Update modeling_alinlight.py
Browse files- modeling_alinlight.py +13 -26
modeling_alinlight.py
CHANGED
|
@@ -13,11 +13,6 @@
|
|
| 13 |
# See the License for the specific language governing permissions and
|
| 14 |
# limitations under the License.
|
| 15 |
|
| 16 |
-
# -*- coding: utf-8 -*-
|
| 17 |
-
# Copyright 2026 EngineerGL Research.
|
| 18 |
-
|
| 19 |
-
# -*- coding: utf-8 -*-
|
| 20 |
-
# Copyright 2026 EngineerGL Research.
|
| 21 |
|
| 22 |
import math
|
| 23 |
import torch
|
|
@@ -49,15 +44,15 @@ class AlinlightRotaryEmbedding(nn.Module):
|
|
| 49 |
self.max_position_embeddings = max_position_embeddings
|
| 50 |
self.scaling_factor = scaling_factor
|
| 51 |
|
| 52 |
-
|
| 53 |
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.float32).to(device) / self.dim))
|
| 54 |
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
| 55 |
|
| 56 |
-
|
| 57 |
self._set_cos_sin_cache(seq_len=max_position_embeddings, device=device, dtype=torch.get_default_dtype())
|
| 58 |
|
| 59 |
def _set_cos_sin_cache(self, seq_len, device, dtype):
|
| 60 |
-
|
| 61 |
t = torch.arange(seq_len, device=device, dtype=torch.int64).type_as(self.inv_freq)
|
| 62 |
t = t / self.scaling_factor
|
| 63 |
freqs = torch.outer(t, self.inv_freq)
|
|
@@ -66,7 +61,7 @@ class AlinlightRotaryEmbedding(nn.Module):
|
|
| 66 |
self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
|
| 67 |
|
| 68 |
def forward(self, x, seq_len=None):
|
| 69 |
-
|
| 70 |
if seq_len > self.cos_cached.shape[0]:
|
| 71 |
self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
|
| 72 |
|
|
@@ -145,15 +140,14 @@ class AlinlightAttention(nn.Module):
|
|
| 145 |
key_states = torch.cat([past_key_value[0], key_states], dim=2)
|
| 146 |
value_states = torch.cat([past_key_value[1], value_states], dim=2)
|
| 147 |
|
| 148 |
-
|
| 149 |
-
# Вместо сложной маски просто обрезаем старые токены, вышедшие за пределы окна.
|
| 150 |
if self.sliding_window is not None and key_states.shape[2] > self.sliding_window:
|
| 151 |
key_states = key_states[:, :, -self.sliding_window:, :]
|
| 152 |
value_states = value_states[:, :, -self.sliding_window:, :]
|
| 153 |
|
| 154 |
past_key_value = (key_states, value_states) if use_cache else None
|
| 155 |
|
| 156 |
-
|
| 157 |
if self.num_key_value_groups > 1:
|
| 158 |
key_states = key_states[:, :, None, :, :].expand(
|
| 159 |
bsz, self.num_key_value_heads, self.num_key_value_groups, key_states.shape[-2], self.head_dim
|
|
@@ -163,10 +157,7 @@ class AlinlightAttention(nn.Module):
|
|
| 163 |
bsz, self.num_key_value_heads, self.num_key_value_groups, value_states.shape[-2], self.head_dim
|
| 164 |
).reshape(bsz, self.num_heads, value_states.shape[-2], self.head_dim)
|
| 165 |
|
| 166 |
-
|
| 167 |
-
# is_causal=True нужно ТОЛЬКО при обучении (или prefill), когда q_len > 1.
|
| 168 |
-
# Когда мы генерируем по 1 токену (q_len == 1), мы смотрим на весь кэш (прошлое).
|
| 169 |
-
# Если поставить True при q_len=1, SDPA может замаскировать всё.
|
| 170 |
is_causal = q_len > 1
|
| 171 |
|
| 172 |
attn_output = F.scaled_dot_product_attention(
|
|
@@ -232,17 +223,17 @@ class AlinlightModel(PreTrainedModel):
|
|
| 232 |
else:
|
| 233 |
inputs_embeds = kwargs.get("inputs_embeds")
|
| 234 |
|
| 235 |
-
|
| 236 |
seq_len = inputs_embeds.shape[1]
|
| 237 |
if past_key_values is not None:
|
| 238 |
seq_len += past_key_values[0][0].shape[2]
|
| 239 |
|
| 240 |
-
|
| 241 |
cos, sin = self.rotary_emb(inputs_embeds, seq_len=seq_len)
|
| 242 |
|
| 243 |
position_ids = kwargs.get("position_ids")
|
| 244 |
if position_ids is None:
|
| 245 |
-
|
| 246 |
position_ids = torch.arange(seq_len - inputs_embeds.shape[1], seq_len, dtype=torch.long, device=inputs_embeds.device)
|
| 247 |
position_ids = position_ids.unsqueeze(0).expand(inputs_embeds.shape[0], -1)
|
| 248 |
|
|
@@ -278,10 +269,6 @@ class AlinlightForCausalLM(PreTrainedModel, GenerationMixin):
|
|
| 278 |
self.model = AlinlightModel(config)
|
| 279 |
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
| 280 |
|
| 281 |
-
# === ВАЖНЕЙШЕЕ ИСПРАВЛЕНИЕ ===
|
| 282 |
-
# Связываем веса ТОЛЬКО если это разрешено в конфиге.
|
| 283 |
-
# Раньше это было принудительно, что ломало генерацию для моделей типа Mistral/Llama3,
|
| 284 |
-
# где tie_word_embeddings=False.
|
| 285 |
if config.tie_word_embeddings:
|
| 286 |
self.lm_head.weight = self.model.embed_tokens.weight
|
| 287 |
|
|
@@ -293,18 +280,18 @@ class AlinlightForCausalLM(PreTrainedModel, GenerationMixin):
|
|
| 293 |
def set_output_embeddings(self, new_embeddings): self.lm_head = new_embeddings
|
| 294 |
|
| 295 |
def prepare_inputs_for_generation(self, input_ids, past_key_values=None, **kwargs):
|
| 296 |
-
|
| 297 |
if past_key_values:
|
| 298 |
input_ids = input_ids[:, -1:]
|
| 299 |
|
| 300 |
position_ids = kwargs.get("position_ids", None)
|
| 301 |
if position_ids is None:
|
| 302 |
if past_key_values:
|
| 303 |
-
|
| 304 |
past_length = past_key_values[0][0].shape[2]
|
| 305 |
position_ids = torch.tensor([[past_length]], dtype=torch.long, device=input_ids.device)
|
| 306 |
else:
|
| 307 |
-
|
| 308 |
position_ids = torch.arange(input_ids.shape[1], dtype=torch.long, device=input_ids.device).unsqueeze(0)
|
| 309 |
|
| 310 |
return {
|
|
|
|
| 13 |
# See the License for the specific language governing permissions and
|
| 14 |
# limitations under the License.
|
| 15 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
|
| 17 |
import math
|
| 18 |
import torch
|
|
|
|
| 44 |
self.max_position_embeddings = max_position_embeddings
|
| 45 |
self.scaling_factor = scaling_factor
|
| 46 |
|
| 47 |
+
|
| 48 |
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.float32).to(device) / self.dim))
|
| 49 |
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
| 50 |
|
| 51 |
+
|
| 52 |
self._set_cos_sin_cache(seq_len=max_position_embeddings, device=device, dtype=torch.get_default_dtype())
|
| 53 |
|
| 54 |
def _set_cos_sin_cache(self, seq_len, device, dtype):
|
| 55 |
+
|
| 56 |
t = torch.arange(seq_len, device=device, dtype=torch.int64).type_as(self.inv_freq)
|
| 57 |
t = t / self.scaling_factor
|
| 58 |
freqs = torch.outer(t, self.inv_freq)
|
|
|
|
| 61 |
self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
|
| 62 |
|
| 63 |
def forward(self, x, seq_len=None):
|
| 64 |
+
|
| 65 |
if seq_len > self.cos_cached.shape[0]:
|
| 66 |
self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
|
| 67 |
|
|
|
|
| 140 |
key_states = torch.cat([past_key_value[0], key_states], dim=2)
|
| 141 |
value_states = torch.cat([past_key_value[1], value_states], dim=2)
|
| 142 |
|
| 143 |
+
|
|
|
|
| 144 |
if self.sliding_window is not None and key_states.shape[2] > self.sliding_window:
|
| 145 |
key_states = key_states[:, :, -self.sliding_window:, :]
|
| 146 |
value_states = value_states[:, :, -self.sliding_window:, :]
|
| 147 |
|
| 148 |
past_key_value = (key_states, value_states) if use_cache else None
|
| 149 |
|
| 150 |
+
|
| 151 |
if self.num_key_value_groups > 1:
|
| 152 |
key_states = key_states[:, :, None, :, :].expand(
|
| 153 |
bsz, self.num_key_value_heads, self.num_key_value_groups, key_states.shape[-2], self.head_dim
|
|
|
|
| 157 |
bsz, self.num_key_value_heads, self.num_key_value_groups, value_states.shape[-2], self.head_dim
|
| 158 |
).reshape(bsz, self.num_heads, value_states.shape[-2], self.head_dim)
|
| 159 |
|
| 160 |
+
|
|
|
|
|
|
|
|
|
|
| 161 |
is_causal = q_len > 1
|
| 162 |
|
| 163 |
attn_output = F.scaled_dot_product_attention(
|
|
|
|
| 223 |
else:
|
| 224 |
inputs_embeds = kwargs.get("inputs_embeds")
|
| 225 |
|
| 226 |
+
|
| 227 |
seq_len = inputs_embeds.shape[1]
|
| 228 |
if past_key_values is not None:
|
| 229 |
seq_len += past_key_values[0][0].shape[2]
|
| 230 |
|
| 231 |
+
|
| 232 |
cos, sin = self.rotary_emb(inputs_embeds, seq_len=seq_len)
|
| 233 |
|
| 234 |
position_ids = kwargs.get("position_ids")
|
| 235 |
if position_ids is None:
|
| 236 |
+
|
| 237 |
position_ids = torch.arange(seq_len - inputs_embeds.shape[1], seq_len, dtype=torch.long, device=inputs_embeds.device)
|
| 238 |
position_ids = position_ids.unsqueeze(0).expand(inputs_embeds.shape[0], -1)
|
| 239 |
|
|
|
|
| 269 |
self.model = AlinlightModel(config)
|
| 270 |
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
| 271 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 272 |
if config.tie_word_embeddings:
|
| 273 |
self.lm_head.weight = self.model.embed_tokens.weight
|
| 274 |
|
|
|
|
| 280 |
def set_output_embeddings(self, new_embeddings): self.lm_head = new_embeddings
|
| 281 |
|
| 282 |
def prepare_inputs_for_generation(self, input_ids, past_key_values=None, **kwargs):
|
| 283 |
+
|
| 284 |
if past_key_values:
|
| 285 |
input_ids = input_ids[:, -1:]
|
| 286 |
|
| 287 |
position_ids = kwargs.get("position_ids", None)
|
| 288 |
if position_ids is None:
|
| 289 |
if past_key_values:
|
| 290 |
+
|
| 291 |
past_length = past_key_values[0][0].shape[2]
|
| 292 |
position_ids = torch.tensor([[past_length]], dtype=torch.long, device=input_ids.device)
|
| 293 |
else:
|
| 294 |
+
|
| 295 |
position_ids = torch.arange(input_ids.shape[1], dtype=torch.long, device=input_ids.device).unsqueeze(0)
|
| 296 |
|
| 297 |
return {
|