None or not isinstance(gate_logits, tuple): return 0 if isinstance(gate_logits, tuple): compute_device = gate_logits[0].device concatenated_gate_logits = torch.cat( [layer_gate.to(compute_device) for layer_gate in gate_logits if layer_gate.shape[1] > 1], dim=0 ) routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1) _, selected_experts = torch.topk(routing_weights, top_k, dim=-1) expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts) if attention_mask is None: # Compute the percentage of tokens routed to each experts tokens_per_expert = torch.mean(expert_mask.float(), dim=0) # Compute the average probability of routing to these experts router_prob_per_expert = torch.mean(routing_weights, dim=0) else: batch_size, sequence_length = attention_mask.shape num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length) # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask expert_attention_mask = ( attention_mask[None, :, :, None, None] .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts)) .reshape(-1, top_k, num_experts) .to(compute_device) ) # Compute the percentage of tokens routed to each experts tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum( expert_attention_mask, dim=0 ) # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert router_per_expert_attention_mask = ( attention_mask[None, :, :, None] .expand((num_hidden_layers, batch_size, sequence_length, num_experts)) .reshape(-1, num_experts) .to(compute_device) ) # Compute the average probability of routing to these experts router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum( router_per_expert_attention_mask, dim=0 ) overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0)) return overall_loss * num_experts # Copied from transformers.models.llama.modeling_llama._get_unpad_data def _get_unpad_data(attention_mask): seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32) indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten() max_seqlen_in_batch = seqlens_in_batch.max().item() cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0)) return ( indices, cu_seqlens, max_seqlen_in_batch, ) class HymbaRMSNorm(nn.Module): def __init__(self, hidden_size, eps=1e-6): """ HymbaRMSNorm 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) class PerheadHymbaRMSNorm(nn.Module): def __init__(self, hidden_size, num_heads, eps=1e-6): """ For per-head kq normalization """ super().__init__() self.weight = nn.Parameter(torch.ones(1, num_heads, 1, hidden_size)) self.variance_epsilon = eps def forward(self, hidden_states): # assert 1==0, f"hiddens_states shape: {hidden_states.shape}" # [bsz, num_heads, seq_len, head_dim] assert hidden_states.shape[1] == self.weight.shape[1], f"hidden_state: {hidden_states.shape}, weight: {self.weight.shape}" assert hidden_states.shape[3] == self.weight.shape[3], f"hidden_state: {hidden_states.shape}, weight: {self.weight.shape}" 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) 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) class HymbaOnlyNorm(nn.Module): def __init__(self, hidden_size, eps=1e-6): """ HymbaRMSNorm 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 hidden_states.to(input_dtype) class LlamaRotaryEmbedding(nn.Module): def __init__(self, config, dim, base=10000, device=None, scaling_factor=1.0): super().__init__() self.scaling_factor = scaling_factor self.dim = dim self.base = base self.config = config self.rope_type = config.rope_type self.factor = 2 max_position_embeddings = self.config.max_position_embeddings if config.rope_type is None or config.rope_type == "default": inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim)) self.max_seq_len_cached = max_position_embeddings elif config.rope_type == 'ntk': assert self.config.orig_max_position_embeddings is not None orig_max_position_embeddings = self.config.orig_max_position_embeddings base = base * ((self.factor * max_position_embeddings / orig_max_position_embeddings) - (self.factor - 1)) ** (self.dim / (self.dim - 2)) inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim)) self.max_seq_len_cached = orig_max_position_embeddings elif config.rope_type == 'dynamic_ntk': inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim)) self.original_inv_freq = inv_freq self.max_seq_len_cached = self.config.orig_max_position_embeddings else: raise ValueError(f"Not support rope_type: {config.rope_type}") self.register_buffer("inv_freq", inv_freq, persistent=False) def _dynamic_frequency_update(self, position_ids, device): """ dynamic RoPE layers should recompute `inv_freq` in the following situations: 1 - growing beyond the cached sequence length (allow scaling) 2 - the current sequence length is in the original scale (avoid losing precision with small sequences) """ seq_len = torch.max(position_ids) + 1 if seq_len > self.max_seq_len_cached: # growth base = self.base * ((self.factor * seq_len / self.config.orig_max_position_embeddings) - (self.factor - 1)) ** (self.dim / (self.dim - 2)) inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim)) self.register_buffer("inv_freq", inv_freq, persistent=False) self.max_seq_len_cached = seq_len if seq_len < self.config.orig_max_position_embeddings and self.max_seq_len_cached > self.config.orig_max_position_embeddings: # reset self.register_buffer("inv_freq", self.original_inv_freq, persistent=False) self.max_seq_len_cached = self.config.orig_max_position_embeddings @torch.no_grad() def forward(self, x, position_ids): if self.rope_type == 'dynamic_ntk': self._dynamic_frequency_update(position_ids, device=x.device) # x: [bs, num_attention_heads, seq_len, head_size] inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) position_ids_expanded = position_ids[:, None, :].float() # Force float32 since bfloat16 loses precision on long contexts # See https://github.com/huggingface/transformers/pull/29285 device_type = x.device.type device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu" with torch.autocast(device_type=device_type, enabled=False): freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) emb = torch.cat((freqs, freqs), dim=-1) cos = emb.cos() sin = emb.sin() 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, 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) if q is not None: q_embed = (q * cos) + (rotate_half(q) * sin) else: q_embed = None if k is not None: k_embed = (k * cos) + (rotate_half(k) * sin) else: k_embed = None return q_embed, k_embed # Copied from transformers.models.llama.modeling_llama.repeat_kv 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) class HybridMambaAttentionDynamicCache(DynamicCache): """ A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache (which has a constant shape regardless of seq_len). This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states` and `ssm_states` for mamba cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`, while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors). For mamba layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors), while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`, and `ssm_states` represents the ssm state and has a shape of `(batch_size, d_inner, d_state)`. """ def __init__(self, config, batch_size, dtype=torch.float16, device=None, layer_type=None): self.dtype = dtype # self.layers_block_type = config.layers_block_type self.has_previous_state = False # only used by mamba intermediate_size = config.mamba_expand * config.hidden_size ssm_state_size = config.mamba_d_state conv_kernel_size = config.mamba_d_conv self.conv_states = [] self.ssm_states = [] self.layer_type = layer_type for i in range(config.num_hidden_layers): if layer_type is None: has_mamba_state = True else: has_mamba_state = self.layer_type[i] == 'h' or self.layer_type[i] == 'm' if has_mamba_state: if hasattr(config, 'conv_dim'): conv_dim = config.conv_dim[i] else: conv_dim = intermediate_size self.conv_states += [ torch.zeros(batch_size, conv_dim, conv_kernel_size, device=device, dtype=dtype) ] self.ssm_states += [ torch.zeros(batch_size, intermediate_size, ssm_state_size, device=device, dtype=dtype) ] else: self.conv_states += [torch.tensor([[]] * batch_size, device=device)] self.ssm_states += [torch.tensor([[]] * batch_size, device=device)] self.key_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)] self.value_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)] self.mamba_past_length = [0 for _ in range(config.num_hidden_layers)] 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]: # Update the cache if self.key_cache[layer_idx].shape[-1] == 0: 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.conv_states[layer_idx].device self.conv_states[layer_idx] = self.conv_states[layer_idx].index_select(0, beam_idx.to(device)) device = self.ssm_states[layer_idx].device self.ssm_states[layer_idx] = self.ssm_states[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 if self.layer_type[layer_idx] == 'm': return self.mamba_past_length[layer_idx] if self.key_cache[layer_idx].shape[-1] == 0: return 0 return self.key_cache[layer_idx].shape[-2] def to_legacy_cache(self) -> Tuple[Tuple[torch.Tensor], Tuple[torch.Tensor]]: raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.") @classmethod def from_legacy_cache(cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None) -> "DynamicCache": raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.") @dataclass class MambaCacheParams: seqlen_offset: int = 0 conv_states: Dict[int, torch.Tensor] = field(default_factory=dict) ssm_states: Dict[int, torch.Tensor] = field(default_factory=dict) # Adapted from transformers.models.mistral.modeling_mistral.MistralAttention with Mistral->Hymba class HymbaAttention(nn.Module): """ Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer and "Generating Long Sequences with Sparse Transformers". """ def __init__(self, config: HymbaConfig, layer_idx: Optional[int] = None, reuse_kv=False, output_hidden_size=None, attn_only_wo_proj=False): super().__init__() self.config = config self.layer_idx = layer_idx if layer_idx is None: logger.warning_once( f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will " "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` " "when creating this class." ) # self.hidden_size = config.hidden_size self.hidden_size = config.attn_hidden_size if config.attn_hidden_size > 0 else config.hidden_size self.num_heads = config.num_attention_heads self.head_dim = self.hidden_size // self.num_heads self.max_position_embeddings = config.max_position_embeddings self.rope_theta = config.rope_theta self.attn_only_wo_proj = attn_only_wo_proj self.kq_head_dim = config.kq_head_dim if config.kq_head_dim > 0 else self.head_dim self.v_head_dim = config.v_head_dim if config.v_head_dim > 0 else self.head_dim self.num_key_value_heads = config.num_key_value_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.is_causal = True self.attention_dropout = config.attention_dropout if (self.head_dim * self.num_heads) != self.hidden_size: raise ValueError( f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}" f" and `num_heads`: {self.num_heads})." ) if not self.attn_only_wo_proj: self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.kq_head_dim, bias=False) self.reuse_kv = reuse_kv if not self.attn_only_wo_proj and not self.reuse_kv: self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.kq_head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.v_head_dim, bias=False) if output_hidden_size is None: output_hidden_size = self.hidden_size if not self.attn_only_wo_proj: self.o_proj = nn.Linear(self.num_heads * self.v_head_dim, output_hidden_size, bias=False) if self.config.kq_norm == "rms": self.k_norm = HymbaRMSNorm(self.kq_head_dim) self.q_norm = HymbaRMSNorm(self.kq_head_dim) elif self.config.kq_norm == "perhead-rms": self.k_norm = PerheadHymbaRMSNorm(self.kq_head_dim, self.num_key_value_heads) self.q_norm = PerheadHymbaRMSNorm(self.kq_head_dim, self.num_heads) elif self.config.kq_norm == "none": self.k_norm = None self.q_norm = None else: raise NotImplementedError(f"Unknown kq_norm: {self.config.kq_norm}") if self.config.rope: self._init_rope() def set_rope(self, rope_type, orig_max_position_embeddings, max_position_embeddings): self.config.rope_type = rope_type self.config.orig_max_position_embeddings = orig_max_position_embeddings self.config.max_position_embeddings = max_position_embeddings self._init_rope() def _init_rope(self): self.rotary_emb = LlamaRotaryEmbedding( config=self.config, dim=self.kq_head_dim, base=self.rope_theta, device=torch.device("cuda"), ) def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int): return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous() 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, output_attentions: bool = False, use_cache: bool = False, kv_last_layer = None, # kv_proj_last_layer = None, use_swa=False, query_states = None, key_states=None, value_states=None, **kwargs, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: raise NotImplementedError("HymbaAttention is an abstract class. Use one of the subclasses.") # Adapted from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 with Mistral->Hymba class HymbaFlashAttention2(HymbaAttention): """ Hymba flash attention module. This module inherits from `HymbaAttention` as the weights of the module stays untouched. The only required change would be on the forward pass where it needs to correctly call the public API of flash attention and deal with padding tokens in case the input contains any of them. """ # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__ def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10() def forward( self, hidden_states: torch.Tensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, kv_last_layer=None, # kv_proj_last_layer = None, use_swa=False, query_states = None, key_states=None, value_states=None, **kwargs, ): if "padding_mask" in kwargs: warnings.warn( "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`" ) # overwrite attention_mask with padding_mask attention_mask = kwargs.pop("padding_mask") if self.attn_only_wo_proj: assert query_states is not None bsz, q_len, _ = query_states.size() else: bsz, q_len, _ = hidden_states.size() if not self.attn_only_wo_proj: query_states = self.q_proj(hidden_states) query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous() if self.q_norm is not None: query_states = self.q_norm(query_states) if self.config.rope: if self.attn_only_wo_proj: cos, sin = self.rotary_emb(query_states, position_ids) else: cos, sin = self.rotary_emb(hidden_states, position_ids) query_states, _ = apply_rotary_pos_emb(query_states, None, cos, sin) if self.reuse_kv: assert kv_last_layer is not None key_states, value_states = kv_last_layer # (batch, num_heads, slen, head_dim) else: if not self.attn_only_wo_proj: key_states = self.k_proj(hidden_states) value_states = self.v_proj(hidden_states) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.kq_head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.v_head_dim).transpose(1, 2) if self.k_norm is not None: key_states = self.k_norm(key_states) if self.config.rope: _, key_states = apply_rotary_pos_emb(None, key_states, cos, sin) kv_seq_len = key_states.shape[-2] if past_key_value is not None and not self.reuse_kv: if self.layer_idx is None: raise ValueError( f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} " "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class " "with a layer index." ) kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) use_sliding_windows = ( _flash_supports_window_size and getattr(self.config, "sliding_window", None) is not None and kv_seq_len > (self.config.sliding_window + self.config.num_memory_tokens if self.config.num_memory_tokens > 0 else self.config.sliding_window) and use_swa ) if not _flash_supports_window_size: logger.warning_once( "The current flash attention version does not support sliding window attention, for a more memory efficient implementation" " make sure to upgrade flash-attn library." ) swa_processed_flag = False if past_key_value is not None and use_cache and not self.reuse_kv: kv_layer_idx = self.layer_idx cache_has_contents = past_key_value.get_seq_length(kv_layer_idx) > 0 if ( getattr(self.config, "sliding_window", None) is not None and kv_seq_len > (self.config.sliding_window + self.config.num_memory_tokens if self.config.num_memory_tokens > 0 else self.config.sliding_window) and cache_has_contents and use_swa ): slicing_tokens = 1 - self.config.sliding_window past_key = past_key_value[kv_layer_idx][0] past_value = past_key_value[kv_layer_idx][1] if self.config.num_memory_tokens > 0: # num_fetched_memory_tokens = min(kv_seq_len - self.config.sliding_window, self.config.num_memory_tokens) num_fetched_memory_tokens = self.config.num_memory_tokens past_key = torch.cat([past_key[:, :, :num_fetched_memory_tokens, :], past_key[:, :, slicing_tokens:, :]], dim=-2).contiguous() past_value = torch.cat([past_value[:, :, :num_fetched_memory_tokens, :], past_value[:, :, slicing_tokens:, :]], dim=-2).contiguous() else: past_key = past_key[:, :, slicing_tokens:, :].contiguous() past_value = past_value[:, :, slicing_tokens:, :].contiguous() past_key_value.key_cache[kv_layer_idx] = past_key past_key_value.value_cache[kv_layer_idx] = past_value if attention_mask is not None: attention_mask = attention_mask[:, slicing_tokens:] attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1) swa_processed_flag = True key_states, value_states = past_key_value.update(key_states, value_states, kv_layer_idx) # repeat k/v heads if n_kv_heads < n_heads key_states_no_repeat = key_states value_states_no_repeat = value_states key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) dropout_rate = 0.0 if not self.training else self.attention_dropout input_dtype = query_states.dtype if input_dtype == torch.float32: if torch.is_autocast_enabled(): target_dtype = torch.get_autocast_gpu_dtype() # Handle the case where the model is quantized elif hasattr(self.config, "_pre_quantization_dtype"): target_dtype = self.config._pre_quantization_dtype else: target_dtype = self.q_proj.weight.dtype logger.warning_once( f"The input hidden states seems to be silently casted in float32, this might be related to" f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in" f" {target_dtype}." ) query_states = query_states.to(target_dtype) key_states = key_states.to(target_dtype) value_states = value_states.to(target_dtype) # Reashape to the expected shape for Flash Attention query_states = query_states.transpose(1, 2) # (batch, slen, num_heads, head_dim) key_states = key_states.transpose(1, 2) # (batch, slen, num_heads, head_dim) value_states = value_states.transpose(1, 2) # (batch, slen, num_heads, head_dim) attn_output = self._flash_attention_forward( query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate, use_sliding_windows=use_sliding_windows and not swa_processed_flag, ) v_dim = value_states.shape[-2] * value_states.shape[-1] attn_output = attn_output.reshape(bsz, q_len, v_dim).contiguous() if self.attn_only_wo_proj: return attn_output, (key_states_no_repeat, value_states_no_repeat) attn_output = self.o_proj(attn_output) if not output_attentions: attn_weights = None return attn_output, attn_weights, past_key_value, (key_states_no_repeat, value_states_no_repeat) def _flash_attention_forward( self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None, use_sliding_windows=False, ): """ Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token first unpad the input, then computes the attention scores and pad the final attention scores. Args: query_states (`torch.Tensor`): Input query states to be passed to Flash Attention API key_states (`torch.Tensor`): Input key states to be passed to Flash Attention API value_states (`torch.Tensor`): Input value states to be passed to Flash Attention API attention_mask (`torch.Tensor`): The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the position of padding tokens and 1 for the position of non-padding tokens. dropout (`int`, *optional*): Attention dropout softmax_scale (`float`, *optional*): The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim) use_sliding_windows (`bool`, *optional*): Whether to activate sliding window attention. """ if not self._flash_attn_uses_top_left_mask: causal = self.is_causal else: causal = self.is_causal and query_length != 1 # Contains at least one padding token in the sequence if attention_mask is not None: if value_states.shape[-1] == query_states.shape[-1] * 2: value_states1 = value_states[...,:query_states.shape[-1]] batch_size = query_states.shape[0] query_states1, key_states1, value_states1, indices_q, cu_seq_lens, max_seq_lens = self._upad_input( query_states, key_states, value_states1, attention_mask, query_length ) cu_seqlens_q, cu_seqlens_k = cu_seq_lens max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens if not use_sliding_windows: attn_output_unpad1 = flash_attn_varlen_func( query_states1, key_states1, value_states1, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=cu_seqlens_k, max_seqlen_q=max_seqlen_in_batch_q, max_seqlen_k=max_seqlen_in_batch_k, dropout_p=dropout, softmax_scale=softmax_scale, causal=causal, ) else: attn_output_unpad1 = flash_attn_varlen_func( query_states1, key_states1, value_states1, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=cu_seqlens_k, max_seqlen_q=max_seqlen_in_batch_q, max_seqlen_k=max_seqlen_in_batch_k, dropout_p=dropout, softmax_scale=softmax_scale, causal=causal, window_size=(self.config.sliding_window, self.config.sliding_window), ) attn_output1 = pad_input(attn_output_unpad1, indices_q, batch_size, query_length) value_states2 = value_states[...,query_states.shape[-1]:] query_states2, key_states2, value_states2, indices_q, cu_seq_lens, max_seq_lens = self._upad_input( query_states, key_states, value_states2, attention_mask, query_length ) cu_seqlens_q, cu_seqlens_k = cu_seq_lens max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens if not use_sliding_windows: attn_output_unpad2 = flash_attn_varlen_func( query_states2, key_states2, value_states2, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=cu_seqlens_k, max_seqlen_q=max_seqlen_in_batch_q, max_seqlen_k=max_seqlen_in_batch_k, dropout_p=dropout, softmax_scale=softmax_scale, causal=causal, ) else: attn_output_unpad2 = flash_attn_varlen_func( query_states2, key_states2, value_states2, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=cu_seqlens_k, max_seqlen_q=max_seqlen_in_batch_q, max_seqlen_k=max_seqlen_in_batch_k, dropout_p=dropout, softmax_scale=softmax_scale, causal=causal, window_size=(self.config.sliding_window, self.config.sliding_window), ) attn_output2 = pad_input(attn_output_unpad2, indices_q, batch_size, query_length) attn_output = torch.cat([attn_output1, attn_output2], dim=-1) else: batch_size = query_states.shape[0] query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input( query_states, key_states, value_states, attention_mask, query_length ) cu_seqlens_q, cu_seqlens_k = cu_seq_lens max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens if not use_sliding_windows: attn_output_unpad = flash_attn_varlen_func( query_states, key_states, value_states, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=cu_seqlens_k, max_seqlen_q=max_seqlen_in_batch_q, max_seqlen_k=max_seqlen_in_batch_k, dropout_p=dropout, softmax_scale=softmax_scale, causal=causal, ) else: attn_output_unpad = flash_attn_varlen_func( query_states, key_states, value_states, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=cu_seqlens_k, max_seqlen_q=max_seqlen_in_batch_q, max_seqlen_k=max_seqlen_in_batch_k, dropout_p=dropout, softmax_scale=softmax_scale, causal=causal, window_size=(self.config.sliding_window, self.config.sliding_window), ) attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length) else: if value_states.shape[-1] == query_states.shape[-1] * 2: if not use_sliding_windows: attn_output1 = flash_attn_func( query_states, key_states, value_states[...,:query_states.shape[-1]], dropout, softmax_scale=softmax_scale, causal=causal, ) attn_output2 = flash_attn_func( query_states, key_states, value_states[...,query_states.shape[-1]:], dropout, softmax_scale=softmax_scale, causal=causal, ) attn_output = torch.cat([attn_output1, attn_output2], dim=-1) else: attn_output1 = flash_attn_func( query_states, key_states, value_states[...,:query_states.shape[-1]], dropout, softmax_scale=softmax_scale, causal=causal, window_size=(self.config.sliding_window, self.config.sliding_window), ) attn_output2 = flash_attn_func( query_states, key_states, value_states[...,query_states.shape[-1]:], dropout, softmax_scale=softmax_scale, causal=causal, window_size=(self.config.sliding_window, self.config.sliding_window), ) attn_output = torch.cat([attn_output1, attn_output2], dim=-1) else: if not use_sliding_windows: attn_output = flash_attn_func( query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal, ) else: attn_output = flash_attn_func( query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal, window_size=(self.config.sliding_window, self.config.sliding_window), ) return attn_output def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length): batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape # On the first iteration we need to properly re-create the padding mask # by slicing it on the proper place if kv_seq_len != attention_mask.shape[-1]: attention_mask_num_tokens = attention_mask.shape[-1] attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :] indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask) key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k) value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k) if query_length == kv_seq_len: query_layer = index_first_axis( query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k ) cu_seqlens_q = cu_seqlens_k max_seqlen_in_batch_q = max_seqlen_in_batch_k indices_q = indices_k elif query_length == 1: max_seqlen_in_batch_q = 1 cu_seqlens_q = torch.arange( batch_size + 1, dtype=torch.int32, device=query_layer.device ) # There is a memcpy here, that is very bad. indices_q = cu_seqlens_q[:-1] query_layer = query_layer.squeeze(1) else: # The -q_len: slice assumes left padding. attention_mask = attention_mask[:, -query_length:] query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask) return ( query_layer, key_layer, value_layer, indices_q, (cu_seqlens_q, cu_seqlens_k), (max_seqlen_in_batch_q, max_seqlen_in_batch_k), ) # Adapted from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->Hymba class HymbaSdpaAttention(HymbaAttention): """ Hymba attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from `HymbaAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to SDPA API. """ # Adapted from HymbaAttention.forward def forward( self, hidden_states: torch.Tensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, kv_last_layer=None, # kv_proj_last_layer = None, use_swa=False, query_states = None, key_states=None, value_states=None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: if output_attentions: return super().forward( hidden_states=hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, ) if self.attn_only_wo_proj: assert query_states is not None bsz, q_len, _ = query_states.size() else: bsz, q_len, _ = hidden_states.size() if not self.attn_only_wo_proj: query_states = self.q_proj(hidden_states) query_states = query_states.view(bsz, q_len, self.num_heads, self.kq_head_dim).transpose(1, 2).contiguous() if self.q_norm is not None: query_states = self.q_norm(query_states) if self.config.rope: if self.attn_only_wo_proj: cos, sin = self.rotary_emb(query_states, position_ids) else: cos, sin = self.rotary_emb(hidden_states, position_ids) query_states, _ = apply_rotary_pos_emb(query_states, None, cos, sin) if self.reuse_kv: assert kv_last_layer is not None key_states, value_states = kv_last_layer # (batch, num_heads, slen, head_dim) else: if not self.attn_only_wo_proj: key_states = self.k_proj(hidden_states) value_states = self.v_proj(hidden_states) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.kq_head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.v_head_dim).transpose(1, 2) if self.k_norm is not None: key_states = self.k_norm(key_states) if self.config.rope: _, key_states = apply_rotary_pos_emb(None, key_states, cos, sin) kv_seq_len = key_states.shape[-2] if past_key_value is not None and not self.reuse_kv and use_cache: kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx) key_states_no_repeat = key_states value_states_no_repeat = value_states key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) if attention_mask is not None: if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): raise ValueError( f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" ) # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, # Reference: https://github.com/pytorch/pytorch/issues/112577. if query_states.device.type == "cuda" and attention_mask is not None: query_states = query_states.contiguous() key_states = key_states.contiguous() value_states = value_states.contiguous() attn_output = torch.nn.functional.scaled_dot_product_attention( query_states, key_states, value_states, attn_mask=attention_mask, dropout_p=self.attention_dropout if self.training else 0.0, # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1. is_causal=self.is_causal and attention_mask is None and q_len > 1, ) attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.reshape(bsz, q_len, self.v_head_dim * self.num_heads) if self.attn_only_wo_proj: return attn_output, (key_states_no_repeat, value_states_no_repeat) attn_output = self.o_proj(attn_output) return attn_output, None, past_key_value, (key_states_no_repeat, value_states_no_repeat) # Adapted from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 with Mistral->Hymba class HymbaFlexAttention(HymbaFlashAttention2): """ Hymba flash attention module. This module inherits from `HymbaAttention` as the weights of the module stays untouched. The only required change would be on the forward pass where it needs to correctly call the public API of flash attention and deal with padding tokens in case the input contains any of them. """ # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__ def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) assert self.config.num_memory_tokens > 0 # assert self.config.sliding_window is not None from torch.nn.attention.flex_attention import flex_attention, create_block_mask, and_masks, or_masks from functools import partial self.create_block_mask = create_block_mask def sliding_window(b, h, q_idx, kv_idx): return q_idx - kv_idx <= self.config.sliding_window def causal_mask(b, h, q_idx, kv_idx): return q_idx >= kv_idx if self.config.sliding_window is not None and self.config.global_attn_idx is not None and self.layer_idx not in self.config.global_attn_idx: attn_mask = and_masks(causal_mask, sliding_window) else: attn_mask = causal_mask if self.config.memory_tokens_interspersed_every > 0: # !If see errors, note that deprecated n_ctx, using seq_length or max_position_embeddings instead num_memory_band = self.config.seq_length // self.config.memory_tokens_interspersed_every qk_length = self.config.seq_length + num_memory_band * self.config.num_memory_tokens num_tokens_per_band = qk_length // num_memory_band for i in range(num_memory_band): left_mask = lambda b, h, q_idx, kv_idx, i=i: kv_idx > i * num_tokens_per_band right_mask = lambda b, h, q_idx, kv_idx, i=i: kv_idx < i * num_tokens_per_band + self.config.num_memory_tokens band_mask = and_masks(left_mask, right_mask) if i == 0: prefix_mask_interspersed = band_mask else: prefix_mask_interspersed = or_masks(prefix_mask_interspersed, band_mask) register_mask = and_masks(causal_mask, prefix_mask_interspersed) else: def prefix_mask(b, h, q_idx, kv_idx): return kv_idx < self.config.num_memory_tokens register_mask = and_masks(causal_mask, prefix_mask) qk_length = self.config.seq_length + self.config.num_memory_tokens self.attn_mask = or_masks(attn_mask, register_mask) self.block_mask = create_block_mask(self.attn_mask, B=None, H=None, Q_LEN=qk_length, KV_LEN=qk_length, _compile=True) self.flex_attention = torch.compile(flex_attention) def recompile_flexattn(self): from torch.nn.attention.flex_attention import flex_attention self.flex_attention = torch.compile(flex_attention) def forward( self, hidden_states: torch.Tensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, kv_last_layer=None, # kv_proj_last_layer = None, use_swa=False, query_states = None, key_states=None, value_states=None, **kwargs, ): if "padding_mask" in kwargs: warnings.warn( "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`" ) attention_mask = kwargs.pop("padding_mask") if self.attn_only_wo_proj: assert query_states is not None bsz, q_len, _ = query_states.size() else: bsz, q_len, _ = hidden_states.size() if not self.attn_only_wo_proj: query_states = self.q_proj(hidden_states) query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous() if self.q_norm is not None: query_states = self.q_norm(query_states) if self.config.rope: if self.attn_only_wo_proj: cos, sin = self.rotary_emb(query_states, position_ids) else: cos, sin = self.rotary_emb(hidden_states, position_ids) query_states, _ = apply_rotary_pos_emb(query_states, None, cos, sin) if self.reuse_kv: assert kv_last_layer is not None key_states, value_states = kv_last_layer # (batch, num_heads, slen, head_dim) else: if not self.attn_only_wo_proj: key_states = self.k_proj(hidden_states) value_states = self.v_proj(hidden_states) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.kq_head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.v_head_dim).transpose(1, 2) if self.k_norm is not None: key_states = self.k_norm(key_states) if self.config.rope: # cos, sin = self.rotary_emb(hidden_states, position_ids) _, key_states = apply_rotary_pos_emb(None, key_states, cos, sin) kv_seq_len = key_states.shape[-2] if past_key_value is not None and not self.reuse_kv: if self.layer_idx is None: raise ValueError( f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} " "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class " "with a layer index." ) kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) use_sliding_windows = ( _flash_supports_window_size and getattr(self.config, "sliding_window", None) is not None and kv_seq_len > (self.config.sliding_window + self.config.num_memory_tokens if self.config.num_memory_tokens > 0 else self.config.sliding_window) and use_swa ) if not _flash_supports_window_size: logger.warning_once( "The current flash attention version does not support sliding window attention, for a more memory efficient implementation" " make sure to upgrade flash-attn library." ) swa_processed_flag = False if past_key_value is not None and use_cache and not self.reuse_kv: kv_layer_idx = self.layer_idx cache_has_contents = past_key_value.get_seq_length(kv_layer_idx) > 0 if ( getattr(self.config, "sliding_window", None) is not None and kv_seq_len > (self.config.sliding_window + self.config.num_memory_tokens if self.config.num_memory_tokens > 0 else self.config.sliding_window) and cache_has_contents and use_swa ): slicing_tokens = 1 - self.config.sliding_window past_key = past_key_value[kv_layer_idx][0] past_value = past_key_value[kv_layer_idx][1] if self.config.num_memory_tokens > 0: # num_fetched_memory_tokens = min(kv_seq_len - self.config.sliding_window, self.config.num_memory_tokens) num_fetched_memory_tokens = self.config.num_memory_tokens past_key = torch.cat([past_key[:, :, :num_fetched_memory_tokens, :], past_key[:, :, slicing_tokens:, :]], dim=-2).contiguous() past_value = torch.cat([past_value[:, :, :num_fetched_memory_tokens, :], past_value[:, :, slicing_tokens:, :]], dim=-2).contiguous() else: past_key = past_key[:, :, slicing_tokens:, :].contiguous() past_value = past_value[:, :, slicing_tokens:, :].contiguous() ### only keep sliding_window tokens in kv cache: Removed as this will impact the kv_seq_len calculation, resulting in errors for all swa cases past_key_value.key_cache[kv_layer_idx] = past_key past_key_value.value_cache[kv_layer_idx] = past_value if attention_mask is not None: attention_mask = attention_mask[:, slicing_tokens:] attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1) swa_processed_flag = True key_states, value_states = past_key_value.update(key_states, value_states, kv_layer_idx) # print(key_states.shape, value_states.shape) else: cache_has_contents = False # repeat k/v heads if n_kv_heads < n_heads key_states_no_repeat = key_states value_states_no_repeat = value_states key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) dropout_rate = 0.0 if not self.training else self.attention_dropout input_dtype = query_states.dtype if input_dtype == torch.float32: if torch.is_autocast_enabled(): target_dtype = torch.get_autocast_gpu_dtype() # Handle the case where the model is quantized elif hasattr(self.config, "_pre_quantization_dtype"): target_dtype = self.config._pre_quantization_dtype else: target_dtype = self.q_proj.weight.dtype logger.warning_once( f"The input hidden states seems to be silently casted in float32, this might be related to" f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in" f" {target_dtype}." ) query_states = query_states.to(target_dtype) key_states = key_states.to(target_dtype) value_states = value_states.to(target_dtype) if past_key_value is not None and use_cache and (not use_swa or query_states.shape[-2] <= self.config.sliding_window): query_states = query_states.transpose(1, 2) # (batch, slen, num_heads, head_dim) key_states = key_states.transpose(1, 2) # (batch, slen, num_heads, head_dim) value_states = value_states.transpose(1, 2) # (batch, slen, num_heads, head_dim) attn_output = self._flash_attention_forward( query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate, use_sliding_windows=use_sliding_windows and not swa_processed_flag, ) v_dim = value_states.shape[-2] * value_states.shape[-1] attn_output = attn_output.reshape(bsz, q_len, v_dim).contiguous() else: if key_states.shape[-2] <= self.block_mask.shape[-2] - 128 or key_states.shape[-2] > self.block_mask.shape[-2]: block_mask = self.create_block_mask(self.attn_mask, B=None, H=None, Q_LEN=key_states.shape[-2], KV_LEN=key_states.shape[-2]) # , _compile=True) else: block_mask = self.block_mask if value_states.shape[-1] == query_states.shape[-1] * 2: attn_output1 = self.flex_attention(query_states, key_states, value_states[...,:query_states.shape[-1]], block_mask=block_mask) attn_output2 = self.flex_attention(query_states, key_states, value_states[...,query_states.shape[-1]:], block_mask=block_mask) attn_output = torch.cat([attn_output1, attn_output2], dim=-1) else: attn_output = self.flex_attention(query_states, key_states, value_states, block_mask=block_mask) attn_output = attn_output.transpose(1, 2).contiguous() ## [batch_size, seq_length, num_head, v_head_dim] if hasattr(self, 'head_mask') and self.head_mask is not None: head_mask = self.head_mask.to(attn_output) head_mask = head_mask.view(1, 1, -1, 1) attn_output = attn_output * head_mask attn_output = attn_output.reshape(bsz, q_len, self.v_head_dim * self.num_heads) if self.attn_only_wo_proj: return attn_output, (key_states_no_repeat, value_states_no_repeat) attn_output = self.o_proj(attn_output) if not output_attentions: attn_weights = None return attn_output, attn_weights, past_key_value, (key_states_no_repeat, value_states_no_repeat) def set_head_mask(self, mask): self.head_mask = mask JAMBA_ATTENTION_CLASSES = { "eager": HymbaAttention, "flash_attention_2": HymbaFlashAttention2, "sdpa": HymbaSdpaAttention, ## the default attention "flex": HymbaFlexAttention, } # Adapted from transformers.models.mamba.modeling_mamba.MambaMixer class HymbaBlock(nn.Module): """ Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`. A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective) ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4, and is why Mamba is called **selective** state spaces) """ def __init__(self, config: HymbaConfig, layer_idx, reuse_kv=None): super().__init__() self.config = config self.layer_idx = layer_idx self.hidden_size = config.hidden_size self.ssm_state_size = config.mamba_d_state self.conv_kernel_size = config.mamba_d_conv self.intermediate_size = int(config.mamba_expand * config.hidden_size) self.reuse_kv = reuse_kv self.attn_hidden_size = config.hidden_size self.num_attention_heads = config.num_attention_heads self.num_key_value_heads = config.num_key_value_heads config.v_head_dim = self.intermediate_size // self.num_attention_heads self.k_hidden_size = int(self.num_key_value_heads/self.num_attention_heads * self.attn_hidden_size) self.v_hidden_size = int(self.num_key_value_heads/self.num_attention_heads * self.attn_hidden_size * config.mamba_expand) self.self_attn = JAMBA_ATTENTION_CLASSES[config.attn_implementation](config, layer_idx, attn_only_wo_proj=True, reuse_kv=reuse_kv) self.time_step_rank = config.mamba_dt_rank self.use_conv_bias = config.mamba_conv_bias self.use_bias = config.mamba_proj_bias self.activation = config.hidden_act self.act = ACT2FN[config.hidden_act] self.apply_inner_layernorms = config.mamba_inner_layernorms self.use_fast_kernels = True # config.use_mamba_kernels if self.reuse_kv: self.latent_dim = self.intermediate_size + self.attn_hidden_size ## mamba plus q else: self.latent_dim = self.intermediate_size + self.attn_hidden_size + self.k_hidden_size + self.v_hidden_size ## mamba plus qkv self.pre_avg_layernorm1 = HymbaRMSNorm(self.intermediate_size, eps=config.rms_norm_eps) self.pre_avg_layernorm2 = HymbaRMSNorm(self.intermediate_size, eps=config.rms_norm_eps) self.in_proj = nn.Linear(self.hidden_size, self.latent_dim + self.intermediate_size, bias=self.use_bias) self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=self.use_bias) num_ssm_param = 1 if not hasattr(config, 'conv_dim'): config.conv_dim = {i:0 for i in range(config.num_hidden_layers)} self.conv1d = nn.Conv1d( in_channels=self.intermediate_size, out_channels=self.intermediate_size, bias=self.use_conv_bias, kernel_size=self.conv_kernel_size, groups=self.intermediate_size, padding=self.conv_kernel_size - 1 ) config.conv_dim[self.layer_idx] = self.intermediate_size self.x_proj = nn.ModuleList([nn.Linear(self.intermediate_size, self.time_step_rank + self.ssm_state_size * 2, bias=False) for _ in range(num_ssm_param)]) self.dt_proj = nn.ModuleList([nn.Linear(self.time_step_rank, self.intermediate_size, bias=True) for _ in range(num_ssm_param)]) A = torch.arange(1, self.ssm_state_size + 1, dtype=torch.float32)[None, :] A = A.expand(self.intermediate_size, -1).contiguous() self.A_log = nn.ParameterList([nn.Parameter(torch.log(A)) for _ in range(num_ssm_param)]) self.D = nn.ParameterList([nn.Parameter(torch.ones(self.intermediate_size)) for _ in range(num_ssm_param)]) if self.apply_inner_layernorms: self.dt_layernorm = HymbaRMSNorm(self.time_step_rank, eps=config.rms_norm_eps) self.B_layernorm = HymbaRMSNorm(self.ssm_state_size, eps=config.rms_norm_eps) self.C_layernorm = HymbaRMSNorm(self.ssm_state_size, eps=config.rms_norm_eps) else: self.dt_layernorm = None self.B_layernorm = None self.C_layernorm = None if not is_fast_path_available: logger.warning_once( "The fast path is not available because on of `(selective_state_update, selective_scan_fn, causal_conv1d_fn, causal_conv1d_update, mamba_inner_fn)`" " is None. To install follow https://github.com/state-spaces/mamba/#installation and" " https://github.com/Dao-AILab/causal-conv1d. If you want to use the naive implementation, set `use_mamba_kernels=False` in the model config" ) def set_attn_mamba_mask(self, attn_branch_mask, mamba_branch_mask): self.attn_branch_mask = attn_branch_mask self.mamba_branch_mask = mamba_branch_mask def _apply_layernorms(self, dt, B, C): if self.dt_layernorm is not None: dt = self.dt_layernorm(dt) if self.B_layernorm is not None: B = self.B_layernorm(B) if self.C_layernorm is not None: C = self.C_layernorm(C) return dt, B, C def cuda_kernels_forward(self, hidden_states: torch.Tensor, cache_params: HybridMambaAttentionDynamicCache = None, attention_mask=None, position_ids=None, kv_last_layer=None, use_cache=False, use_swa=False): projected_states = self.in_proj(hidden_states).transpose(1, 2) ## (bs, latent_dim, seq_len) if ( self.training and cache_params is None and not self.apply_inner_layernorms ): # Doesn't support outputting the states -> used for training contextualized_states = mamba_inner_fn( projected_states, self.conv1d.weight, self.conv1d.bias if self.use_conv_bias else None, self.x_proj.weight, self.dt_proj.weight, self.out_proj.weight, self.out_proj.bias.float() if self.use_bias else None, -torch.exp(self.A_log.float()), None, # input-dependent B None, # input-dependent C self.D.float(), delta_bias=self.dt_proj.bias.float(), delta_softplus=True, ) else: batch_size, seq_len, _ = hidden_states.shape use_precomputed_states = ( cache_params is not None and cache_params.has_previous_state and seq_len == 1 and cache_params.conv_states[self.layer_idx].shape[0] == cache_params.ssm_states[self.layer_idx].shape[0] == batch_size and use_cache ) hidden_states, gate = projected_states.tensor_split((self.latent_dim,), dim=1) conv_weights = self.conv1d.weight.view(self.conv1d.weight.size(0), self.conv1d.weight.size(2)) if self.reuse_kv: query_states, hidden_states = hidden_states.tensor_split((self.attn_hidden_size,), dim=1) query_states = query_states.transpose(1,2) else: query_states, key_states, value_states, hidden_states = hidden_states.tensor_split((self.attn_hidden_size, self.attn_hidden_size + self.k_hidden_size, self.attn_hidden_size + self.k_hidden_size + self.v_hidden_size), dim=1) query_states = query_states.transpose(1,2) key_states = key_states.transpose(1,2) value_states = value_states.transpose(1,2) if use_precomputed_states: hidden_states = causal_conv1d_update( hidden_states.squeeze(-1), cache_params.conv_states[self.layer_idx], conv_weights, self.conv1d.bias, self.activation, ) hidden_states = hidden_states.unsqueeze(-1) cache_params.mamba_past_length[self.layer_idx] += seq_len else: if cache_params is not None: conv_states = nn.functional.pad( hidden_states, (self.conv_kernel_size - hidden_states.shape[-1], 0) ) cache_params.conv_states[self.layer_idx].copy_(conv_states) cache_params.mamba_past_length[self.layer_idx] += seq_len hidden_states = causal_conv1d_fn( hidden_states, conv_weights, self.conv1d.bias, activation=self.activation ) if self.reuse_kv: assert kv_last_layer is not None attn_outputs, attn_key_value = self.self_attn(attention_mask=attention_mask, position_ids=position_ids, query_states=query_states, kv_last_layer=kv_last_layer, use_swa=use_swa, use_cache=use_cache, past_key_value=cache_params) else: attn_outputs, attn_key_value = self.self_attn(attention_mask=attention_mask, position_ids=position_ids, query_states=query_states, key_states=key_states, value_states=value_states, use_swa=use_swa, use_cache=use_cache, past_key_value=cache_params) ## Mamba head index = 0 ssm_parameters = self.x_proj[index](hidden_states.transpose(1, 2)) time_step, B, C = torch.split( ssm_parameters, [self.time_step_rank, self.ssm_state_size, self.ssm_state_size], dim=-1 ) time_step, B, C = self._apply_layernorms(time_step, B, C) if hasattr(self.dt_proj[index], "base_layer"): time_proj_bias = self.dt_proj[index].base_layer.bias self.dt_proj[index].base_layer.bias = None else: time_proj_bias = self.dt_proj[index].bias self.dt_proj[index].bias = None discrete_time_step = self.dt_proj[index](time_step).transpose(1, 2) # [batch, intermediate_size, seq_len] if hasattr(self.dt_proj[index], "base_layer"): self.dt_proj[index].base_layer.bias = time_proj_bias else: self.dt_proj[index].bias = time_proj_bias A = -torch.exp(self.A_log[index].float()) time_proj_bias = time_proj_bias.float() if time_proj_bias is not None else None if use_precomputed_states: scan_outputs = selective_state_update( cache_params.ssm_states[self.layer_idx], hidden_states[..., 0], discrete_time_step[..., 0], A, B[:, 0], C[:, 0], self.D[index], gate[..., 0], time_proj_bias, dt_softplus=True, ).unsqueeze(-1) else: outputs = selective_scan_fn( hidden_states, discrete_time_step, A, B.transpose(1, 2), C.transpose(1, 2), self.D[index].float(), z=gate, delta_bias=time_proj_bias, delta_softplus=True, return_last_state=True, ) if len(outputs) == 3: scan_outputs, ssm_state, _ = outputs else: scan_outputs, ssm_state = outputs if ssm_state is not None and cache_params is not None: cache_params.ssm_states[self.layer_idx].copy_(ssm_state) scan_outputs = scan_outputs.transpose(1, 2) hidden_states = (self.pre_avg_layernorm1(attn_outputs) + self.pre_avg_layernorm2(scan_outputs)) / 2 contextualized_states = self.out_proj(hidden_states) return contextualized_states, attn_key_value def mixer_forward(self, hidden_states, cache_params: HybridMambaAttentionDynamicCache = None, attention_mask=None, position_ids=None, kv_last_layer=None, use_cache=False, use_swa=False): if self.use_fast_kernels: if not is_fast_path_available or "cuda" not in self.x_proj[0].weight.device.type: # if not is_fast_path_available or "cuda" not in self.x_proj.weight.device.type: raise ValueError( "Fast Mamba kernels are not available. Make sure to they are installed and that the mamba module is on a CUDA device" ) return self.cuda_kernels_forward(hidden_states, cache_params, attention_mask=attention_mask, position_ids=position_ids, kv_last_layer=kv_last_layer, use_cache=use_cache, use_swa=use_swa) else: raise ValueError("Support Mamba kernel only") def forward( self, hidden_states: torch.Tensor, past_key_value: Optional[HybridMambaAttentionDynamicCache] = None, **kwargs, ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor]]]: res, attn_key_value = self.mixer_forward(hidden_states, cache_params=past_key_value, attention_mask=kwargs['attention_mask'], kv_last_layer=kwargs['kv_last_layer'], position_ids=kwargs['position_ids'], use_cache=kwargs['use_cache'], use_swa=kwargs['use_swa']) return res, attn_key_value, past_key_value class HymbaMLP(nn.Module): def __init__(self, config: HymbaConfig): super().__init__() # self.config = config self.act_fn_name = config.mlp_hidden_act self.act_fn = ACT2FN[self.act_fn_name] self.ffn_dim = config.intermediate_size self.hidden_dim = config.hidden_size if self.act_fn_name == "silu": self.gate_proj = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False) self.down_proj = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False) self.up_proj = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False) def forward(self, x): if self.act_fn_name == "silu": return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) elif self.act_fn_name == "relu2": return self.down_proj(self.act_fn(self.up_proj(x))) else: raise NotImplementedError(f"No such hidden_act: {self.act_fn_name}") # Adapted from transformers.models.mixtral.modeling_mixtral.MixtralSparseMoeBlock with Mistral->Hymba class HymbaSparseMoeBlock(nn.Module): """ This implementation is strictly equivalent to standard MoE with full capacity (no dropped tokens). It's faster since it formulates MoE operations in terms of block-sparse operations to accomodate imbalanced assignments of tokens to experts, whereas standard MoE either (1) drop tokens at the cost of reduced performance or (2) set capacity factor to number of experts and thus waste computation and memory on padding. """ def __init__(self, config: HymbaConfig, num_experts: int, num_experts_per_tok: int): super().__init__() self.hidden_dim = config.hidden_size self.ffn_dim = config.intermediate_size # these values are decided on runtime depending on the layer index self.num_experts = num_experts self.top_k = num_experts_per_tok if num_experts > 1: # expert routing self.router = nn.Linear(self.hidden_dim, self.num_experts, bias=False) else: self.router = None self.experts = nn.ModuleList([HymbaMLP(config) for _ in range(self.num_experts)]) def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: """ """ if len(hidden_states.shape) == 3: batch_size, sequence_length, hidden_dim = hidden_states.shape bs_times_seq_len = batch_size * sequence_length elif len(hidden_states.shape) == 2: assert self.num_experts == 1 bs_times_seq_len, hidden_dim = hidden_states.shape else: batch_size, sequence_length, _, hidden_dim = hidden_states.shape bs_times_seq_len = batch_size * sequence_length if self.num_experts == 1: # in this case we have a single MLP block and don't need to do any routing final_hidden_states = self.experts[0](hidden_states) router_logits = torch.ones( (bs_times_seq_len, 1), device=hidden_states.device, dtype=hidden_states.dtype, requires_grad=hidden_states.requires_grad, ) return final_hidden_states, router_logits # in this case we have multiple experts and need to do routing hidden_states = hidden_states.view(-1, hidden_dim) # router_logits: (batch * sequence_length, n_experts) router_logits = self.router(hidden_states) routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float) routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1) # we cast back to the input dtype routing_weights = routing_weights.to(hidden_states.dtype) final_hidden_states = torch.zeros( (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device ) # One hot encode the selected experts to create an expert mask # this will be used to easily index which expert is going to be sollicitated expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0) # Loop over all available experts in the model and perform the computation on each expert for expert_idx in range(self.num_experts): expert_layer = self.experts[expert_idx] idx, top_x = torch.where(expert_mask[expert_idx]) if top_x.shape[0] == 0: continue # in torch it is faster to index using lists than torch tensors top_x_list = top_x.tolist() idx_list = idx.tolist() # Index the correct hidden states and compute the expert hidden state for # the current expert. We need to make sure to multiply the output hidden # states by `routing_weights` on the corresponding tokens (top-1 and top-2) current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim) current_hidden_states = expert_layer(current_state) * routing_weights[top_x_list, idx_list, None] # However `index_add_` only support torch tensors for indexing so we'll use # the `top_x` tensor here. final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype)) final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim) return final_hidden_states, router_logits class HymbaDecoderLayer(nn.Module): def __init__(self, config: HymbaConfig, num_experts: int, layer_idx: int, reuse_kv: bool = False): super().__init__() self.config = config self.layer_idx = layer_idx self.reuse_kv = reuse_kv self.mamba = HymbaBlock(config=config, layer_idx=layer_idx, reuse_kv=reuse_kv) self.input_layernorm = HymbaRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.intermediate_size = config.intermediate_size if self.intermediate_size > 0: num_experts_per_tok = config.num_experts_per_tok if num_experts > 1 else 1 self.moe = HymbaSparseMoeBlock(config, num_experts=num_experts, num_experts_per_tok=num_experts_per_tok) self.pre_moe_layernorm = HymbaRMSNorm(config.hidden_size, eps=config.rms_norm_eps) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, attention_mask_raw: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[HybridMambaAttentionDynamicCache] = None, output_attentions: Optional[bool] = False, output_router_logits: Optional[bool] = False, use_cache: Optional[bool] = False, kv_last_layer = None, use_swa=False, **kwargs, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: if "padding_mask" in kwargs: warnings.warn( "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`" ) """ Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`, *optional*): attention mask of size `(batch, sequence_length)` where padding elements are indicated by 0. past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_router_logits (`bool`, *optional*): Whether or not to return the logits of all the routers. They are useful for computing the router loss, and should not be returned during inference. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). """ residual = hidden_states hidden_states = self.input_layernorm(hidden_states) hidden_states, attn_key_value, present_key_value = self.mamba( hidden_states=hidden_states, past_key_value=past_key_value, attention_mask=attention_mask, position_ids=position_ids, kv_last_layer=kv_last_layer, use_cache=use_cache, use_swa=use_swa ) bs, seqlen, _ = hidden_states.shape past_seqlen = self._get_past_seqlen(past_key_value, seqlen) num_attention_heads = self.mamba.config.num_attention_heads self_attn_weights = torch.empty(bs, num_attention_heads, seqlen, past_seqlen, device="meta") # residual connection after mamba hidden_states = residual + hidden_states if self.intermediate_size > 0: residual = hidden_states hidden_states = self.pre_moe_layernorm(hidden_states) hidden_states, router_logits = self.moe(hidden_states) hidden_states = residual + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) if use_cache: outputs += (present_key_value,) if output_router_logits: outputs += (router_logits,) outputs += (attn_key_value,) return outputs def _get_past_seqlen(self, past_key_value, seqlen): if past_key_value is None: return seqlen past_seqlen = past_key_value.get_seq_length() if past_seqlen == 0: return seqlen return past_seqlen class HymbaPreTrainedModel(PreTrainedModel): config_class = HymbaConfig base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = ["HymbaDecoderLayer"] _skip_keys_device_placement = "past_key_values" _supports_flash_attn_2 = True _supports_sdpa = True _supports_cache_class = True def _init_weights(self, module): std = self.config.initializer_range if isinstance(module, (nn.Linear, nn.Conv1d)): module.weight.data.normal_(mean=0.0, std=std) if module.bias is not None: module.bias.data.zero_() elif isinstance(module, nn.Embedding): module.weight.data.normal_(mean=0.0, std=std) if module.padding_idx is not None: module.weight.data[module.padding_idx].zero_() @staticmethod def _convert_to_standard_cache( past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]], batch_size: int ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]: """ Standardizes the format of the cache so as to match most implementations, i.e. have the seqlen as the third dim also for mamba layers """ attn_layer_index = [k.shape == v.shape for k, v in past_key_value].index(True) seqlen = past_key_value[attn_layer_index][0].shape[2] standard_past_key_value = () for k, v in past_key_value: if k.shape != v.shape: # mamba layer # expand doesn't use more memory, so it's fine to do it here standard_past_key_value += ((k.expand(-1, -1, seqlen, -1), v.expand(-1, -1, seqlen, -1)),) else: standard_past_key_value += ((k, v),) return standard_past_key_value @staticmethod def _convert_to_hymba_cache( past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]], ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]: """ Converts the cache to the format expected by Hymba, i.e. dummy seqlen dimesion with size 1 for mamba layers """ hymba_past_key_value = () for k, v in past_key_value: if k.shape != v.shape: # mamba layer hymba_past_key_value += ((k[:, :, :1, :], v[:, :, :1, :]),) else: hymba_past_key_value += ((k, v),) return hymba_past_key_value HYMBA_INPUTS_DOCSTRING = r""" Args: To be added later. Please refer to the forward function. """ # Adapted from transformers.models.mistral.modeling_mistral.MistralModel with MISTRAL->JAMBA, Mistral->Hymba class HymbaModel(HymbaPreTrainedModel): """ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`HymbaDecoderLayer`] Args: config: HymbaConfig """ def __init__(self, config: HymbaConfig): super().__init__(config) config.attn_implementation = config.attn_implementation_new config._attn_implementation = config.attn_implementation_new self.config = 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.inter_layer_kv_reuse = config.kv_reuse_every_i_layer > 0 or config.kv_reuse_group is not None self.kv_reuse_group = config.kv_reuse_group self.kv_reuse_every_i_layer = config.kv_reuse_every_i_layer decoder_layers = [] if self.kv_reuse_group is not None: self.kv_reuse_group = [{'producer': group[0], 'consumer': group[1:]} for group in self.kv_reuse_group] layer_type = [] for i in range(config.num_hidden_layers): if self.inter_layer_kv_reuse: if self.kv_reuse_group is not None: reuse_kv = False for group_id, item in enumerate(self.kv_reuse_group): if i in item['consumer']: reuse_kv = True else: if i % config.kv_reuse_every_i_layer == 0: reuse_kv = False else: reuse_kv = True else: reuse_kv = False layer_type.append('h') decoder_layer = HymbaDecoderLayer(config, num_experts=1, layer_idx=i, reuse_kv=reuse_kv) decoder_layers.append(decoder_layer) config.layer_type = layer_type if config.sliding_window is not None: self.sliding_window = config.sliding_window self.global_attn_idx = config.global_attn_idx else: self.sliding_window = None self.global_attn_idx = None self._attn_layer_index = [] self._hymba_layer_index = [isinstance(layer, HymbaDecoderLayer) for layer in decoder_layers].index(True) self.layers = nn.ModuleList(decoder_layers) self._attn_implementation = config.attn_implementation self.final_layernorm = HymbaRMSNorm(config.hidden_size, eps=config.rms_norm_eps) if self.config.num_memory_tokens > 0: self.memory_tokens = nn.Parameter(torch.randn(self.config.num_memory_tokens, self.config.hidden_size)) self.gradient_checkpointing = False self.post_init() # Ignore copy @add_start_docstrings_to_model_forward(HYMBA_INPUTS_DOCSTRING) def forward( self, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Union[List[torch.FloatTensor], HybridMambaAttentionDynamicCache]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, output_router_logits: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, MoeModelOutputWithPast]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_router_logits = ( output_router_logits if output_router_logits is not None else self.config.output_router_logits ) 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 return_dict = return_dict if return_dict is not None else self.config.use_return_dict # retrieve input_ids and inputs_embeds if input_ids is not None and inputs_embeds is not None: raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") elif input_ids is not None: batch_size, seq_length = input_ids.shape elif inputs_embeds is not None: batch_size, seq_length, _ = inputs_embeds.shape else: raise ValueError("You have to specify either input_ids or inputs_embeds") past_key_values_length = 0 if self.gradient_checkpointing and self.training: if use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." ) use_cache = False if use_cache: use_legacy_cache = False # past_key_values_length = past_key_values.get_usable_length(seq_length, self._attn_layer_index) if past_key_values is not None: past_key_values_length = past_key_values.get_usable_length(seq_length, 0) else: use_cache = False if position_ids is None: device = input_ids.device if input_ids is not None else inputs_embeds.device position_ids = torch.arange( past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device ) position_ids = position_ids.unsqueeze(0).view(-1, seq_length) else: if self.config.num_memory_tokens > 0 and past_key_values is not None and past_key_values.get_seq_length() == 0: position_ids = position_ids.view(-1, seq_length + self.config.num_memory_tokens).long() else: position_ids = position_ids.view(-1, seq_length).long() if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) if self.config.num_memory_tokens > 0 and (past_key_values is None or past_key_values.get_seq_length() == 0): ori_b, ori_n = inputs_embeds.shape[0], inputs_embeds.shape[1] if self.config.memory_tokens_interspersed_every > 0: mem_every = self.config.memory_tokens_interspersed_every next_seq_len = math.ceil(ori_n / mem_every) * mem_every # print(f"before padding: {inputs_embeds.shape}") inputs_embeds = pad_at_dim(inputs_embeds, (0, next_seq_len - ori_n), dim = -2, value = 0.) # print(f"after padding: {inputs_embeds.shape}") inputs_embeds = rearrange(inputs_embeds, 'b (n m) d -> (b n) m d', m = mem_every) # m is the segment length mem = repeat(self.memory_tokens, 'n d -> b n d', b = inputs_embeds.shape[0]) # prepend the memory to every segment of m by repeating the memory tokens inputs_embeds, mem_packed_shape = pack((mem, inputs_embeds), 'b * d') if self.config.memory_tokens_interspersed_every > 0: inputs_embeds = rearrange(inputs_embeds, '(b n) m d -> b (n m) d', b = ori_b) if position_ids is not None and position_ids.shape[1] != inputs_embeds.shape[1]: position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device).unsqueeze(0) attention_mask_raw = attention_mask if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache: is_padding_right = attention_mask[:, -1].sum().item() != batch_size if is_padding_right: raise ValueError( "You are attempting to perform batched generation with padding_side='right'" " this may lead to unexpected behaviour for Flash Attention version of Hymba. Make sure to " " call `tokenizer.padding_side = 'left'` before tokenizing the input. " ) if self._attn_implementation == "flash_attention_2" or self._attn_implementation == "flex": attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None attention_mask_swa = attention_mask elif self._attn_implementation == "sdpa" and not output_attentions: attention_mask_input = attention_mask attention_mask = _prepare_4d_causal_attention_mask_for_sdpa( attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length, ) if self.sliding_window is not None: attention_mask_swa = _prepare_4d_causal_attention_mask_for_sdpa( attention_mask_input, (batch_size, seq_length), inputs_embeds, past_key_values_length, sliding_window=self.sliding_window ) else: # 4d mask is passed through the layers attention_mask = _prepare_4d_causal_attention_mask( attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length, ) if self.sliding_window is not None: attention_mask_swa = _prepare_4d_causal_attention_mask( attention_mask_input, (batch_size, seq_length), inputs_embeds, past_key_values_length, sliding_window=self.sliding_window ) hidden_states = inputs_embeds # decoder layers all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None all_router_logits = () if output_router_logits else None next_decoder_cache = None kv_last_layer = None shared_kv_cache_dict = {} for i, decoder_layer in enumerate(self.layers): if output_hidden_states: all_hidden_states += (hidden_states,) if self.inter_layer_kv_reuse and self.kv_reuse_group is not None: no_reuse_flag = True for group_id, item in enumerate(self.kv_reuse_group): if i in item['consumer']: kv_last_layer = shared_kv_cache_dict[group_id] no_reuse_flag = False # print(f'[Layer-{i}]: Reuse KV cache from Layer-{self.kv_reuse_group[group_id]["producer"]}') break if no_reuse_flag: kv_last_layer = None if self.gradient_checkpointing and self.training: layer_outputs = self._gradient_checkpointing_func( decoder_layer.__call__, hidden_states, attention_mask if (self.sliding_window is None or i in self.global_attn_idx) else attention_mask_swa, attention_mask_raw, position_ids, past_key_values, output_attentions, output_router_logits, use_cache, kv_last_layer, ) else: layer_outputs = decoder_layer( hidden_states, attention_mask=attention_mask if (self.sliding_window is None or i in self.global_attn_idx) else attention_mask_swa, attention_mask_raw=attention_mask_raw, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, output_router_logits=output_router_logits, use_cache=use_cache, kv_last_layer=kv_last_layer if self.inter_layer_kv_reuse else None, use_swa=self.sliding_window is not None and i not in self.global_attn_idx, ) hidden_states = layer_outputs[0] if use_cache: next_decoder_cache = layer_outputs[2 if output_attentions else 1] if output_attentions: all_self_attns += (layer_outputs[1],) if output_router_logits: all_router_logits += (layer_outputs[3],) if self.inter_layer_kv_reuse: kv_last_layer = layer_outputs[-1] if self.kv_reuse_group is not None: for group_id, item in enumerate(self.kv_reuse_group): if i == item['producer']: shared_kv_cache_dict[group_id] = kv_last_layer break del shared_kv_cache_dict hidden_states = self.final_layernorm(hidden_states) # add hidden states from the last decoder layer if output_hidden_states: all_hidden_states += (hidden_states,) if self.config.num_memory_tokens > 0 and (past_key_values is None or past_key_values.get_seq_length() == 0): if self.config.memory_tokens_interspersed_every > 0: hidden_states = rearrange(hidden_states, 'b (n m) d -> (b n) m d', m = (self.config.num_memory_tokens + self.config.memory_tokens_interspersed_every)) mem, hidden_states = unpack(hidden_states, mem_packed_shape, 'b * d') if self.config.memory_tokens_interspersed_every > 0: hidden_states = rearrange(hidden_states, '(b n) m d -> b (n m) d', b = ori_b) hidden_states = hidden_states[:, :ori_n, :] if past_key_values and not past_key_values.has_previous_state: past_key_values.has_previous_state = True next_cache = None if use_cache: next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache if not return_dict: return tuple( v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits] if v is not None ) return MoeModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=next_cache, hidden_states=all_hidden_states, attentions=all_self_attns, router_logits=all_router_logits, ) # Adapted from transformers.models.mixtral.modeling_mixtral.MixtralForCausalLM with MIXTRAL->JAMBA, Mixtral->Hymba class HymbaForCausalLM(HymbaPreTrainedModel): _tied_weights_keys = ["lm_head.weight"] def __init__(self, config: HymbaConfig): super().__init__(config) self.config = config self.model = HymbaModel(config) self.vocab_size = config.vocab_size self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) self.router_aux_loss_coef = config.router_aux_loss_coef self.num_experts = config.num_experts self.num_experts_per_tok = config.num_experts_per_tok # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self): return self.model.embed_tokens def set_input_embeddings(self, value): self.model.embed_tokens = value def get_output_embeddings(self): return self.lm_head def set_output_embeddings(self, new_embeddings): self.lm_head = new_embeddings def set_decoder(self, decoder): self.model = decoder def get_decoder(self): return self.model @add_start_docstrings_to_model_forward(HYMBA_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) def forward( self, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[List[torch.FloatTensor]] = 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, output_router_logits: Optional[bool] = None, return_dict: Optional[bool] = None, calc_logits_for_entire_prompt: Optional[bool] = True, ) -> Union[Tuple, MoeCausalLMOutputWithPast]: r""" Args: 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]`. calc_logits_for_entire_prompt (`bool`, *optional*): Whether or not to calculate the logits for the entire prompt, or just the last token. Only last token logits are needed for generation, and calculating them only for that token can save memory, which becomes pretty significant for long sequences. Returns: ```""" output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_router_logits = ( output_router_logits if output_router_logits is not None else self.config.output_router_logits ) output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) outputs = 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, output_router_logits=output_router_logits, return_dict=return_dict, ) hidden_states = outputs[0] if calc_logits_for_entire_prompt: logits = self.lm_head(hidden_states) else: logits = self.lm_head(hidden_states[..., -1:, :]) logits = logits.float() loss = None if labels is not None: # Shift so that tokens < n predict n shift_logits = logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() # Flatten the tokens loss_fct = CrossEntropyLoss() shift_logits = shift_logits.view(-1, self.config.vocab_size) shift_labels = shift_labels.view(-1) # Enable model parallelism shift_labels = shift_labels.to(shift_logits.device) loss = loss_fct(shift_logits, shift_labels) aux_loss = None if output_router_logits: aux_loss = load_balancing_loss_func( outputs.router_logits if return_dict else outputs[-1], self.num_experts, self.num_experts_per_tok, attention_mask, ) if labels is not None: loss += self.router_aux_loss_coef * aux_loss.to(loss.device) # make sure to reside in the same device if not return_dict: output = (logits,) + outputs[1:] if output_router_logits: output = (aux_loss,) + output return (loss,) + output if loss is not None else output # print("hidden_states.shape:", hidden_states.shape, "input_ids.shape:", input_ids.shape, "logits.shape:", logits.shape) return MoeCausalLMOutputWithPast( loss=loss, aux_loss=aux_loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, router_logits=outputs.router_logits, ) def prepare_inputs_for_generation( self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, output_router_logits=False, **kwargs, ): if self.config.num_memory_tokens > 0: attention_mask = torch.cat([torch.ones(input_ids.shape[0], self.config.num_memory_tokens, device=attention_mask.device), attention_mask], dim=1) if past_key_values is not None and past_key_values.get_seq_length() > 0: if isinstance(past_key_values, Tuple): if past_key_values[self.model._hymba_layer_index][0].shape[2] > 1: past_key_values = self._convert_to_hymba_cache(past_key_values) if isinstance(past_key_values, Cache): cache_length = past_key_values.get_seq_length() past_length = past_key_values.seen_tokens max_cache_length = past_key_values.get_max_length() past_length = cache_length else: cache_length = past_length = past_key_values[self.model._attn_layer_index][0].shape[2] max_cache_length = None # Keep only the unprocessed tokens: # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as # input) if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]: input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :] # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard # input_ids based on the past_length. elif self.config.num_memory_tokens <= 0 and past_length < input_ids.shape[1]: input_ids = input_ids[:, past_length:] elif self.config.num_memory_tokens > 0 and past_length < input_ids.shape[1] + self.config.num_memory_tokens: new_query_id = past_length - self.config.num_memory_tokens input_ids = input_ids[:, new_query_id:] if self.config.sliding_window is not None and (self.config.global_attn_idx is None or len(self.config.global_attn_idx) == 0): input_ids = input_ids[:, -1:] # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens. # If we are about to go beyond the maximum cache length, we need to crop the input attention mask. if ( max_cache_length is not None and attention_mask is not None and cache_length + input_ids.shape[1] > max_cache_length ): attention_mask = attention_mask[:, -max_cache_length:] else: past_key_values = HybridMambaAttentionDynamicCache( self.config, input_ids.shape[0], self.dtype, device=self.device, layer_type=self.config.layer_type ) position_ids = kwargs.get("position_ids", None) if attention_mask is not None and position_ids is None: # create position_ids on the fly for batch generation position_ids = attention_mask.long().cumsum(-1) - 1 position_ids.masked_fill_(attention_mask == 0, 1) if past_key_values.get_seq_length() > 0: position_ids = position_ids[:, -input_ids.shape[1] :] # if `inputs_embeds` are passed, we only want to use them in the 1st generation step if inputs_embeds is not None and past_key_values is None: model_inputs = {"inputs_embeds": inputs_embeds} else: model_inputs = {"input_ids": input_ids} model_inputs.update( { "position_ids": position_ids, "past_key_values": past_key_values, "use_cache": kwargs.get("use_cache"), "attention_mask": attention_mask, "output_router_logits": output_router_logits, "calc_logits_for_entire_prompt": self.config.calc_logits_for_entire_prompt, } ) return model_inputs @staticmethod def _reorder_cache(past_key_values, beam_idx): reordered_past = () for layer_past in past_key_values: reordered_past += ( tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past), ) return reordered_past