# Copyright 2026 the HuggingFace Team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math from collections.abc import Callable from typing import Optional import torch import torch.nn as nn import torch.nn.functional as F from transformers.activations import ACT2FN from transformers.cache_utils import Cache, DynamicCache from transformers.generation import GenerationMixin from transformers.masking_utils import create_causal_mask from transformers.modeling_flash_attention_utils import FlashAttentionKwargs from transformers.modeling_layers import GradientCheckpointingLayer from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel from transformers.processing_utils import Unpack from transformers.utils import TransformersKwargs, can_return_tuple, logging from transformers.utils.generic import check_model_inputs from .configuration_deepseek_v32 import DeepseekV32Config logger = logging.get_logger(__name__) class DeepseekV32RMSNorm(nn.Module): def __init__(self, hidden_size, eps: float = 1e-6) -> None: """ DeepseekV32RMSNorm is equivalent to T5LayerNorm """ super().__init__() self.weight = nn.Parameter(torch.ones(hidden_size)) self.variance_epsilon = eps def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: input_dtype = hidden_states.dtype hidden_states = hidden_states.to(torch.float32) variance = hidden_states.pow(2).mean(-1, keepdim=True) hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) return self.weight * hidden_states.to(input_dtype) def extra_repr(self): return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1): """Applies Rotary Position Embedding to the query and key tensors. Args: q (`torch.Tensor`): The query tensor. k (`torch.Tensor`): The key tensor. cos (`torch.Tensor`): The cosine part of the rotary embedding. sin (`torch.Tensor`): The sine part of the rotary embedding. unsqueeze_dim (`int`, *optional*, defaults to 1): The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. Returns: `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. """ cos = cos.unsqueeze(unsqueeze_dim) sin = sin.unsqueeze(unsqueeze_dim) q_embed = (q * cos) + (rotate_half(q) * sin) k_embed = (k * cos) + (rotate_half(k) * sin) return q_embed, k_embed def 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_interleave(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): r""" TODO let's just use the original freqcis computation to not have the view transpose + reshape! This is not optimized! 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`): The position indices of the tokens corresponding to the query and key tensors. For example, this can be used to pass offsetted position ids when working with a KV-cache. 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) b, h, s, d = q.shape q = q.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d) b, h, s, d = k.shape k = k.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d) q_embed = (q * cos) + (rotate_half(q) * sin) k_embed = (k * cos) + (rotate_half(k) * sin) return q_embed, k_embed def yarn_get_mscale(scale=1, mscale=1): if scale <= 1: return 1.0 return 0.1 * mscale * math.log(scale) + 1.0 def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: """ This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) """ batch, num_key_value_heads, slen, head_dim = hidden_states.shape if n_rep == 1: return hidden_states hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) def eager_attention_forward( module: nn.Module, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, attention_mask: torch.Tensor | None, scaling: float, dropout: float = 0.0, **kwargs: Unpack[TransformersKwargs], ): key_states = repeat_kv(key, module.num_key_value_groups) value_states = repeat_kv(value, module.num_key_value_groups) attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling if attention_mask is not None: causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] attn_weights = attn_weights + causal_mask attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) attn_output = torch.matmul(attn_weights, value_states) attn_output = attn_output.transpose(1, 2).contiguous() return attn_output, attn_weights class DeepseekV32Indexer(nn.Module): def __init__(self, config: "DeepseekV32Config", index_layer_idx: int): super().__init__() self.config = config self.layer_idx = index_layer_idx self.hidden_size: int = config.hidden_size self.num_heads: int = config.index_n_heads self.num_local_heads: int = config.index_n_heads # world_size handling can be added as needed self.head_dim: int = config.index_head_dim self.qk_rope_head_dim: int = config.qk_rope_head_dim self.index_topk: int = config.index_topk self.q_lora_rank: int = config.q_lora_rank self.wq_b = nn.Linear(self.q_lora_rank, self.num_heads * self.head_dim, bias=False) self.wk = nn.Linear(self.hidden_size, self.head_dim, bias=False) self.k_norm = nn.LayerNorm(self.head_dim) self.weights_proj = nn.Linear(self.hidden_size, self.num_heads, dtype=torch.get_default_dtype(), bias=False) self.softmax_scale = self.head_dim**-0.5 @torch.no_grad() def forward( self, hidden_states: torch.Tensor, # [B, S, hidden] q_resid: torch.Tensor, # [B, S, q_lora_rank] position_embeddings: tuple[torch.Tensor, torch.Tensor], attention_mask: torch.Tensor | None, past_key_values_index: "Cache", cache_position: torch.LongTensor | None, ) -> torch.LongTensor: B, S, _ = hidden_states.shape cos, sin = position_embeddings # Queries q_states = self.wq_bj(q_resid) # [B, S, H*D] q_states = q_states.view(B, S, self.num_heads, self.head_dim) # [B, S, H, D] q_rot, q_pass = torch.split(q_states, [self.qk_rope_head_dim, self.head_dim - self.qk_rope_head_dim], dim=-1) q_rot = apply_rotary_pos_emb_interleave(q_rot, cos, sin) # [B, S, H, rope_D] q_states = torch.cat([q_rot, q_pass], dim=-1) # [B, S, H, D] # Keys k = self.k_norm(self.wk(hidden_states)) # [B, S, D] k_rot, k_pass = torch.split(k, [self.qk_rope_head_dim, self.head_dim - self.qk_rope_head_dim], dim=-1) # MLA uses single-head rope stream, then expands later; keep [B, 1, S, rope_D] here k_rot = k_rot.unsqueeze(1) # [B, 1, S, rope_D] k_rot = apply_rotary_pos_emb_interleave(k_rot, cos, sin) # [B, 1, S, rope_D] k_states = torch.cat( [ k_rot.expand(B, self.num_heads, S, -1), # expand rope k_pass.view(B, 1, S, -1).expand(B, self.num_heads, S, -1), ], dim=-1, ) # [B, H, S, D] # Quantize (per provided utilities) # Update indexer cache (layer idx belongs to the attention layer using this indexer) # We store as: keys = k_fp8 (as [B, 1, S, D] or [B, H, S, D]? We keep [B, 1, S, D] like original) # For compactness, collapse heads to 1 for the indexer (you can keep H if your fp8_index expects it). k_1h = k_states.mean(dim=1, keepdim=True) # [B, 1, S, D] (cheap head merge; adjust if needed) k_cache = past_key_values_index.update(k_1h, self.layer_idx, cache_kwargs={"cache_position": cache_position}) # Weights per head head_weights = self.weights_proj(hidden_states) * (self.num_heads**-0.5) # [B, S, H] head_weights = head_weights.unsqueeze(-1) * self.softmax_scale # [B, S, H, *] logits = torch.matmul(k_cache.unsqueeze(1), q_states.transpose(-1, -2)) # [B, M, N, H] # ReLU and sum over heads -> [B, M, N] logits.clamp_min_(0) index_scores = logits.sum(dim=-1) # [B, M, N] if attention_mask is not None: index_scores = index_scores + attention_mask T = index_scores.shape[-1] topk = min(self.index_topk, T) topk_indices = index_scores.topk(topk, dim=-1).indices # [..., topk] return topk_indices class DeepseekV32Attention(nn.Module): """ DeepSeek V3.2 sparse attention mechanism with indexer. This implements the native sparse attention from [DeepSeek V3.2](https://huggingface.co/deepseek-ai/DeepSeek-V3.2) which uses an indexer to select top-k tokens for attention computation, making it more efficient for long sequences. """ def __init__(self, config: DeepseekV32Config, layer_idx: int): super().__init__() self.config = config self.layer_idx = layer_idx self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads self.attention_dropout = config.attention_dropout self.num_heads = config.num_attention_heads self.q_lora_rank = config.q_lora_rank self.qk_rope_head_dim = config.qk_rope_head_dim self.kv_lora_rank = config.kv_lora_rank self.v_head_dim = config.v_head_dim self.qk_nope_head_dim = config.qk_nope_head_dim self.qk_head_dim = config.qk_head_dim self.index_topk = config.index_topk self.is_causal = True # Query projection if self.q_lora_rank is None: self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.qk_head_dim, bias=False) else: self.q_a_proj = nn.Linear(config.hidden_size, config.q_lora_rank, bias=config.attention_bias) self.q_a_layernorm = DeepseekV32RMSNorm(config.q_lora_rank) self.q_b_proj = nn.Linear(config.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False) # Key-Value projections self.kv_a_proj_with_mqa = nn.Linear( config.hidden_size, self.kv_lora_rank + self.qk_rope_head_dim, bias=config.attention_bias, ) self.kv_a_layernorm = DeepseekV32RMSNorm(self.kv_lora_rank) self.kv_b_proj = nn.Linear( self.kv_lora_rank, self.num_heads * (self.qk_nope_head_dim + self.v_head_dim), bias=False, ) # Output projection self.o_proj = nn.Linear( self.num_heads * self.v_head_dim, config.hidden_size, bias=config.attention_bias, ) # Indexer components for sparse attention self.wq_b = nn.Linear(config.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False) self.wk = nn.Linear(config.hidden_size, self.qk_head_dim, bias=config.attention_bias) self.k_norm = DeepseekV32RMSNorm(self.qk_head_dim) self.weights_proj = nn.Linear(config.hidden_size, self.num_heads, bias=False) self.scaling = self.qk_head_dim ** (-0.5) if self.config.rope_scaling.get("rope_type", "default") != "default": mscale_all_dim = self.config.rope_scaling.get("mscale_all_dim", 0) scaling_factor = self.config.rope_scaling["factor"] if mscale_all_dim: mscale = yarn_get_mscale(scaling_factor, mscale_all_dim) self.scaling = self.scaling * mscale * mscale self.indexer = DeepseekV32Indexer(config, layer_idx) def forward( self, hidden_states: torch.Tensor, position_embeddings: tuple[torch.Tensor, torch.Tensor], attention_mask: torch.Tensor | None, past_key_values: Cache | None = None, cache_position: torch.LongTensor | None = None, **kwargs: Unpack[FlashAttentionKwargs], ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]: batch_size, seq_length = hidden_states.shape[:-1] # For training or when index_topk is not effective, fall back to standard attention # This is a simplified implementation - in practice, you'd implement the full sparse indexer if self.training or seq_length <= self.index_topk: logger.warning_once( "DeepSeek V3.2 sparse attention is not fully implemented in this version. " "Falling back to standard attention. For production use, please use vLLM or " "other optimized inference engines.", ) return self._standard_attention( hidden_states, position_embeddings, attention_mask, past_key_values, cache_position, **kwargs ) # Sparse attention implementation would go here # This requires custom CUDA kernels for efficient top-k selection and indexing return self._dsa_attention( hidden_states, position_embeddings, attention_mask, past_key_values, cache_position, **kwargs ) def _standard_attention( self, hidden_states: torch.Tensor, position_embeddings: tuple[torch.Tensor, torch.Tensor], attention_mask: torch.Tensor | None, past_key_values: Cache | None = None, cache_position: torch.LongTensor | None = None, **kwargs: Unpack[FlashAttentionKwargs], ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]: """Standard attention fallback (same as DeepSeek V3)""" batch_size, seq_length = hidden_states.shape[:-1] query_shape = (batch_size, seq_length, -1, self.qk_head_dim) key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim) if self.q_lora_rank is None: q_states = self.q_proj(hidden_states) else: q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states))) q_states = q_states.view(query_shape).transpose(1, 2) q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1) compressed_kv = self.kv_a_proj_with_mqa(hidden_states) k_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1) k_pass = self.kv_b_proj(self.kv_a_layernorm(k_pass)).view(key_shape).transpose(1, 2) k_pass, value_states = torch.split(k_pass, [self.qk_nope_head_dim, self.v_head_dim], dim=-1) k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim) cos, sin = position_embeddings if self.config.rope_interleave: q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin) else: q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin) k_rot = k_rot.expand(*k_pass.shape[:-1], -1) query_states = torch.cat((q_pass, q_rot), dim=-1) key_states = torch.cat((k_pass, k_rot), dim=-1) if past_key_values is not None: cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs) if self.config._attn_implementation == "flash_attention_2" and self.qk_head_dim != self.v_head_dim: value_states = F.pad(value_states, [0, self.qk_head_dim - self.v_head_dim]) attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface( self.config._attn_implementation, eager_attention_forward ) attn_output, attn_weights = attention_interface( self, query_states, key_states, value_states, attention_mask, dropout=0.0 if not self.training else self.attention_dropout, scaling=self.scaling, **kwargs, ) if self.config._attn_implementation == "flash_attention_2" and self.qk_head_dim != self.v_head_dim: attn_output = attn_output[:, :, :, : self.v_head_dim] attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous() attn_output = self.o_proj(attn_output) return attn_output, attn_weights def _dsa_attention( self, hidden_states: torch.Tensor, position_embeddings: tuple[torch.Tensor, torch.Tensor], attention_mask: torch.Tensor | None, past_key_values: Cache | None = None, cache_position: torch.LongTensor | None = None, **kwargs: Unpack[FlashAttentionKwargs] ): B, S, _ = hidden_states.shape cos, sin = position_embeddings # ----- Q path ----- q_resid = self.q_a_layernorm(self.q_a_proj(hidden_states)) # [B, S, q_lora_rank] q_states = self.q_b_proj(q_resid).view(B, S, self.num_heads, self.qk_head_dim) # [B, S, H, D] # Split into pass/rot then apply RoPE on q_rot q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1) q_rot = apply_rotary_pos_emb(q_rot, cos, sin) # [B, S, H, rope_D] q_states = torch.cat([q_pass, q_rot], dim=-1) # [B, S, H, D] # Layout for matmul: [B, H, S, D] q_states = q_states.transpose(1, 2).contiguous() # [B, H, S, D] # ----- KV path (compressed + rope stream) ----- kv_all = self.kv_a_proj_with_mqa(hidden_states) # [B, S, kv_rank + rope_D] kv_compressed, k_rot = torch.split(kv_all, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1) kv_compressed = self.kv_a_layernorm(kv_compressed) # [B, S, kv_rank] # Pre-project to K_pass and V kv_proj = self.kv_b_proj(kv_compressed) # [B, S, H*(qk_nope + v)] kv_proj = kv_proj.view(B, S, self.num_heads, self.qk_nope_head_dim + self.v_head_dim) k_pass, v_states = torch.split( kv_proj, [self.qk_nope_head_dim, self.v_head_dim], dim=-1 ) # [B,S,H,nope], [B,S,H,V] # Rope on K side: keep a single-head rope stream like MLA, then expand k_rot = k_rot.view(B, 1, S, self.qk_rope_head_dim) # [B, 1, S, rope_D] k_rot = apply_rotary_pos_emb(k_rot, cos, sin) # [B, 1, S, rope_D] # Concatenate K = [K_pass, K_rot(expanded)] k_states = torch.cat( ( k_pass.transpose(1, 2), # [B, H, S, nope_D] k_rot.expand(B, self.num_heads, S, -1), ), # [B, H, S, rope_D] dim=-1, ) # [B, H, S, D] v_states = v_states.transpose(1, 2).contiguous() # [B, H, S, V] # ----- Cache update/usage ----- if past_key_values is not None: # Store compressed stream & rope stream (as in original MLA path) # We cache `kv_compressed` under `keys` and `k_rot` under `values` in MlaLayer. # Shapes must be [B, H, t, *] and [B, 1, t, rope_D]. kv_comp_cache = kv_compressed.view(B, 1, S, self.kv_lora_rank).expand(B, self.num_heads, S, -1) k_rot_cache = k_rot # [B, 1, S, rope_D] cached_kv, cached_pe = past_key_values.update( kv_comp_cache, k_rot_cache, layer_idx=self.layer_idx, cache_kwargs={"cache_position": cache_position} ) # Decode path makes use of cached projections; Prefill can use full K/V directly. # ----- Two paths (prefill vs decode) ----- if attention_mask is not None: # Prefill (full attention over local window): standard scaled dot-product with top-k pruning from indexer # Build scores: [B, H, S, S_total] # K layout already [B, H, T, D] scores = (q_states.float() @ k_states.float().transpose(-1, -2)) * self.scaling # [B, H, S, T] # Indexer top-k if past_key_values is not None: topk_idx = self.indexer( hidden_states, q_resid, position_embeddings, attention_mask, past_key_values_index=past_key_values, # we reuse same Cache with IndexerLayer? (separate cache recommended) cache_position=cache_position, ) # Build mask to keep only top-k per (B,S,head?) # Expect topk_idx shape to broadcast to [B, H, S, T]. We scatter along last dim. keep_mask = torch.full_like(scores, float("-inf")) # If topk_idx is [B,S,topk], expand for heads: if topk_idx.dim() == 3: topk_idx = topk_idx.unsqueeze(1).expand(B, self.num_heads, S, -1) keep_mask.scatter_(-1, topk_idx, 0.0) scores = scores + keep_mask probs = nn.functional.softmax(scores, dim=-1, dtype=torch.float32).type_as(hidden_states) # [B, H, S, T] attn_output = probs @ v_states # [B, H, S, V] elif past_key_values is not None: # Decode: use cached compressed KV & rope stream to recompose attention scores efficiently # Compose q_pass and q_rot pieces as in MLA math, but via matmul # 1) Rebuild "nope" term via kv_b weights (dequant on the fly) wkv_b = self.kv_b_proj.weight.view( self.num_heads, self.qk_nope_head_dim + self.v_head_dim, self.kv_lora_rank ) w_k_nope = wkv_b[:, : self.qk_nope_head_dim, :] # [H, nope_D, kv_rank] w_v = wkv_b[:, self.qk_nope_head_dim :, :] # [H, V, kv_rank] # q_pass: [B,H,S,nope_D]; cached_kv: [B,H,T,kv_rank] q_pass = q_states[..., : self.qk_nope_head_dim] # [B,H,S,nope_D] kv_comp = past_key_values[self.layer_idx][0] # keys -> [B,H,T,kv_rank] pe_full = past_key_values[self.layer_idx][1] # values -> [B,1,T,rope_D] # Project q_pass with w_k_nope: [B,H,S,kv_rank] qk_nope = torch.matmul(q_pass, w_k_nope.transpose(-1, -2)) # [B,H,S,kv_rank] # Scores_nope = qk_nope @ kv_comp^T scores_nope = torch.matmul(qk_nope.float(), kv_comp.float().transpose(-1, -2)) # [B,H,S,T] # 2) Rope term: q_rot @ k_rot^T q_rot_only = q_states[..., -self.qk_rope_head_dim :] # [B,H,S,rope_D] k_rot_only = pe_full.expand(B, self.num_heads, -1, -1) # [B,H,T,rope_D] scores_rot = torch.matmul(q_rot_only.float(), k_rot_only.float().transpose(-1, -2)) # [B,H,S,T] scores = (scores_nope + scores_rot) * self.scaling # Indexer top-k (decode) topk_idx = self.indexer( hidden_states, q_resid, position_embeddings, attention_mask, past_key_values_index=past_key_values, cache_position=cache_position, ) # For decode single-step S==1 typically; build a [B,H,1,T] mask keep_mask = torch.full_like(scores, float("-inf")) if topk_idx.dim() == 3: topk_idx = topk_idx.unsqueeze(1).expand(B, self.num_heads, S, -1) keep_mask.scatter_(-1, topk_idx, 0.0) scores = scores + keep_mask probs = nn.functional.softmax(scores, dim=-1, dtype=torch.float32).type_as(hidden_states) # [B,H,S,T] # Rebuild V for decode fast-path: v = (kv_comp @ w_v^T) # kv_comp: [B,H,T,kv_rank], w_v: [H, V, kv_rank] v_from_comp = torch.matmul(kv_comp, w_v.transpose(-1, -2)) # [B,H,T,V] attn_output = torch.matmul(probs, v_from_comp) # [B,H,S,V] # Output projection attn_output = attn_output.transpose(1, 2).reshape(B, S, -1).contiguous() # [B,S,H*V] attn_output = self.o_proj(attn_output) # [B,S,hidden] return attn_output, None class DeepseekV32MLP(nn.Module): def __init__(self, config, intermediate_size=None): super().__init__() self.config = config self.hidden_size = config.hidden_size self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False) self.act_fn = ACT2FN[config.hidden_act] def forward(self, x): down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) return down_proj class DeepseekV32TopkRouter(nn.Module): def __init__(self, config: DeepseekV32Config): super().__init__() self.config = config self.top_k = config.num_experts_per_tok self.n_routed_experts = config.n_routed_experts self.routed_scaling_factor = config.routed_scaling_factor self.n_group = config.n_group self.topk_group = config.topk_group self.norm_topk_prob = config.norm_topk_prob self.weight = nn.Parameter(torch.empty((self.n_routed_experts, config.hidden_size))) self.register_buffer("e_score_correction_bias", torch.zeros((self.n_routed_experts), dtype=torch.float32)) def forward(self, hidden_states): hidden_states = hidden_states.view(-1, self.config.hidden_size) router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32)) return router_logits class DeepseekV32MoE(nn.Module): """ A mixed expert module containing shared experts. """ def __init__(self, config): super().__init__() self.config = config self.experts = nn.ModuleList( [ DeepseekV32MLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(config.n_routed_experts) ] ) self.gate = DeepseekV32TopkRouter(config) self.shared_experts = DeepseekV32MLP( config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts ) def moe(self, hidden_states: torch.Tensor, topk_indices: torch.Tensor, topk_weights: torch.Tensor): r""" CALL FOR CONTRIBUTION! I don't have time to optimise this right now, but expert weights need to be fused to not have to do a loop here (deepseek has 256 experts soooo yeah). """ final_hidden_states = torch.zeros_like(hidden_states, dtype=topk_weights.dtype) expert_mask = torch.nn.functional.one_hot(topk_indices, num_classes=len(self.experts)) expert_mask = expert_mask.permute(2, 0, 1) for expert_idx in range(len(self.experts)): expert = self.experts[expert_idx] mask = expert_mask[expert_idx] token_indices, weight_indices = torch.where(mask) if token_indices.numel() > 0: expert_weights = topk_weights[token_indices, weight_indices] expert_input = hidden_states[token_indices] expert_output = expert(expert_input) weighted_output = expert_output * expert_weights.unsqueeze(-1) final_hidden_states.index_add_(0, token_indices, weighted_output) # in original deepseek, the output of the experts are gathered once we leave this module # thus the moe module is itelsf an IsolatedParallel module # and all expert are "local" meaning we shard but we don't gather return final_hidden_states.type(hidden_states.dtype) def forward(self, hidden_states): residuals = hidden_states orig_shape = hidden_states.shape topk_indices, topk_weights = self.gate(hidden_states) hidden_states = hidden_states.view(-1, hidden_states.shape[-1]) hidden_states = self.moe(hidden_states, topk_indices, topk_weights).view(*orig_shape) hidden_states = hidden_states + self.shared_experts(residuals) return hidden_states class DeepseekV32DecoderLayer(GradientCheckpointingLayer): def __init__(self, config: DeepseekV32Config, layer_idx: int): super().__init__() self.hidden_size = config.hidden_size self.self_attn = DeepseekV32Attention(config, layer_idx) if layer_idx >= config.first_k_dense_replace: self.mlp = DeepseekV32MoE(config) else: self.mlp = DeepseekV32MLP(config) self.input_layernorm = DeepseekV32RMSNorm(config.hidden_size, config.rms_norm_eps) self.post_attention_layernorm = DeepseekV32RMSNorm(config.hidden_size, config.rms_norm_eps) def forward( self, hidden_states: torch.Tensor, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, past_key_values: Cache | None = None, use_cache: bool | None = False, cache_position: torch.LongTensor | None = None, position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None, **kwargs: Unpack[TransformersKwargs], ) -> torch.Tensor: residual = hidden_states hidden_states = self.input_layernorm(hidden_states) # Self Attention hidden_states, _ = self.self_attn( hidden_states=hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, use_cache=use_cache, cache_position=cache_position, position_embeddings=position_embeddings, **kwargs, ) hidden_states = residual + hidden_states # Fully Connected residual = hidden_states hidden_states = self.post_attention_layernorm(hidden_states) hidden_states = self.mlp(hidden_states) hidden_states = residual + hidden_states return hidden_states class DeepseekV32PreTrainedModel(PreTrainedModel): config: DeepseekV32Config base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = ["DeepseekV32DecoderLayer"] _skip_keys_device_placement = ["past_key_values"] _supports_flash_attn = True _supports_sdpa = True _supports_flex_attn = True _can_compile_fullgraph = False _supports_attention_backend = True _can_record_outputs = { "hidden_states":DeepseekV32DecoderLayer, "attentions": DeepseekV32Attention, } @torch.no_grad() def _init_weights(self, module): super()._init_weights(module) if isinstance(module, DeepseekV32TopkRouter): module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) class DeepseekV32RotaryEmbedding(nn.Module): inv_freq: torch.Tensor # fix linting for `register_buffer` def __init__(self, config: DeepseekV32Config, device=None): super().__init__() self.max_seq_len_cached = config.max_position_embeddings self.original_max_seq_len = config.max_position_embeddings self.config = config self.rope_type = self.config.rope_scaling.get("rope_type", "default") rope_init_fn: Callable = self.compute_default_rope_parameters if self.rope_type != "default": rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type] inv_freq, self.attention_scaling = rope_init_fn(self.config, device) self.register_buffer("inv_freq", inv_freq, persistent=False) self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False) @staticmethod def compute_default_rope_parameters( config: DeepseekV32Config | None = None, device: Optional["torch.device"] = None, seq_len: int | None = None, ) -> tuple["torch.Tensor", float]: """ Computes the inverse frequencies according to the original RoPE implementation Args: config ([`~transformers.PreTrainedConfig`]): The model configuration. device (`torch.device`): The device to use for initialization of the inverse frequencies. seq_len (`int`, *optional*): The current sequence length. Unused for this type of RoPE. Returns: Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE). """ base = config.rope_theta partial_rotary_factor = config.rope_scaling.get("partial_rotary_factor", 1.0) head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads dim = int(head_dim * partial_rotary_factor) attention_factor = 1.0 # Unused in this type of RoPE # Compute the inverse frequencies inv_freq = 1.0 / ( base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim) ) return inv_freq, attention_factor @torch.no_grad() @dynamic_rope_update # power user: used with advanced RoPE types (e.g. dynamic rope) def forward(self, x, position_ids): inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device) position_ids_expanded = position_ids[:, None, :].float() device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu" with torch.autocast(device_type=device_type, enabled=False): # Force float32 freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) emb = torch.cat((freqs, freqs), dim=-1) cos = emb.cos() * self.attention_scaling sin = emb.sin() * self.attention_scaling return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) class DeepseekV32Model(DeepseekV32PreTrainedModel): _keys_to_ignore_on_load_unexpected = [r"model\.layers\.78.*"] def __init__(self, config: DeepseekV32Config): super().__init__(config) self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) self.layers = nn.ModuleList( [DeepseekV32DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] ) self.norm = DeepseekV32RMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.rotary_emb = DeepseekV32RotaryEmbedding(config=config) self.gradient_checkpointing = False # Initialize weights and apply final processing self.post_init() @check_model_inputs def forward( self, input_ids: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, past_key_values: Cache | None = None, inputs_embeds: torch.FloatTensor | None = None, cache_position: torch.LongTensor | None = None, use_cache: bool | None = None, **kwargs: Unpack[TransformersKwargs], ) -> BaseModelOutputWithPast: if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") if inputs_embeds is None: inputs_embeds: torch.Tensor = self.embed_tokens(input_ids) if use_cache and past_key_values is None: past_key_values = DynamicCache(config=self.config) if cache_position is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 cache_position: torch.Tensor = ( torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens ) if position_ids is None: position_ids = cache_position.unsqueeze(0) causal_mask = create_causal_mask( config=self.config, input_embeds=inputs_embeds, attention_mask=attention_mask, cache_position=cache_position, past_key_values=past_key_values, position_ids=position_ids, ) hidden_states = inputs_embeds position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids) for decoder_layer in self.layers[: self.config.num_hidden_layers]: hidden_states = decoder_layer( hidden_states, attention_mask=causal_mask, position_embeddings=position_embeddings, position_ids=position_ids, past_key_values=past_key_values, use_cache=use_cache, cache_position=cache_position, **kwargs, ) hidden_states = self.norm(hidden_states) return BaseModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=past_key_values, ) class DeepseekV32ForCausalLM(DeepseekV32PreTrainedModel, GenerationMixin): _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"} _tp_plan = {"lm_head": "colwise_gather_output"} _pp_plan = {"lm_head": (["hidden_states"], ["logits"])} def __init__(self, config): super().__init__(config) self.model = DeepseekV32Model(config) self.vocab_size = config.vocab_size self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) # Initialize weights and apply final processing self.post_init() @can_return_tuple def forward( self, input_ids: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, past_key_values: Cache | None = None, inputs_embeds: torch.FloatTensor | None = None, labels: torch.LongTensor | None = None, use_cache: bool | None = None, cache_position: torch.LongTensor | None = None, logits_to_keep: int | torch.Tensor = 0, **kwargs: Unpack[TransformersKwargs], ) -> CausalLMOutputWithPast: outputs: BaseModelOutputWithPast = self.model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, cache_position=cache_position, **kwargs, ) hidden_states = outputs.last_hidden_state # Only compute necessary logits, and do not upcast them to float if we are not computing the loss slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep logits = self.lm_head(hidden_states[:, slice_indices, :]) loss = None if labels is not None: loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs) return CausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) __all__ = ["DeepseekV32PreTrainedModel", "DeepseekV32Model", "DeepseekV32ForCausalLM"]