modeling_livemem_gdn2.py
| 1 | """GDN2 memory side-branch with an optional per-token write gate. |
| 2 | |
| 3 | This subclasses fla's `GatedDeltaNet2` and overrides `forward` to apply a |
| 4 | `write_mask` that freezes the recurrence on read tokens. The GDN-2 update is |
| 5 | |
| 6 | S_t = (I - k_t (b_t * k_t)^T) Diag(exp(g_t)) S_{t-1} + k_t (w_t * v_t)^T |
| 7 | |
| 8 | so setting g_t = b_t = w_t = 0 gives S_t = S_{t-1} (read-only); the output |
| 9 | o_t = q_t · S_t is still produced (the token reads the memory). Verified |
| 10 | against `fla.ops.gdn2.naive.naive_recurrent_gdn2`. |
| 11 | |
| 12 | - Design Y: write_mask = is_evicted (open the gate only on the compress/evict |
| 13 | region; live + Q + A are frozen read-only). |
| 14 | - Design X: write_mask = None -> standard continuous scan over all tokens. |
| 15 | """ |
| 16 | from __future__ import annotations |
| 17 | |
| 18 | import torch |
| 19 | import torch.nn.functional as F |
| 20 | from einops import rearrange, repeat |
| 21 | |
| 22 | from fla.layers.gdn2 import GatedDeltaNet2 |
| 23 | from fla.layers.utils import ( |
| 24 | get_layer_cache, |
| 25 | get_unpad_data, |
| 26 | index_first_axis, |
| 27 | pad_input, |
| 28 | update_layer_cache, |
| 29 | ) |
| 30 | from fla.ops.gdn2 import chunk_gdn2, fused_recurrent_gdn2 |
| 31 | |
| 32 | |
| 33 | class LiveMemGatedDeltaNet2(GatedDeltaNet2): |
| 34 | def forward( |
| 35 | self, |
| 36 | hidden_states: torch.Tensor, |
| 37 | write_mask: torch.Tensor | None = None, |
| 38 | attention_mask: torch.Tensor | None = None, |
| 39 | past_key_values=None, |
| 40 | use_cache: bool | None = False, |
| 41 | output_attentions: bool | None = False, |
| 42 | **kwargs, |
| 43 | ): |
| 44 | """Mirrors fla.GatedDeltaNet2.forward, adding `write_mask` support. |
| 45 | |
| 46 | `write_mask`: float/bool [batch, seq_len]; 0 freezes the recurrence |
| 47 | (g=b=w=0 -> S_t = S_{t-1}). Padding (`attention_mask`) + `write_mask` |
| 48 | together are not supported (the unpad reorder would desync the mask); |
| 49 | the training path always passes `attention_mask=None`. |
| 50 | """ |
| 51 | if attention_mask is not None: |
| 52 | assert len(attention_mask.shape) == 2, ( |
| 53 | "Expected attention_mask as a [batch_size, seq_len] 0/1 padding mask." |
| 54 | ) |
| 55 | assert write_mask is None, ( |
| 56 | "write_mask is incompatible with a 2D padding mask (unpad reorder)." |
| 57 | ) |
| 58 | |
| 59 | cu_seqlens = kwargs.get("cu_seqlens") |
| 60 | batch_size, q_len, _ = hidden_states.shape |
| 61 | if cu_seqlens is not None and cu_seqlens.ndim == 2: |
| 62 | if use_cache: |
| 63 | raise ValueError("batched cu_seqlens training path does not support cache") |
| 64 | outs = [] |
| 65 | for b in range(batch_size): |
| 66 | cu = cu_seqlens[b] |
| 67 | cu = cu[cu >= 0].contiguous() |
| 68 | wm = write_mask[b:b + 1] if write_mask is not None else None |
| 69 | o, _, _ = self.forward( |
| 70 | hidden_states[b:b + 1], |
| 71 | write_mask=wm, |
| 72 | attention_mask=None, |
| 73 | past_key_values=None, |
| 74 | use_cache=False, |
| 75 | output_attentions=output_attentions, |
| 76 | cu_seqlens=cu, |
| 77 | ) |
| 78 | outs.append(o) |
| 79 | return torch.cat(outs, dim=0), None, past_key_values |
| 80 | |
| 81 | mode = "fused_recurrent" if (q_len <= 64 and not self.training) else self.mode |
| 82 | if self.training: |
| 83 | assert mode == "chunk", "Only chunk mode is supported in training." |
| 84 | |
| 85 | last_state = get_layer_cache(self, past_key_values) |
| 86 | indices = None |
| 87 | if cu_seqlens is None and attention_mask is not None: |
| 88 | indices, cu_seqlens, _ = get_unpad_data(attention_mask[:, -q_len:]) |
| 89 | hidden_states = index_first_axis( |
| 90 | rearrange(hidden_states, "b s ... -> (b s) ..."), indices |
| 91 | ).unsqueeze(0) |
| 92 | |
| 93 | if self.use_short_conv: |
| 94 | conv_state_q, conv_state_k, conv_state_v = None, None, None |
| 95 | if last_state is not None: |
| 96 | conv_state_q, conv_state_k, conv_state_v = last_state["conv_state"] |
| 97 | q, conv_state_q = self.q_conv1d( |
| 98 | x=self.q_proj(hidden_states), cache=conv_state_q, |
| 99 | output_final_state=use_cache, cu_seqlens=cu_seqlens, |
| 100 | ) |
| 101 | k, conv_state_k = self.k_conv1d( |
| 102 | x=self.k_proj(hidden_states), cache=conv_state_k, |
| 103 | output_final_state=use_cache, cu_seqlens=cu_seqlens, |
| 104 | ) |
| 105 | v, conv_state_v = self.v_conv1d( |
| 106 | x=self.v_proj(hidden_states), cache=conv_state_v, |
| 107 | output_final_state=use_cache, cu_seqlens=cu_seqlens, |
| 108 | ) |
| 109 | else: |
| 110 | q = F.silu(self.q_proj(hidden_states)) |
| 111 | k = F.silu(self.k_proj(hidden_states)) |
| 112 | v = F.silu(self.v_proj(hidden_states)) |
| 113 | |
| 114 | g = F.softplus(self.f_proj(hidden_states).float() + self.dt_bias) |
| 115 | b = self.b_proj(hidden_states).sigmoid() |
| 116 | w = self.w_proj(hidden_states).sigmoid() |
| 117 | |
| 118 | q, k, g = (rearrange(x, "... (h d) -> ... h d", d=self.head_k_dim) for x in (q, k, g)) |
| 119 | v = rearrange(v, "... (h d) -> ... h d", d=self.head_v_dim) |
| 120 | b = rearrange(b, "... (h d) -> ... h d", d=self.head_k_dim) |
| 121 | w = rearrange(w, "... (h d) -> ... h d", d=self.head_v_dim) |
| 122 | g = -self.A_log.float().exp().unsqueeze(-1) * g |
| 123 | |
| 124 | # --- write gate (the only addition over fla) ----------------------- |
| 125 | # Zeroing g/b/w on read tokens freezes the state at S_{t-1}. |
| 126 | if write_mask is not None: |
| 127 | wm = write_mask.to(g.dtype).view(write_mask.shape[0], write_mask.shape[1], 1, 1) |
| 128 | g = g * wm |
| 129 | b = b * wm.to(b.dtype) |
| 130 | w = w * wm.to(w.dtype) |
| 131 | # ------------------------------------------------------------------- |
| 132 | |
| 133 | if self.num_v_heads > self.num_heads: |
| 134 | q, k, g, b = ( |
| 135 | repeat(x, "... h d -> ... (h g) d", g=self.num_v_heads // self.num_heads) |
| 136 | for x in (q, k, g, b) |
| 137 | ) |
| 138 | |
| 139 | if self.allow_neg_eigval: |
| 140 | b = b * 2.0 |
| 141 | |
| 142 | recurrent_state = last_state["recurrent_state"] if last_state is not None else None |
| 143 | if mode == "chunk": |
| 144 | o, recurrent_state = chunk_gdn2( |
| 145 | q=q, k=k, v=v, g=g, b=b, w=w, |
| 146 | initial_state=recurrent_state, output_final_state=use_cache, |
| 147 | use_qk_l2norm_in_kernel=True, cu_seqlens=cu_seqlens, |
| 148 | ) |
| 149 | elif mode == "fused_recurrent": |
| 150 | o, recurrent_state = fused_recurrent_gdn2( |
| 151 | q=q, k=k, v=v, g=g, b=b, w=w, |
| 152 | initial_state=recurrent_state, output_final_state=use_cache, |
| 153 | use_qk_l2norm_in_kernel=True, cu_seqlens=cu_seqlens, |
| 154 | ) |
| 155 | else: |
| 156 | raise NotImplementedError(f"Unsupported mode `{mode}`.") |
| 157 | |
| 158 | update_layer_cache( |
| 159 | self, past_key_values, |
| 160 | recurrent_state=recurrent_state, |
| 161 | conv_state=(conv_state_q, conv_state_k, conv_state_v) if self.use_short_conv else None, |
| 162 | offset=q_len, |
| 163 | ) |
| 164 | |
| 165 | o = self.o_norm(o, rearrange(self.g_proj(hidden_states), "... (h d) -> ... h d", d=self.head_v_dim)) |
| 166 | o = rearrange(o, "b t h d -> b t (h d)") |
| 167 | o = self.o_proj(o) |
| 168 | if attention_mask is not None: |
| 169 | o = pad_input(o.squeeze(0), indices, batch_size, q_len) |
| 170 | |
| 171 | return o, None, past_key_values |
| 172 | |