# coding=utf-8 # Copyright 2024 Zyphra Technologies and the HuggingFace Inc. team. All rights reserved. # # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math import re from itertools import cycle from typing import Callable, Optional, Union import torch import torch.utils.checkpoint from torch import nn from ...activations import ACT2FN from ...modeling_flash_attention_utils import FlashAttentionKwargs from ...modeling_outputs import BaseModelOutputWithPast from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel from ...processing_utils import Unpack from ...utils import ( logging, ) from ...utils.import_utils import ( is_causal_conv1d_available, is_mamba_ssm_available, ) from ..llama.modeling_llama import LlamaRotaryEmbedding, apply_rotary_pos_emb from ..mamba2.modeling_mamba2 import pad_tensor_by_size, reshape_into_chunks, segment_sum from ..zamba.modeling_zamba import ( ZambaAttention, ZambaAttentionDecoderLayer, ZambaForCausalLM, ZambaForSequenceClassification, ZambaHybridDynamicCache, ZambaHybridLayer, ZambaMambaDecoderLayer, ZambaModel, ZambaRMSNorm, eager_attention_forward, ) from .configuration_zamba2 import Zamba2Config if is_mamba_ssm_available(): from mamba_ssm.ops.triton.selective_state_update import selective_state_update from mamba_ssm.ops.triton.ssd_combined import mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined else: selective_state_update, mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined = None, None, None if is_causal_conv1d_available(): from causal_conv1d import causal_conv1d_fn, causal_conv1d_update else: causal_conv1d_update, causal_conv1d_fn = None, None is_fast_path_available = all((selective_state_update, causal_conv1d_fn, causal_conv1d_update)) _CONFIG_FOR_DOC = "Zyphra/Zamba2-2.7B" logger = logging.get_logger(__name__) class Zamba2RMSNormGated(torch.nn.Module): def __init__(self, hidden_size, group_size, eps=1e-6): super().__init__() self.weight = nn.Parameter(torch.ones(hidden_size)) self.variance_epsilon = eps self.group_size = group_size def forward(self, hidden_states, gate=None): input_dtype = hidden_states.dtype hidden_states = hidden_states.to(torch.float32) if gate is not None: hidden_states = hidden_states * nn.functional.silu(gate.to(torch.float32)) *prefix_dims, last_dim = hidden_states.shape group_count = last_dim // self.group_size hidden_states_group = hidden_states.view(*prefix_dims, group_count, self.group_size) variance = hidden_states_group.pow(2).mean(-1, keepdim=True) hidden_states_group = hidden_states_group * torch.rsqrt(variance + self.variance_epsilon) hidden_states = hidden_states_group.view(*prefix_dims, group_count * self.group_size) return self.weight * hidden_states.to(input_dtype) class Zamba2RMSNorm(ZambaRMSNorm): pass class Zamba2HybridDynamicCache(ZambaHybridDynamicCache): """ 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: Zamba2Config, batch_size: int, dtype: torch.dtype = torch.float16, device: Optional[str] = None ): self.dtype = dtype self.layers_block_type = config.layers_block_type self.has_previous_state = False self.intermediate_size = int(config.mamba_expand * config.hidden_size) self.ssm_state_size = config.mamba_d_state self.conv_kernel_size = config.mamba_d_conv self.n_mamba_heads = config.n_mamba_heads self.transformer_layers = [] self._modules = {} self._parameters = {} self._buffers = {} self.conv_states = {} self.ssm_states = {} for i in range(config.num_hidden_layers): self.conv_states[i] = torch.zeros( batch_size, self.intermediate_size + 2 * config.mamba_ngroups * config.mamba_d_state, self.conv_kernel_size, device=device, dtype=dtype, ) self.ssm_states[i] = torch.zeros( batch_size, self.n_mamba_heads, config.mamba_headdim, self.ssm_state_size, device=device, dtype=dtype ) if self.layers_block_type[i] == "hybrid": self.transformer_layers.append(i) 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)] def update_conv_state( self, layer_idx: int, new_conv_state: torch.Tensor, cache_position: torch.LongTensor ) -> torch.Tensor: conv_state = self.conv_states[layer_idx] cache_position = cache_position.clamp(0, self.conv_kernel_size - 1) conv_state = conv_state.roll(shifts=-1, dims=-1) conv_state[:, :, cache_position] = new_conv_state.to(conv_state.device) self.conv_states[layer_idx].zero_() self.conv_states[layer_idx] += conv_state return self.conv_states[layer_idx] def reset(self): self.conv_states.zero_() self.ssm_states.zero_() 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.transformer_layers[0] if layer_idx not in self.transformer_layers 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] class Zamba2RotaryEmbedding(LlamaRotaryEmbedding): pass class Zamba2Attention(ZambaAttention): """ Multi-headed attention from 'Attention Is All You Need' paper. Adapted from transformers.models.mistral.modeling_mistral.MistralAttention: The input dimension here is attention_hidden_size = 2 * hidden_size, and head_dim = attention_hidden_size // num_heads. The extra factor of 2 comes from the input being the concatenation of original_hidden_states with the output of the previous (mamba) layer (see fig. 2 in https://huggingface.co/papers/2405.16712). Additionally, replaced attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim) with attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim/2) Finally, this attention layer contributes to tied transformer blocks aimed to increasing compute without increasing model size. Because this layer is tied, un-tied adapters (formally the same as LoRA but used in the base model) modules are added to the q, k, v projectors to increase expressivity with a small memory overhead (see Fig. 2 of https://huggingface.co/papers/2411.15242). """ def __init__( self, config: Zamba2Config, layer_idx: Optional[int] = None, num_fwd_mem_blocks: Optional[int] = None, block_id: Optional[int] = None, ): super().__init__(config, layer_idx) self.num_fwd_mem_blocks = num_fwd_mem_blocks self.layer_block_map = config.hybrid_layer_ids self.block_id = block_id if config.use_shared_attention_adapter: self.linear_q_adapter_list = nn.ModuleList([]) self.linear_k_adapter_list = nn.ModuleList([]) self.linear_v_adapter_list = nn.ModuleList([]) for i in range(self.num_fwd_mem_blocks): if i % config.num_mem_blocks == block_id: linear_q_adapter = nn.Sequential( nn.Linear(self.attention_hidden_size, self.config.adapter_rank, bias=False), nn.Linear(self.config.adapter_rank, self.attention_hidden_size, bias=False), ) linear_k_adapter = nn.Sequential( nn.Linear(self.attention_hidden_size, self.config.adapter_rank, bias=False), nn.Linear(self.config.adapter_rank, self.attention_hidden_size, bias=False), ) linear_v_adapter = nn.Sequential( nn.Linear(self.attention_hidden_size, self.config.adapter_rank, bias=False), nn.Linear(self.config.adapter_rank, self.attention_hidden_size, bias=False), ) else: linear_q_adapter = nn.Identity() linear_k_adapter = nn.Identity() linear_v_adapter = nn.Identity() self.linear_q_adapter_list.append(linear_q_adapter) self.linear_k_adapter_list.append(linear_k_adapter) self.linear_v_adapter_list.append(linear_v_adapter) self.layer_dic = {value: index for index, value in enumerate(self.layer_block_map)} def forward( self, hidden_states: torch.Tensor, layer_idx: int, attention_mask: Optional[torch.Tensor] = None, past_key_value: Optional[Zamba2HybridDynamicCache] = None, position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = 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) key_states = self.k_proj(hidden_states) value_states = self.v_proj(hidden_states) if self.config.use_shared_attention_adapter: adapter_layer_idx = self.layer_dic[layer_idx] query_states = query_states + self.linear_q_adapter_list[adapter_layer_idx](hidden_states) key_states = key_states + self.linear_k_adapter_list[adapter_layer_idx](hidden_states) value_states = value_states + self.linear_v_adapter_list[adapter_layer_idx](hidden_states) query_states = query_states.view(hidden_shape).transpose(1, 2) key_states = key_states.view(hidden_shape).transpose(1, 2) value_states = value_states.view(hidden_shape).transpose(1, 2) if self.config.use_mem_rope: cos, sin = position_embeddings query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) if past_key_value is not None: key_states, value_states = past_key_value.update(key_states, value_states, layer_idx) attention_interface: Callable = eager_attention_forward if self.config._attn_implementation != "eager": attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation] attn_output, attn_weights = attention_interface( self, query_states, key_states, value_states, attention_mask, dropout=0.0 if not self.training else self.attention_dropout, scaling=self.scaling, **kwargs, ) attn_output = attn_output.reshape(*input_shape, -1).contiguous() attn_output = self.o_proj(attn_output) return attn_output, attn_weights class Zamba2MambaMixer(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: Zamba2Config, layer_idx: Optional[int] = None): super().__init__() self.config = config 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 * self.hidden_size) self.layer_idx = layer_idx self.use_conv_bias = config.use_conv_bias self.activation = "silu" self.act = nn.SiLU() self.use_mem_eff_path = config.use_mem_eff_path self.n_groups = config.mamba_ngroups self.head_dim = config.mamba_headdim self.num_heads = self.config.n_mamba_heads self.chunk_size = config.chunk_size self.time_step_limit = config.time_step_limit self.time_step_min = config.time_step_min self.time_step_max = config.time_step_max self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size self.conv1d = nn.Conv1d( in_channels=self.conv_dim, out_channels=self.conv_dim, bias=True, kernel_size=config.mamba_d_conv, groups=self.conv_dim, padding=config.mamba_d_conv - 1, ) # projection of the input hidden states projection_size = self.intermediate_size + self.conv_dim + self.num_heads self.in_proj = nn.Linear( self.hidden_size, projection_size, bias=config.add_bias_linear, ) # selective projection used to make dt, B and C input dependent # time step projection (discretization) # instantiate once and copy inv_dt in init_weights of PretrainedModel self.dt_bias = nn.Parameter(torch.ones(self.num_heads)) # S4D real initialization. These are not discretized! # The core is to load them, compute the discrete states, then write the updated state. Keeps the memory bounded A = torch.arange(1, self.num_heads + 1) self.A_log = nn.Parameter(torch.log(A)) self.A_log._no_weight_decay = True self.norm = Zamba2RMSNormGated( self.intermediate_size, group_size=self.intermediate_size // self.n_groups, eps=1e-5 ) self.D = nn.Parameter(torch.ones(self.num_heads)) self.D._no_weight_decay = True self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.add_bias_linear) if not is_fast_path_available: logger.warning_once( "The fast path is not available because on of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)`" " is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and" " https://github.com/Dao-AILab/causal-conv1d" ) def cuda_kernels_forward( self, hidden_states: torch.Tensor, cache_params: Optional[Zamba2HybridDynamicCache] = None, attention_mask: Optional[torch.Tensor] = None, ): # set up dimensions for reshapes later batch_size, seq_len, _ = hidden_states.shape groups_time_state_size = self.n_groups * self.ssm_state_size d_to_remove = 2 * self.intermediate_size + 2 * self.n_groups * self.ssm_state_size + self.num_heads # getting projected states from cache if it exists if cache_params is not None and cache_params.has_previous_state: in_projected_states = self.in_proj(hidden_states.squeeze(1)) # (B 2D) d_mlp = (in_projected_states.shape[-1] - d_to_remove) // 2 split_projection_dim = [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads] _, _, gate, hidden_states_B_C, dt = torch.split(in_projected_states, split_projection_dim, dim=-1) hidden_states_B_C = causal_conv1d_update( hidden_states_B_C, cache_params.conv_states[self.layer_idx], self.conv1d.weight.squeeze(1), self.conv1d.bias, self.activation, ) hidden_states, B, C = torch.split( hidden_states_B_C, [self.intermediate_size, groups_time_state_size, groups_time_state_size], dim=-1, ) A = -torch.exp(self.A_log.float()) # (nheads,) A = A[:, None, ...][:, :, None].expand(-1, self.head_dim, self.ssm_state_size).to(dtype=torch.float32) dt = dt[:, :, None].expand(-1, -1, self.head_dim) dt_bias = self.dt_bias[:, None, ...].expand(-1, self.head_dim) D = self.D[:, None, ...].expand(-1, self.head_dim) B = B.view(batch_size, self.n_groups, B.shape[1] // self.n_groups) C = C.view(batch_size, self.n_groups, C.shape[1] // self.n_groups) hidden_states_reshaped = hidden_states.view(batch_size, self.num_heads, self.head_dim) hidden_states = selective_state_update( cache_params.ssm_states[self.layer_idx], hidden_states_reshaped, dt, A, B, C, D, z=None, dt_bias=dt_bias, dt_softplus=True, ) hidden_states = hidden_states.view(batch_size, self.num_heads * self.head_dim) hidden_states = self.norm(hidden_states, gate) out = self.out_proj(hidden_states)[:, None, ...] # if no cache is found, calling the kernel else: if attention_mask is not None and not torch.all(attention_mask == 1): # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66 dtype = hidden_states.dtype hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype) # 1. Gated MLP's linear projection projected_states = self.in_proj(hidden_states) A = -torch.exp(self.A_log.float()) # (num_heads) or (intermediate_size, state_size) dt_limit_kwargs = {} if self.time_step_limit is None else {"dt_limit": self.time_step_limit} if attention_mask is not None: input_not_masked = torch.all(attention_mask == 1) else: input_not_masked = True if self.use_mem_eff_path and self.training and cache_params is None and input_not_masked: out, ssm_state = mamba_split_conv1d_scan_combined( projected_states, self.conv1d.weight.squeeze(1), self.conv1d.bias, self.dt_bias, A, D=self.D, chunk_size=self.chunk_size, seq_idx=None, activation=self.activation, rmsnorm_weight=self.norm.weight, rmsnorm_eps=self.norm.variance_epsilon, outproj_weight=self.out_proj.weight, outproj_bias=self.out_proj.bias, headdim=self.head_dim, ngroups=self.n_groups, norm_before_gate=False, return_final_states=True, **dt_limit_kwargs, ) else: gate, hidden_states_B_C, time_step = torch.split( projected_states, [self.intermediate_size, self.conv_dim, self.num_heads], dim=-1, ) # 1D Convolution if cache_params is not None: hidden_states_B_C_t = hidden_states_B_C.transpose(1, 2) conv_state = nn.functional.pad( hidden_states_B_C_t, (self.conv_kernel_size - hidden_states_B_C_t.shape[-1], 0) ) cache_params.conv_states[self.layer_idx].copy_(conv_state) if causal_conv1d_fn is None or self.activation not in ["silu", "swish"]: hidden_states_B_C = self.act( self.conv1d(hidden_states_B_C.transpose(1, 2)).transpose(1, 2)[:, :seq_len] ) # (B, L, self.d_inner + 2 * ngroups * d_state) else: hidden_states_B_C = causal_conv1d_fn( x=hidden_states_B_C.transpose(1, 2), weight=self.conv1d.weight.squeeze(1), bias=self.conv1d.bias, activation=self.activation, ).transpose(1, 2)[:, :seq_len] hidden_states, B, C = torch.split( hidden_states_B_C, [self.intermediate_size, groups_time_state_size, groups_time_state_size], dim=-1, ) if attention_mask is not None and not torch.all(attention_mask == 1): # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66 dtype = hidden_states.dtype hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype) scan_output, ssm_state = mamba_chunk_scan_combined( hidden_states.view(batch_size, seq_len, -1, self.head_dim), time_step, A, B.view(batch_size, seq_len, self.n_groups, -1), C.view(batch_size, seq_len, self.n_groups, -1), chunk_size=self.chunk_size, D=self.D, z=None, seq_idx=None, return_final_states=True, dt_bias=self.dt_bias, dt_softplus=True, **dt_limit_kwargs, ) if ssm_state is not None and cache_params is not None: cache_params.ssm_states[self.layer_idx].copy_(ssm_state) scan_output = scan_output.view(batch_size, seq_len, -1) # Multiply "gate" branch and apply extra normalization layer scan_output = self.norm(scan_output, gate) out = self.out_proj(scan_output) return out # fmt: off def torch_forward(self, input_states, cache_params: Optional[Zamba2HybridDynamicCache]=None, attention_mask: Optional[torch.Tensor]=None): batch_size, seq_len, _ = input_states.shape dtype = input_states.dtype # Gated MLP's linear projection if cache_params is not None and cache_params.has_previous_state: projected_states = self.in_proj(input_states.squeeze(1)) else: if attention_mask is not None and not torch.all(attention_mask==1): # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66 input_states = (input_states * attention_mask[:, :, None]).to(dtype) projected_states = self.in_proj(input_states) d_mlp = (projected_states.shape[-1] - 2 * self.intermediate_size - 2 * self.n_groups * self.ssm_state_size- self.num_heads) // 2 _, _, gate, hidden_states, dt = projected_states.split( [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1 ) # Convolution sequence transformation if cache_params is not None: ssm_state = cache_params.ssm_states[self.layer_idx].clone() ssm_state = ssm_state.to(hidden_states.device) if cache_params.has_previous_state: gate = gate.unsqueeze(1) conv_state = cache_params.conv_states[self.layer_idx] # [batch, intermediate_size, conv_kernel_size] conv_state = torch.roll(conv_state, shifts=-1, dims=-1) # handle batched generation - states are copied through conv_state[:, :, -1] = hidden_states[:, 0, :] if hidden_states.ndim == 3 else hidden_states cache_params.conv_states[self.layer_idx].copy_(conv_state) hidden_states = torch.sum(conv_state.to(projected_states.device) * self.conv1d.weight[:, 0, :], dim=-1) if self.use_conv_bias: hidden_states += self.conv1d.bias hidden_states = self.act(hidden_states).to(dtype)[:, None, ...] # [batch, 1, intermediate_size] : decoding else: hidden_states = hidden_states.transpose(1,2) conv_state = nn.functional.pad( hidden_states, (self.conv_kernel_size - hidden_states.shape[-1], 0) ) cache_params.conv_states[self.layer_idx].copy_(conv_state) hidden_states = self.act(self.conv1d(hidden_states).transpose(1,2))[:, :seq_len, :] # [batch, intermediate_size, seq_len] if attention_mask is not None and not torch.all(attention_mask==1): dtype = hidden_states.dtype # tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66 hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype) else: ssm_state = torch.zeros( (batch_size, self.num_heads, self.head_dim, self.ssm_state_size), device=hidden_states.device, dtype=dtype ) hidden_states = self.act(self.conv1d(hidden_states.transpose(1, 2))[..., :seq_len].transpose(1, 2)) hidden_states, B, C = torch.split(hidden_states, [self.intermediate_size, self.n_groups * self.ssm_state_size, self.n_groups * self.ssm_state_size], dim=-1) A = -torch.exp(self.A_log.float()) # [num_heads] if cache_params is not None and cache_params.has_previous_state: # Note: there is no need to pad parameter matrices here, as there is just one new token # for batched generation dt = dt[:, None, ...] if dt.ndim == 2 else dt[:, 0, :][:, None, ...] dt = dt.transpose(1, 2).expand(batch_size, dt.shape[-1], self.head_dim) # [num_heads] -> [num_heads, head_dim] dt_bias = self.dt_bias[..., None].expand(self.dt_bias.shape[0], self.head_dim) dt = torch.nn.functional.softplus(dt + dt_bias.to(dt.dtype)) dt = torch.clamp(dt, self.time_step_min) #, self.time_step_max) A = A[..., None, None].expand(self.num_heads, self.head_dim, self.ssm_state_size).to(dtype=torch.float32) # [bsz, num_heads, head_dim, state_size] dA = torch.exp(dt[..., None] * A) # Discretize B # [bsz, n_groups * state_size] -> [bsz, n_groups, 1, state_size] -> # -> [bsz, n_groups, group to head repetition factor, state_size] -> [bsz, num_heads, state_size] B = B.reshape(batch_size, self.n_groups, -1)[..., None, :] B = B.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, B.shape[-1]).contiguous() B = B.reshape(batch_size, -1, B.shape[-1]) # [bsz, num_heads, head_dim, state_size] dB = dt[..., None] * B[..., None, :] # Discretize x into dB # [bsz, intermediate_size] -> [bsz, num_heads, head_dim] hidden_states = hidden_states.reshape(batch_size, -1, self.head_dim) dBx = dB * hidden_states[..., None] # State calculation cache_params.ssm_states[self.layer_idx].copy_( cache_params.ssm_states[self.layer_idx] * dA + dBx ) # Subsequent output # [bsz, n_groups * state_size] -> [bsz, num_heads, state_size] C = C.reshape(batch_size, self.n_groups, -1)[..., None, :] C = C.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, C.shape[-1]).contiguous() C = C.reshape(batch_size, -1, C.shape[-1]) # [bsz, num_heads, head_dim] ssm_states = cache_params.ssm_states[self.layer_idx].to(C.dtype) # Shape: [b, h, d, n] # Reshape ssm_states to merge the first two dimensions ssm_states_reshaped = ssm_states.view(batch_size * self.num_heads, self.head_dim, self.ssm_state_size) # Shape: [b*h, d, n] C_reshaped = C.view(batch_size * self.num_heads, self.ssm_state_size, 1) # Shape: [b*h, n, 1] y = torch.bmm(ssm_states_reshaped, C_reshaped) y = y.view(batch_size, self.num_heads, self.head_dim) # D skip connection # [num_heads] -> [num_heads, head_dim] D = self.D[..., None].expand(self.D.shape[0], self.head_dim) y = (y + hidden_states * D).to(y.dtype) # [bsz, num_heads, head_dim] -> [bsz, 1, intermediate_size] y = y.reshape(batch_size, -1)[:, None, ...] else: # begin ssd naive implementation without einsums dt = nn.functional.softplus(dt + self.dt_bias) dt = torch.clamp(dt, self.time_step_min) hidden_states = hidden_states.reshape(batch_size, seq_len, -1, self.head_dim).float() B = B.reshape(batch_size, seq_len, -1, self.ssm_state_size).float() C = C.reshape(batch_size, seq_len, -1, self.ssm_state_size).float() B = B.repeat_interleave(self.num_heads // self.n_groups, dim=2, output_size=self.num_heads) C = C.repeat_interleave(self.num_heads // self.n_groups, dim=2, output_size=self.num_heads) pad_size = (self.chunk_size - seq_len % self.chunk_size) % self.chunk_size D_residual = self.D[..., None] * pad_tensor_by_size(hidden_states, pad_size) # Discretize x and A hidden_states = hidden_states * dt[..., None] A = A.to(hidden_states.dtype) * dt # Rearrange into blocks/chunks hidden_states, A, B, C = [reshape_into_chunks(t, pad_size, self.chunk_size) for t in (hidden_states, A, B, C)] # [bsz, -1, chunk_size, num_heads] -> [bsz, num_heads, -1, chunk_size] A = A.permute(0, 3, 1, 2) A_cumsum = torch.cumsum(A, dim=-1) # 1. Compute the output for each intra-chunk (diagonal blocks) # This is the analog of a causal mask L = torch.exp(segment_sum(A)) # First, contraction of C and B to get G (attention-weights like) G_intermediate = C[:, :, :, None, :, :] * B[:, :, None, :, : ,:] # shape: (b, c, l, s, h, n) G = G_intermediate.sum(dim=-1) # shape: (b, c, l, s, h) # Step 2: Compute M, equivalent to applying attention mask to weights M_intermediate = G[..., None] * L.permute(0, 2, 3, 4, 1)[..., None] M = M_intermediate.sum(dim=-1) # Step 3: Compute Y_diag (apply to values) Y_diag = (M[..., None] * hidden_states[:, :, None]).sum(3) # (right term of low-rank factorization of off-diagonal blocks; B terms) decay_states = torch.exp(A_cumsum[:, :, :, -1:] - A_cumsum) B_decay_contraction = B * decay_states.permute(0, 2, 3, 1)[..., None] # permute back B * decay states states = (B_decay_contraction.permute(0, 1, 3, 2, 4)[..., None] * hidden_states.permute(0, 1, 3, 2, 4)[..., None, :]).sum(dim=3).permute(0, 1, 2, 4, 3) if cache_params is not None and cache_params.has_previous_state: previous_states = cache_params.ssm_states[self.layer_idx][:, None, ...] else: previous_states = torch.zeros_like(states[:, :1]) states = torch.cat([previous_states, states], dim=1) decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0)))) states_permuted = states.permute(0, 2, 1, 3, 4) result = (decay_chunk[..., None, None] * states_permuted[:, :, None, ...]).sum(dim=2) new_states = result.permute(0, 2, 1, 3, 4) states, ssm_state = new_states[:, :-1], new_states[:, -1] # Compute state -> output conversion per chunk # (left term of low-rank factorization of off-diagonal blocks; C terms) state_decay_out = torch.exp(A_cumsum) # compute Yoff C_times_states = (C[..., None, :] * states[:, :, None, ...]) state_decay_out_permuted = state_decay_out.permute(0, 2, 3, 1) Y_off = (C_times_states.sum(-1) * state_decay_out_permuted[..., None]) # Add output of intra-chunk and inter-chunk terms (diagonal and off-diagonal blocks) y = Y_diag + Y_off # [bsz, -1, self.chunk_size, num_heads, head_dim] -> [bsz, (padded) seq_len, num_heads, head_dim] y = y.reshape(batch_size, -1, self.num_heads, self.head_dim) y = y + D_residual # Cutting off padded chunks if pad_size > 0: y = y[:, :seq_len, :, :] y = y.reshape(batch_size, seq_len, -1) if ssm_state is not None and cache_params is not None: cache_params.ssm_states[self.layer_idx].copy_(ssm_state) scan_output = self.norm(y, gate) # end ssd naive # 4. Final linear projection contextualized_states = self.out_proj(scan_output.to(dtype)) # [batch, seq_len, hidden_size] return contextualized_states # fmt: on def forward( self, hidden_states, cache_params: Optional[Zamba2HybridDynamicCache] = None, attention_mask: Optional[torch.Tensor] = None, ): if is_fast_path_available and "cuda" in self.in_proj.weight.device.type: return self.cuda_kernels_forward(hidden_states, cache_params, attention_mask) return self.torch_forward(hidden_states, cache_params, attention_mask) class Zamba2MLP(nn.Module): def __init__(self, config: Zamba2Config, num_fwd_mem_blocks=None, block_id: Optional[int] = None): """ This MLP layer contributes to tied transformer blocks aimed to increasing compute without increasing model size. Because this layer is tied, un-tied adapter modules (formally same as LoRA, but used in the base model) are added to the up and gate projectors to increase expressivity with a small memory overhead. """ super().__init__() self.config = config self.hidden_size = config.hidden_size self.intermediate_size = config.intermediate_size self.num_fwd_mem_blocks = num_fwd_mem_blocks self.block_id = block_id self.gate_up_proj = nn.Linear(self.hidden_size, 2 * self.intermediate_size, bias=config.add_bias_linear) self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.add_bias_linear) self.act_fn = ACT2FN[config.hidden_act] self.gate_up_proj_adapter_list = nn.ModuleList([]) for i in range(self.num_fwd_mem_blocks): if i % config.num_mem_blocks == block_id: gate_up_proj_adapter = nn.Sequential( nn.Linear(self.config.hidden_size, self.config.adapter_rank, bias=False), nn.Linear(self.config.adapter_rank, 2 * self.intermediate_size, bias=False), ) else: gate_up_proj_adapter = nn.Identity() self.gate_up_proj_adapter_list.append(gate_up_proj_adapter) layer_block_map = config.hybrid_layer_ids self.layer_dic = {value: index for index, value in enumerate(layer_block_map)} def forward(self, hidden_state, layer_idx=None): gate_up_state = self.gate_up_proj(hidden_state) layer_idx = self.layer_dic[layer_idx] gate_up_state = gate_up_state + self.gate_up_proj_adapter_list[layer_idx](hidden_state) gate_up_state = torch.chunk(gate_up_state, 2, dim=-1) hidden_state = self.act_fn(gate_up_state[0]) * gate_up_state[1] output = self.down_proj(hidden_state) return output class Zamba2AttentionDecoderLayer(ZambaAttentionDecoderLayer): def __init__(self, config: Zamba2Config, block_id: Optional[int] = None, layer_idx: Optional[int] = None): self.block_id = block_id num_gs = len(config.hybrid_layer_ids) super().__init__(config, layer_idx) self.self_attn = Zamba2Attention(config, layer_idx=-1, num_fwd_mem_blocks=num_gs, block_id=block_id) self.feed_forward = Zamba2MLP(config, num_fwd_mem_blocks=num_gs, block_id=block_id) def forward( self, hidden_states: torch.Tensor, original_hidden_states: torch.Tensor, layer_idx: int, attention_mask: Optional[torch.Tensor] = None, past_key_value: Optional[Zamba2HybridDynamicCache] = None, output_attentions: Optional[bool] = False, position_embeddings: Optional[torch.LongTensor] = None, **kwargs: Unpack[FlashAttentionKwargs], ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: hidden_states (`torch.FloatTensor`): output of previous Mamba layer of shape `(batch, seq_len, embed_dim)` original_hidden_states (`torch.FloatTensor`): word embedding output of shape `(batch, seq_len, embed_dim)`. This is concatenated with `hidden_states` (which is the output of the previous (mamba) layer). The concatenated tensor is then used as input of the pre-attention RMSNorm (see fig. 2 in https://huggingface.co/papers/2405.16712). attention_mask (`torch.FloatTensor`, *optional*): attention mask of size `(batch, sequence_length)` where padding elements are indicated by 0. past_key_value (`Zamba2HybridDynamicCache`, *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. 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`). position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*): Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`, with `head_dim` being the embedding dimension of each attention head. """ hidden_states = torch.concatenate([hidden_states, original_hidden_states], dim=-1) hidden_states = self.input_layernorm(hidden_states) hidden_states, self_attn_weights = self.self_attn( hidden_states=hidden_states, layer_idx=layer_idx, attention_mask=attention_mask, past_key_value=past_key_value, output_attentions=output_attentions, position_embeddings=position_embeddings, **kwargs, ) hidden_states = self.pre_ff_layernorm(hidden_states) hidden_states = self.feed_forward(hidden_states, layer_idx) outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) return outputs class Zamba2MambaDecoderLayer(ZambaMambaDecoderLayer): def __init__(self, config: Zamba2Config, layer_idx: int): super().__init__(config, layer_idx) self.mamba = Zamba2MambaMixer(config=config, layer_idx=layer_idx) self.input_layernorm = Zamba2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) class Zamba2HybridLayer(ZambaHybridLayer): def __init__( self, shared_transformer: Zamba2AttentionDecoderLayer, linear: nn.Linear, mamba: Zamba2MambaDecoderLayer ): super().__init__(shared_transformer, linear, mamba) del self.shared_transf self.shared_transformer = shared_transformer def forward( self, hidden_states: torch.Tensor, original_hidden_states: Optional[torch.Tensor] = None, layer_idx: Optional[int] = None, attention_mask: Optional[torch.Tensor] = None, causal_mask: Optional[torch.Tensor] = None, past_key_value: Optional[Zamba2HybridDynamicCache] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, position_embeddings: Optional[torch.LongTensor] = None, ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` original_hidden_states (`torch.FloatTensor`): word embedding output that will be concatenated with hidden activations to form the input of the shared transformer layer. layer_idx (`int`): layer number. attention_mask (`torch.FloatTensor`, *optional*): attention mask of size `(batch, sequence_length)` where padding elements are indicated by 0. past_key_value (`Zamba2HybridDynamicCache`, *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. 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`). position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*): Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`, with `head_dim` being the embedding dimension of each attention head. """ layer_outputs = self.shared_transformer( hidden_states, original_hidden_states=original_hidden_states, layer_idx=layer_idx, attention_mask=causal_mask, past_key_value=past_key_value, output_attentions=output_attentions, position_embeddings=position_embeddings, ) transformer_hidden_states = layer_outputs[0] if output_attentions: self_attn_weights = layer_outputs[1] transformer_hidden_states = self.linear(transformer_hidden_states) layer_outputs = self.mamba_decoder( hidden_states, transformer_hidden_states=transformer_hidden_states, attention_mask=attention_mask, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, position_embeddings=position_embeddings, ) if output_attentions: layer_outputs = (layer_outputs[0], self_attn_weights) + layer_outputs[2:] return layer_outputs class Zamba2PreTrainedModel(PreTrainedModel): config: Zamba2Config base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = ["Zamba2AttentionDecoderLayer", "Zamba2MambaDecoderLayer"] _skip_keys_device_placement = "past_key_values" _supports_flash_attn = True _supports_flex_attn = True _supports_sdpa = True # Note: only supports Zamba2HybridDynamicCache _is_stateful = True def _init_weights(self, module): super()._init_weights(module) if isinstance(module, Zamba2MambaMixer): dt = torch.exp( torch.rand(self.config.n_mamba_heads) * (math.log(self.config.time_step_max) - math.log(self.config.time_step_min)) + math.log(self.config.time_step_min) ).clamp(min=self.config.time_step_floor) # # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759 inv_dt = dt + torch.log(-torch.expm1(-dt)) module.dt_bias.data.copy_(inv_dt) A = torch.arange(1, module.num_heads + 1) module.A_log.data.copy_(torch.log(A)) module.D.data.fill_(1.0) class Zamba2Model(ZambaModel, Zamba2PreTrainedModel): """ Model consisting of *config.num_hidden_layers* layers. Args: config: Zamba2Config """ def __init__(self, config: Zamba2Config): Zamba2PreTrainedModel.__init__(self, config) 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) blocks = [Zamba2AttentionDecoderLayer(config, block_id=k) for k in range(config.num_mem_blocks)] mamba_layers = [] linear_layers = [] self.layers_block_type = config.layers_block_type for i in range(config.num_hidden_layers): if config.layers_block_type[i] == "mamba": mamba_layers.append(Zamba2MambaDecoderLayer(config, layer_idx=i)) elif config.layers_block_type[i] == "hybrid": linear_layers.append(nn.Linear(self.config.hidden_size, self.config.hidden_size, bias=False)) mamba_layers.append(Zamba2MambaDecoderLayer(config, layer_idx=i)) mamba_layers = iter(mamba_layers) linear_layers = iter(linear_layers) blocks = cycle(blocks) layers = self.get_layers(blocks, linear_layers, mamba_layers) self.layers = nn.ModuleList(layers) self._attn_implementation = config._attn_implementation self.final_layernorm = Zamba2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) if config.use_mem_rope: if config.use_long_context: logger.warning_once( "`use_long_context` set to `True`: using rescaled `rope_theta` and extended `max_position_embeddings`." ) self.rotary_emb = Zamba2RotaryEmbedding(config) self.gradient_checkpointing = False # Initialize weights and apply final processing self.post_init() def get_layers(self, blocks, linear_layers, mamba_layers): layers = [] self._tied_weights_keys = [] self.first_transformer_layer_id = 0 for layer_id, layer_type in enumerate(self.layers_block_type): if layer_type == "hybrid": if self.first_transformer_layer_id == 0: self.first_transformer_layer_id = layer_id block = next(blocks) if self.config.num_mem_blocks * len(self.config.hybrid_layer_ids) > 1: prefix_pattern = rf"^layers\.{layer_id}\.shared_transformer\." main_keys_pattern = re.compile( prefix_pattern + r"(?:" + r"self_attn\.(?:q_proj|k_proj|v_proj|o_proj)\.weight|" + r"feed_forward\.(?:gate_up_proj|down_proj)\.weight|" + r"(?:input_layernorm|pre_ff_layernorm)\.weight" + r")$" ) self._tied_weights_keys.append(main_keys_pattern) adapter_id = 0 for _layer_type in self.layers_block_type: if _layer_type == "hybrid" and adapter_id % self.config.num_mem_blocks == block.block_id: adapter_pattern = re.compile( r"^shared_transformer\.feed_forward\.gate_up_proj_adapter_list\." + str(adapter_id) + r"\.(?:0|1)\.weight$" ) self._tied_weights_keys.append(adapter_pattern) adapter_id += 1 if self.config.use_shared_attention_adapter: adapter_id = 0 for _layer_type in self.layers_block_type: if _layer_type == "hybrid" and adapter_id % self.config.num_mem_blocks == block.block_id: attn_adapter_pattern = re.compile( r"^shared_transformer\.self_attn\." + r"(?:linear_q_adapter_list|linear_k_adapter_list|linear_v_adapter_list)\." + str(adapter_id) + r"\.(?:0|1)\.weight$" ) self._tied_weights_keys.append(attn_adapter_pattern) adapter_id += 1 layers.append(Zamba2HybridLayer(block, next(linear_layers), next(mamba_layers))) else: layers.append(next(mamba_layers)) return layers 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[Zamba2HybridDynamicCache] = None, inputs_embeds: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, ) -> Union[tuple, 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 return_dict = return_dict if return_dict is not None else self.config.use_return_dict if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError( "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" ) 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) hidden_states = inputs_embeds original_hidden_states = torch.clone(inputs_embeds) # original_hidden_states: word embedding output that will be concatenated with hidden activations to form the input of the shared transformer layer if use_cache and past_key_values is None: batch_size = input_ids.shape[0] if input_ids is not None else inputs_embeds.shape[0] past_key_values = Zamba2HybridDynamicCache(self.config, batch_size, dtype=self.dtype, device=self.device) if cache_position is None: past_seen_tokens = ( past_key_values.get_seq_length(layer_idx=self.first_transformer_layer_id) 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) causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position) # create position embeddings to be shared across the decoder layers if self.config.use_mem_rope: position_embeddings = self.rotary_emb(hidden_states, position_ids) else: position_embeddings = None all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None for layer_idx, layer in enumerate(self.layers): if output_hidden_states: all_hidden_states += (hidden_states,) if self.gradient_checkpointing and self.training: layer_outputs = self._gradient_checkpointing_func( layer.__call__, hidden_states, original_hidden_states, layer_idx, attention_mask, causal_mask, past_key_values, output_attentions, use_cache, position_embeddings, ) else: layer_outputs = layer( hidden_states, original_hidden_states=original_hidden_states, layer_idx=layer_idx, attention_mask=attention_mask, causal_mask=causal_mask, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, position_embeddings=position_embeddings, ) hidden_states = layer_outputs[0] if output_attentions: if layer_outputs[1] is not None: # append attentions only of attention layers. Mamba layers return `None` as the attention weights all_self_attns += (layer_outputs[1],) 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 past_key_values is not None and not past_key_values.has_previous_state: past_key_values.has_previous_state = True output = BaseModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=past_key_values if use_cache else None, hidden_states=all_hidden_states, attentions=all_self_attns, ) return output if return_dict else output.to_tuple() class Zamba2ForCausalLM(ZambaForCausalLM): pass class Zamba2ForSequenceClassification(ZambaForSequenceClassification): pass __all__ = [ "Zamba2ForCausalLM", "Zamba2ForSequenceClassification", "Zamba2Model", "Zamba2PreTrainedModel", ]