Initial model upload with custom modeling and generation code
Browse files- config.json +2 -1
- modeling_qwen2.py +81 -203
config.json
CHANGED
@@ -28,5 +28,6 @@
|
|
28 |
"vocab_size": 151936,
|
29 |
"auto_map": {
|
30 |
"AutoModelForCausalLM": "modeling_qwen2.Qwen2ForCausalLM"
|
31 |
-
}
|
|
|
32 |
}
|
|
|
28 |
"vocab_size": 151936,
|
29 |
"auto_map": {
|
30 |
"AutoModelForCausalLM": "modeling_qwen2.Qwen2ForCausalLM"
|
31 |
+
},
|
32 |
+
"trust_remote_code": true
|
33 |
}
|
modeling_qwen2.py
CHANGED
@@ -12,17 +12,16 @@
|
|
12 |
# See the License for the specific language governing permissions and
|
13 |
# limitations under the License.
|
14 |
|
15 |
-
# This is a cleaned version of the
|
16 |
-
#
|
17 |
|
18 |
import logging
|
19 |
-
from typing import Callable, List, Optional, Tuple, Union
|
20 |
|
21 |
import torch
|
22 |
from torch import nn
|
23 |
from transformers.activations import ACT2FN
|
24 |
from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
|
25 |
-
from transformers.generation import GenerationMixin
|
26 |
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
|
27 |
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
|
28 |
from transformers.modeling_outputs import (
|
@@ -39,6 +38,9 @@ from transformers.utils import (
|
|
39 |
replace_return_docstrings,
|
40 |
)
|
41 |
|
|
|
|
|
|
|
42 |
logger = logging.getLogger(__name__)
|
43 |
|
44 |
_CHECKPOINT_FOR_DOC = "meta-qwen2/Qwen2-2-7b-hf"
|
@@ -88,35 +90,8 @@ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
|
88 |
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
89 |
|
90 |
|
91 |
-
def eager_attention_forward(
|
92 |
-
module: nn.Module,
|
93 |
-
query: torch.Tensor,
|
94 |
-
key: torch.Tensor,
|
95 |
-
value: torch.Tensor,
|
96 |
-
attention_mask: Optional[torch.Tensor],
|
97 |
-
scaling: float,
|
98 |
-
dropout: float = 0.0,
|
99 |
-
**kwargs,
|
100 |
-
):
|
101 |
-
key_states = repeat_kv(key, module.num_key_value_groups)
|
102 |
-
value_states = repeat_kv(value, module.num_key_value_groups)
|
103 |
-
|
104 |
-
attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
|
105 |
-
if attention_mask is not None:
|
106 |
-
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
107 |
-
attn_weights = attn_weights + causal_mask
|
108 |
-
|
109 |
-
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
|
110 |
-
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|
111 |
-
attn_output = torch.matmul(attn_weights, value_states)
|
112 |
-
attn_output = attn_output.transpose(1, 2).contiguous()
|
113 |
-
|
114 |
-
return attn_output, attn_weights
|
115 |
-
|
116 |
-
|
117 |
class Qwen2Attention(nn.Module):
|
118 |
-
|
119 |
-
|
120 |
def __init__(self, config: Qwen2Config, layer_idx: int):
|
121 |
super().__init__()
|
122 |
self.config = config
|
@@ -136,6 +111,7 @@ class Qwen2Attention(nn.Module):
|
|
136 |
position_embeddings: Tuple[torch.Tensor, torch.Tensor],
|
137 |
attention_mask: Optional[torch.Tensor],
|
138 |
past_key_value: Optional[Cache] = None,
|
|
|
139 |
cache_position: Optional[torch.LongTensor] = None,
|
140 |
is_causal: bool = True,
|
141 |
**kwargs: Unpack[FlashAttentionKwargs],
|
@@ -155,42 +131,34 @@ class Qwen2Attention(nn.Module):
|
|
155 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
156 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
157 |
|
158 |
-
|
159 |
-
if
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
if self.config._attn_implementation != "eager":
|
168 |
-
if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
|
169 |
-
logger.warning_once(
|
170 |
-
"`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
|
171 |
-
'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
|
172 |
-
)
|
173 |
-
else:
|
174 |
-
attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
|
175 |
-
self.is_causal = is_causal
|
176 |
attn_output, attn_weights = attention_interface(
|
177 |
-
self,
|
178 |
query_states,
|
179 |
key_states,
|
180 |
value_states,
|
181 |
-
attention_mask,
|
182 |
-
dropout=
|
183 |
-
|
184 |
-
sliding_window=sliding_window,
|
185 |
**kwargs,
|
186 |
)
|
187 |
-
|
188 |
-
attn_output = attn_output.reshape(bsz,
|
189 |
-
attn_output = attn_output.reshape(bsz, q_len, self.config.hidden_size).contiguous()
|
190 |
attn_output = self.o_proj(attn_output)
|
191 |
-
|
|
|
|
|
192 |
|
|
|
193 |
|
|
|
194 |
class Qwen2RMSNorm(nn.Module):
|
195 |
def __init__(self, hidden_size, eps=1e-6):
|
196 |
super().__init__()
|
@@ -204,10 +172,6 @@ class Qwen2RMSNorm(nn.Module):
|
|
204 |
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
205 |
return self.weight * hidden_states.to(input_dtype)
|
206 |
|
207 |
-
def extra_repr(self):
|
208 |
-
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
209 |
-
|
210 |
-
|
211 |
class Qwen2DecoderLayer(nn.Module):
|
212 |
def __init__(self, config: Qwen2Config, layer_idx: int):
|
213 |
super().__init__()
|
@@ -216,11 +180,6 @@ class Qwen2DecoderLayer(nn.Module):
|
|
216 |
self.mlp = Qwen2MLP(config)
|
217 |
self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
218 |
self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
219 |
-
if config.sliding_window and config._attn_implementation != "flash_attention_2":
|
220 |
-
logger.warning_once(
|
221 |
-
f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
|
222 |
-
"unexpected results may be encountered."
|
223 |
-
)
|
224 |
|
225 |
def forward(
|
226 |
self,
|
@@ -238,10 +197,11 @@ class Qwen2DecoderLayer(nn.Module):
|
|
238 |
residual = hidden_states
|
239 |
hidden_states = self.input_layernorm(hidden_states)
|
240 |
|
241 |
-
hidden_states, self_attn_weights = self.self_attn(
|
242 |
hidden_states=hidden_states,
|
243 |
attention_mask=attention_mask,
|
244 |
past_key_value=past_key_value,
|
|
|
245 |
cache_position=cache_position,
|
246 |
position_embeddings=position_embeddings,
|
247 |
is_causal=is_causal,
|
@@ -257,10 +217,11 @@ class Qwen2DecoderLayer(nn.Module):
|
|
257 |
outputs = (hidden_states,)
|
258 |
if output_attentions:
|
259 |
outputs += (self_attn_weights,)
|
|
|
|
|
260 |
|
261 |
return outputs
|
262 |
|
263 |
-
|
264 |
class Qwen2RotaryEmbedding(nn.Module):
|
265 |
def __init__(self, config: Qwen2Config, device=None):
|
266 |
super().__init__()
|
@@ -304,24 +265,6 @@ class Qwen2RotaryEmbedding(nn.Module):
|
|
304 |
sin = sin * self.attention_scaling
|
305 |
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
306 |
|
307 |
-
|
308 |
-
QWEN2_START_DOCSTRING = r"""
|
309 |
-
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
310 |
-
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
311 |
-
etc.)
|
312 |
-
|
313 |
-
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
314 |
-
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
315 |
-
and behavior.
|
316 |
-
|
317 |
-
Parameters:
|
318 |
-
config ([`Qwen2Config`]):
|
319 |
-
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
320 |
-
load the weights associated with the model, only the configuration. Check out the
|
321 |
-
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
322 |
-
"""
|
323 |
-
|
324 |
-
|
325 |
@add_start_docstrings(
|
326 |
"The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
|
327 |
QWEN2_START_DOCSTRING,
|
@@ -334,11 +277,7 @@ class Qwen2PreTrainedModel(PreTrainedModel):
|
|
334 |
_skip_keys_device_placement = ["past_key_values"]
|
335 |
_supports_flash_attn_2 = True
|
336 |
_supports_sdpa = True
|
337 |
-
_supports_flex_attn = True
|
338 |
_supports_cache_class = True
|
339 |
-
_supports_quantized_cache = True
|
340 |
-
_supports_static_cache = True
|
341 |
-
_supports_attention_backend = True
|
342 |
|
343 |
def _init_weights(self, module):
|
344 |
std = self.config.initializer_range
|
@@ -351,36 +290,6 @@ class Qwen2PreTrainedModel(PreTrainedModel):
|
|
351 |
if module.padding_idx is not None:
|
352 |
module.weight.data[module.padding_idx].zero_()
|
353 |
|
354 |
-
|
355 |
-
QWEN2_INPUTS_DOCSTRING = r"""
|
356 |
-
Args:
|
357 |
-
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
358 |
-
Indices of input sequence tokens in the vocabulary.
|
359 |
-
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
360 |
-
Mask to avoid performing attention on padding token indices.
|
361 |
-
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
362 |
-
Indices of positions of each input sequence tokens in the position embeddings.
|
363 |
-
past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
|
364 |
-
Pre-computed hidden-states (key and values in the self-attention blocks) that can be used to speed up sequential decoding.
|
365 |
-
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
366 |
-
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
|
367 |
-
use_cache (`bool`, *optional*):
|
368 |
-
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding.
|
369 |
-
output_attentions (`bool`, *optional*):
|
370 |
-
Whether or not to return the attentions tensors of all attention layers.
|
371 |
-
output_hidden_states (`bool`, *optional*):
|
372 |
-
Whether or not to return the hidden states of all layers.
|
373 |
-
return_dict (`bool`, *optional*):
|
374 |
-
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
375 |
-
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
376 |
-
Indices depicting the position of the input sequence tokens in the sequence.
|
377 |
-
"""
|
378 |
-
|
379 |
-
|
380 |
-
@add_start_docstrings(
|
381 |
-
"The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
|
382 |
-
QWEN2_START_DOCSTRING,
|
383 |
-
)
|
384 |
class Qwen2Model(Qwen2PreTrainedModel):
|
385 |
def __init__(self, config: Qwen2Config):
|
386 |
super().__init__(config)
|
@@ -401,7 +310,6 @@ class Qwen2Model(Qwen2PreTrainedModel):
|
|
401 |
def set_input_embeddings(self, value):
|
402 |
self.embed_tokens = value
|
403 |
|
404 |
-
@add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
|
405 |
def forward(
|
406 |
self,
|
407 |
input_ids: torch.LongTensor = None,
|
@@ -431,114 +339,87 @@ class Qwen2Model(Qwen2PreTrainedModel):
|
|
431 |
use_cache = False
|
432 |
if inputs_embeds is None:
|
433 |
inputs_embeds = self.embed_tokens(input_ids)
|
434 |
-
|
435 |
-
|
|
|
|
|
|
|
|
|
|
|
436 |
if cache_position is None:
|
437 |
-
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
438 |
cache_position = torch.arange(
|
439 |
-
|
440 |
)
|
441 |
if position_ids is None:
|
442 |
position_ids = cache_position.unsqueeze(0)
|
443 |
-
|
444 |
-
|
445 |
-
)
|
446 |
hidden_states = inputs_embeds
|
447 |
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
448 |
all_hidden_states = () if output_hidden_states else None
|
449 |
all_self_attns = () if output_attentions else None
|
|
|
450 |
|
451 |
for decoder_layer in self.layers:
|
452 |
if output_hidden_states:
|
453 |
all_hidden_states += (hidden_states,)
|
454 |
-
|
455 |
-
|
456 |
-
|
457 |
-
|
458 |
-
|
459 |
-
|
460 |
-
|
461 |
-
|
462 |
-
|
463 |
-
|
464 |
-
|
465 |
-
|
466 |
-
|
467 |
-
else:
|
468 |
-
layer_outputs = decoder_layer(
|
469 |
-
hidden_states,
|
470 |
-
attention_mask=causal_mask,
|
471 |
-
position_ids=position_ids,
|
472 |
-
past_key_value=past_key_values,
|
473 |
-
output_attentions=output_attentions,
|
474 |
-
use_cache=use_cache,
|
475 |
-
cache_position=cache_position,
|
476 |
-
position_embeddings=position_embeddings,
|
477 |
-
is_causal=is_causal,
|
478 |
-
**flash_attn_kwargs,
|
479 |
-
)
|
480 |
hidden_states = layer_outputs[0]
|
|
|
|
|
481 |
if output_attentions:
|
482 |
all_self_attns += (layer_outputs[1],)
|
483 |
|
484 |
hidden_states = self.norm(hidden_states)
|
485 |
if output_hidden_states:
|
486 |
all_hidden_states += (hidden_states,)
|
487 |
-
|
|
|
|
|
|
|
|
|
|
|
488 |
last_hidden_state=hidden_states,
|
489 |
-
past_key_values=
|
490 |
hidden_states=all_hidden_states,
|
491 |
attentions=all_self_attns,
|
492 |
)
|
493 |
-
return output if return_dict else output.to_tuple()
|
494 |
|
495 |
-
def _update_causal_mask(
|
496 |
-
|
497 |
-
|
498 |
-
|
499 |
-
|
500 |
-
past_key_values: Cache,
|
501 |
-
output_attentions: bool,
|
502 |
-
):
|
503 |
-
# Standard causal mask creation logic from transformers, no changes needed here.
|
504 |
if self.config._attn_implementation == "flash_attention_2":
|
505 |
if attention_mask is not None and 0.0 in attention_mask:
|
506 |
return attention_mask
|
507 |
return None
|
508 |
-
|
509 |
-
|
510 |
-
|
511 |
-
|
512 |
-
|
513 |
-
):
|
514 |
-
return None
|
515 |
-
dtype, device = input_tensor.dtype, input_tensor.device
|
516 |
-
min_dtype = torch.finfo(dtype).min
|
517 |
-
sequence_length = input_tensor.shape[1]
|
518 |
-
if isinstance(past_key_values, StaticCache):
|
519 |
-
target_length = past_key_values.get_max_cache_shape()
|
520 |
-
else:
|
521 |
-
target_length = attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else past_seen_tokens + sequence_length
|
522 |
-
causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
|
523 |
-
if sequence_length != 1:
|
524 |
-
causal_mask = torch.triu(causal_mask, diagonal=1)
|
525 |
-
causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
|
526 |
causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
|
|
|
527 |
if attention_mask is not None:
|
528 |
-
|
529 |
-
|
530 |
-
|
531 |
-
padding_mask = padding_mask == 0
|
532 |
-
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(padding_mask, min_dtype)
|
533 |
-
if self.config._attn_implementation == "sdpa" and attention_mask is not None and attention_mask.device.type == "cuda" and not output_attentions:
|
534 |
-
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
|
535 |
return causal_mask
|
536 |
|
537 |
-
|
538 |
-
class KwargsForCausalLM(FlashAttentionKwargs, ): ...
|
539 |
-
|
540 |
-
|
541 |
-
class Qwen2ForCausalLM(Qwen2PreTrainedModel, GenerationMixin):
|
542 |
_tied_weights_keys = ["lm_head.weight"]
|
543 |
|
544 |
def __init__(self, config):
|
@@ -573,7 +454,7 @@ class Qwen2ForCausalLM(Qwen2PreTrainedModel, GenerationMixin):
|
|
573 |
input_ids: torch.LongTensor = None,
|
574 |
attention_mask: Optional[torch.Tensor] = None,
|
575 |
position_ids: Optional[torch.LongTensor] = None,
|
576 |
-
past_key_values: Optional[
|
577 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
578 |
labels: Optional[torch.LongTensor] = None,
|
579 |
use_cache: Optional[bool] = None,
|
@@ -582,7 +463,7 @@ class Qwen2ForCausalLM(Qwen2PreTrainedModel, GenerationMixin):
|
|
582 |
return_dict: Optional[bool] = None,
|
583 |
cache_position: Optional[torch.LongTensor] = None,
|
584 |
is_causal: bool = True,
|
585 |
-
**kwargs: Unpack[
|
586 |
) -> Union[Tuple, CausalLMOutputWithPast]:
|
587 |
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
588 |
output_hidden_states = (
|
@@ -611,14 +492,11 @@ class Qwen2ForCausalLM(Qwen2PreTrainedModel, GenerationMixin):
|
|
611 |
loss = None
|
612 |
|
613 |
if labels is not None:
|
614 |
-
# Shift so that tokens < n predict n
|
615 |
shift_logits = logits[..., :-1, :].contiguous()
|
616 |
shift_labels = labels[..., 1:].contiguous()
|
617 |
-
# Flatten the tokens
|
618 |
loss_fct = torch.nn.CrossEntropyLoss()
|
619 |
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
620 |
shift_labels = shift_labels.view(-1)
|
621 |
-
# Ensure labels are on the same device as logits
|
622 |
shift_labels = shift_labels.to(shift_logits.device)
|
623 |
loss = loss_fct(shift_logits, shift_labels)
|
624 |
|
|
|
12 |
# See the License for the specific language governing permissions and
|
13 |
# limitations under the License.
|
14 |
|
15 |
+
# This is a cleaned version of the model script for public release.
|
16 |
+
# It imports the MDMGenerationMixin from the accompanying generation_utils.py file.
|
17 |
|
18 |
import logging
|
19 |
+
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
|
20 |
|
21 |
import torch
|
22 |
from torch import nn
|
23 |
from transformers.activations import ACT2FN
|
24 |
from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
|
|
|
25 |
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
|
26 |
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
|
27 |
from transformers.modeling_outputs import (
|
|
|
38 |
replace_return_docstrings,
|
39 |
)
|
40 |
|
41 |
+
# Import the custom generation mixin from the local file in the repo
|
42 |
+
from .generation_utils import MDMGenerationMixin
|
43 |
+
|
44 |
logger = logging.getLogger(__name__)
|
45 |
|
46 |
_CHECKPOINT_FOR_DOC = "meta-qwen2/Qwen2-2-7b-hf"
|
|
|
90 |
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
91 |
|
92 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
93 |
class Qwen2Attention(nn.Module):
|
94 |
+
# ... (rest of the class is unchanged)
|
|
|
95 |
def __init__(self, config: Qwen2Config, layer_idx: int):
|
96 |
super().__init__()
|
97 |
self.config = config
|
|
|
111 |
position_embeddings: Tuple[torch.Tensor, torch.Tensor],
|
112 |
attention_mask: Optional[torch.Tensor],
|
113 |
past_key_value: Optional[Cache] = None,
|
114 |
+
output_attentions: Optional[bool] = False,
|
115 |
cache_position: Optional[torch.LongTensor] = None,
|
116 |
is_causal: bool = True,
|
117 |
**kwargs: Unpack[FlashAttentionKwargs],
|
|
|
131 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
132 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
133 |
|
134 |
+
attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get(self.config._attn_implementation, None)
|
135 |
+
if attention_interface is None:
|
136 |
+
raise ValueError(f"Attention implementation {self.config._attn_implementation} not found.")
|
137 |
+
|
138 |
+
if self.config._attn_implementation == "sdpa" and output_attentions:
|
139 |
+
logger.warning_once("Using SDPA with `output_attentions=True` requires eager attention.")
|
140 |
+
attention_interface = ALL_ATTENTION_FUNCTIONS["eager"]
|
141 |
+
|
142 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
143 |
attn_output, attn_weights = attention_interface(
|
|
|
144 |
query_states,
|
145 |
key_states,
|
146 |
value_states,
|
147 |
+
attention_mask=attention_mask,
|
148 |
+
dropout=self.attention_dropout if self.training else 0.0,
|
149 |
+
is_causal=is_causal,
|
|
|
150 |
**kwargs,
|
151 |
)
|
152 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
153 |
+
attn_output = attn_output.reshape(bsz, q_len, self.config.hidden_size)
|
|
|
154 |
attn_output = self.o_proj(attn_output)
|
155 |
+
|
156 |
+
if not output_attentions:
|
157 |
+
attn_weights = None
|
158 |
|
159 |
+
return attn_output, attn_weights, past_key_value
|
160 |
|
161 |
+
# ... (Qwen2RMSNorm, Qwen2DecoderLayer, Qwen2RotaryEmbedding, Qwen2PreTrainedModel, Qwen2Model are unchanged)
|
162 |
class Qwen2RMSNorm(nn.Module):
|
163 |
def __init__(self, hidden_size, eps=1e-6):
|
164 |
super().__init__()
|
|
|
172 |
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
173 |
return self.weight * hidden_states.to(input_dtype)
|
174 |
|
|
|
|
|
|
|
|
|
175 |
class Qwen2DecoderLayer(nn.Module):
|
176 |
def __init__(self, config: Qwen2Config, layer_idx: int):
|
177 |
super().__init__()
|
|
|
180 |
self.mlp = Qwen2MLP(config)
|
181 |
self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
182 |
self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
|
|
|
|
|
|
|
|
|
183 |
|
184 |
def forward(
|
185 |
self,
|
|
|
197 |
residual = hidden_states
|
198 |
hidden_states = self.input_layernorm(hidden_states)
|
199 |
|
200 |
+
hidden_states, self_attn_weights, present_key_value = self.self_attn(
|
201 |
hidden_states=hidden_states,
|
202 |
attention_mask=attention_mask,
|
203 |
past_key_value=past_key_value,
|
204 |
+
output_attentions=output_attentions,
|
205 |
cache_position=cache_position,
|
206 |
position_embeddings=position_embeddings,
|
207 |
is_causal=is_causal,
|
|
|
217 |
outputs = (hidden_states,)
|
218 |
if output_attentions:
|
219 |
outputs += (self_attn_weights,)
|
220 |
+
if use_cache:
|
221 |
+
outputs += (present_key_value,)
|
222 |
|
223 |
return outputs
|
224 |
|
|
|
225 |
class Qwen2RotaryEmbedding(nn.Module):
|
226 |
def __init__(self, config: Qwen2Config, device=None):
|
227 |
super().__init__()
|
|
|
265 |
sin = sin * self.attention_scaling
|
266 |
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
267 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
268 |
@add_start_docstrings(
|
269 |
"The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
|
270 |
QWEN2_START_DOCSTRING,
|
|
|
277 |
_skip_keys_device_placement = ["past_key_values"]
|
278 |
_supports_flash_attn_2 = True
|
279 |
_supports_sdpa = True
|
|
|
280 |
_supports_cache_class = True
|
|
|
|
|
|
|
281 |
|
282 |
def _init_weights(self, module):
|
283 |
std = self.config.initializer_range
|
|
|
290 |
if module.padding_idx is not None:
|
291 |
module.weight.data[module.padding_idx].zero_()
|
292 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
293 |
class Qwen2Model(Qwen2PreTrainedModel):
|
294 |
def __init__(self, config: Qwen2Config):
|
295 |
super().__init__(config)
|
|
|
310 |
def set_input_embeddings(self, value):
|
311 |
self.embed_tokens = value
|
312 |
|
|
|
313 |
def forward(
|
314 |
self,
|
315 |
input_ids: torch.LongTensor = None,
|
|
|
339 |
use_cache = False
|
340 |
if inputs_embeds is None:
|
341 |
inputs_embeds = self.embed_tokens(input_ids)
|
342 |
+
|
343 |
+
past_key_values_length = 0
|
344 |
+
if use_cache:
|
345 |
+
if past_key_values is None:
|
346 |
+
past_key_values = DynamicCache()
|
347 |
+
past_key_values_length = past_key_values.get_seq_length()
|
348 |
+
|
349 |
if cache_position is None:
|
|
|
350 |
cache_position = torch.arange(
|
351 |
+
past_key_values_length, past_key_values_length + inputs_embeds.shape[1], device=inputs_embeds.device
|
352 |
)
|
353 |
if position_ids is None:
|
354 |
position_ids = cache_position.unsqueeze(0)
|
355 |
+
|
356 |
+
causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, is_causal)
|
|
|
357 |
hidden_states = inputs_embeds
|
358 |
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
359 |
all_hidden_states = () if output_hidden_states else None
|
360 |
all_self_attns = () if output_attentions else None
|
361 |
+
next_decoder_cache = () if use_cache else None
|
362 |
|
363 |
for decoder_layer in self.layers:
|
364 |
if output_hidden_states:
|
365 |
all_hidden_states += (hidden_states,)
|
366 |
+
|
367 |
+
layer_outputs = decoder_layer(
|
368 |
+
hidden_states,
|
369 |
+
attention_mask=causal_mask,
|
370 |
+
position_ids=position_ids,
|
371 |
+
past_key_value=past_key_values,
|
372 |
+
output_attentions=output_attentions,
|
373 |
+
use_cache=use_cache,
|
374 |
+
cache_position=cache_position,
|
375 |
+
position_embeddings=position_embeddings,
|
376 |
+
is_causal=is_causal,
|
377 |
+
**flash_attn_kwargs,
|
378 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
379 |
hidden_states = layer_outputs[0]
|
380 |
+
if use_cache:
|
381 |
+
next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
|
382 |
if output_attentions:
|
383 |
all_self_attns += (layer_outputs[1],)
|
384 |
|
385 |
hidden_states = self.norm(hidden_states)
|
386 |
if output_hidden_states:
|
387 |
all_hidden_states += (hidden_states,)
|
388 |
+
|
389 |
+
next_cache = next_decoder_cache if use_cache else None
|
390 |
+
|
391 |
+
if not return_dict:
|
392 |
+
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
|
393 |
+
return BaseModelOutputWithPast(
|
394 |
last_hidden_state=hidden_states,
|
395 |
+
past_key_values=next_cache,
|
396 |
hidden_states=all_hidden_states,
|
397 |
attentions=all_self_attns,
|
398 |
)
|
|
|
399 |
|
400 |
+
def _update_causal_mask(self, attention_mask, input_tensor, cache_position, is_causal):
|
401 |
+
if not is_causal:
|
402 |
+
return attention_mask
|
403 |
+
|
404 |
+
seq_len = input_tensor.shape[1]
|
|
|
|
|
|
|
|
|
405 |
if self.config._attn_implementation == "flash_attention_2":
|
406 |
if attention_mask is not None and 0.0 in attention_mask:
|
407 |
return attention_mask
|
408 |
return None
|
409 |
+
|
410 |
+
dtype = input_tensor.dtype
|
411 |
+
device = input_tensor.device
|
412 |
+
|
413 |
+
causal_mask = torch.triu(torch.full((seq_len, seq_len), torch.finfo(dtype).min, device=device), 1)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
414 |
causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
|
415 |
+
|
416 |
if attention_mask is not None:
|
417 |
+
causal_mask = causal_mask.clone()
|
418 |
+
causal_mask = causal_mask + attention_mask[:, None, None, :]
|
419 |
+
|
|
|
|
|
|
|
|
|
420 |
return causal_mask
|
421 |
|
422 |
+
class Qwen2ForCausalLM(Qwen2PreTrainedModel, MDMGenerationMixin):
|
|
|
|
|
|
|
|
|
423 |
_tied_weights_keys = ["lm_head.weight"]
|
424 |
|
425 |
def __init__(self, config):
|
|
|
454 |
input_ids: torch.LongTensor = None,
|
455 |
attention_mask: Optional[torch.Tensor] = None,
|
456 |
position_ids: Optional[torch.LongTensor] = None,
|
457 |
+
past_key_values: Optional[Cache] = None,
|
458 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
459 |
labels: Optional[torch.LongTensor] = None,
|
460 |
use_cache: Optional[bool] = None,
|
|
|
463 |
return_dict: Optional[bool] = None,
|
464 |
cache_position: Optional[torch.LongTensor] = None,
|
465 |
is_causal: bool = True,
|
466 |
+
**kwargs: Unpack[FlashAttentionKwargs],
|
467 |
) -> Union[Tuple, CausalLMOutputWithPast]:
|
468 |
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
469 |
output_hidden_states = (
|
|
|
492 |
loss = None
|
493 |
|
494 |
if labels is not None:
|
|
|
495 |
shift_logits = logits[..., :-1, :].contiguous()
|
496 |
shift_labels = labels[..., 1:].contiguous()
|
|
|
497 |
loss_fct = torch.nn.CrossEntropyLoss()
|
498 |
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
499 |
shift_labels = shift_labels.view(-1)
|
|
|
500 |
shift_labels = shift_labels.to(shift_logits.device)
|
501 |
loss = loss_fct(shift_logits, shift_labels)
|
502 |
|