Update modeling_helpingai.py
Browse files- modeling_helpingai.py +1020 -969
modeling_helpingai.py
CHANGED
@@ -1,969 +1,1020 @@
|
|
1 |
-
from typing import Callable, Optional, Union
|
2 |
-
|
3 |
-
import torch
|
4 |
-
from torch import nn
|
5 |
-
|
6 |
-
from transformers.activations import ACT2FN
|
7 |
-
from transformers.cache_utils import Cache, DynamicCache
|
8 |
-
from transformers.generation import GenerationMixin
|
9 |
-
from transformers.integrations import use_kernel_forward_from_hub
|
10 |
-
from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask
|
11 |
-
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
|
12 |
-
from transformers.modeling_layers import (
|
13 |
-
GenericForQuestionAnswering,
|
14 |
-
GenericForSequenceClassification,
|
15 |
-
GenericForTokenClassification,
|
16 |
-
GradientCheckpointingLayer,
|
17 |
-
)
|
18 |
-
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
|
19 |
-
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
|
20 |
-
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
|
21 |
-
from transformers.processing_utils import Unpack
|
22 |
-
from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple
|
23 |
-
from transformers.utils.deprecation import deprecate_kwarg
|
24 |
-
from transformers.utils.generic import check_model_inputs
|
25 |
-
from .configuration_helpingai import HelpingAIConfig
|
26 |
-
|
27 |
-
|
28 |
-
@use_kernel_forward_from_hub("RMSNorm")
|
29 |
-
class HelpingAIRMSNorm(nn.Module):
|
30 |
-
def __init__(self, hidden_size, eps=1e-6):
|
31 |
-
"""
|
32 |
-
HelpingAIRMSNorm is equivalent to T5LayerNorm
|
33 |
-
"""
|
34 |
-
super().__init__()
|
35 |
-
self.weight = nn.Parameter(torch.ones(hidden_size))
|
36 |
-
self.variance_epsilon = eps
|
37 |
-
|
38 |
-
def forward(self, hidden_states):
|
39 |
-
input_dtype = hidden_states.dtype
|
40 |
-
hidden_states = hidden_states.to(torch.float32)
|
41 |
-
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
42 |
-
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
43 |
-
return self.weight * hidden_states.to(input_dtype)
|
44 |
-
|
45 |
-
def extra_repr(self):
|
46 |
-
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
47 |
-
|
48 |
-
|
49 |
-
class HelpingAISemanticEmotionReasoning(nn.Module):
|
50 |
-
"""
|
51 |
-
Structured Emotional Reasoning (SER) layer for emotional understanding and processing.
|
52 |
-
Maps emotions to semantic representations and provides contextual emotion analysis.
|
53 |
-
"""
|
54 |
-
def __init__(self, config: HelpingAIConfig):
|
55 |
-
super().__init__()
|
56 |
-
self.config = config
|
57 |
-
self.emotion_hidden_size = config.emotion_hidden_size
|
58 |
-
self.hidden_size = config.hidden_size
|
59 |
-
|
60 |
-
# Emotion detection and mapping
|
61 |
-
self.emotion_detector = nn.Linear(self.hidden_size, self.emotion_hidden_size)
|
62 |
-
self.emotion_mapper = nn.Linear(self.emotion_hidden_size, self.emotion_hidden_size)
|
63 |
-
|
64 |
-
# Contextual emotion analysis
|
65 |
-
self.emotion_context = nn.MultiheadAttention(
|
66 |
-
embed_dim=self.emotion_hidden_size,
|
67 |
-
num_heads=min(8, self.emotion_hidden_size // 64),
|
68 |
-
batch_first=True
|
69 |
-
)
|
70 |
-
|
71 |
-
# Emotion classification heads
|
72 |
-
self.primary_emotion = nn.Linear(self.emotion_hidden_size, 32) # Primary emotions
|
73 |
-
self.emotion_intensity = nn.Linear(self.emotion_hidden_size, 1) # Intensity score
|
74 |
-
self.emotion_valence = nn.Linear(self.emotion_hidden_size, 1) # Positive/negative
|
75 |
-
|
76 |
-
# Output projection
|
77 |
-
self.emotion_output = nn.Linear(self.emotion_hidden_size, self.hidden_size)
|
78 |
-
self.emotion_norm = HelpingAIRMSNorm(self.emotion_hidden_size, eps=config.rms_norm_eps)
|
79 |
-
|
80 |
-
# Activation
|
81 |
-
self.act_fn = ACT2FN[config.hidden_act]
|
82 |
-
|
83 |
-
def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, dict]:
|
84 |
-
# Detect emotional content
|
85 |
-
emotion_features = self.act_fn(self.emotion_detector(hidden_states))
|
86 |
-
emotion_mapped = self.emotion_mapper(emotion_features)
|
87 |
-
emotion_mapped = self.emotion_norm(emotion_mapped)
|
88 |
-
|
89 |
-
# Contextual emotion analysis
|
90 |
-
emotion_context, attention_weights = self.emotion_context(
|
91 |
-
emotion_mapped, emotion_mapped, emotion_mapped
|
92 |
-
)
|
93 |
-
|
94 |
-
# Emotion analysis outputs
|
95 |
-
primary_emotions = self.primary_emotion(emotion_context)
|
96 |
-
emotion_intensity = torch.sigmoid(self.emotion_intensity(emotion_context))
|
97 |
-
emotion_valence = torch.tanh(self.emotion_valence(emotion_context))
|
98 |
-
|
99 |
-
# Project back to hidden size
|
100 |
-
emotion_output = self.emotion_output(emotion_context)
|
101 |
-
|
102 |
-
# Emotion metadata
|
103 |
-
emotion_metadata = {
|
104 |
-
"primary_emotions": primary_emotions,
|
105 |
-
"intensity": emotion_intensity,
|
106 |
-
"valence": emotion_valence,
|
107 |
-
"attention_weights": attention_weights
|
108 |
-
}
|
109 |
-
|
110 |
-
return emotion_output, emotion_metadata
|
111 |
-
|
112 |
-
|
113 |
-
class HelpingAIPerspectiveEmotionThreading(nn.Module):
|
114 |
-
"""
|
115 |
-
Parallel Empathic Threads (PET) layer for multi-threaded emotional reasoning.
|
116 |
-
Processes multiple perspective threads: relatable, supportive, motivational, analytical.
|
117 |
-
"""
|
118 |
-
def __init__(self, config: HelpingAIConfig):
|
119 |
-
super().__init__()
|
120 |
-
self.config = config
|
121 |
-
self.hidden_size = config.hidden_size
|
122 |
-
self.perspective_threads = config.perspective_threads
|
123 |
-
self.thread_hidden_size = config.emotion_hidden_size
|
124 |
-
|
125 |
-
# Thread-specific processors
|
126 |
-
self.thread_projections = nn.ModuleList([
|
127 |
-
nn.Linear(self.hidden_size, self.thread_hidden_size)
|
128 |
-
for _ in range(self.perspective_threads)
|
129 |
-
])
|
130 |
-
|
131 |
-
# Thread names for interpretability
|
132 |
-
self.thread_names = ["relatable", "supportive", "motivational", "analytical"][:self.perspective_threads]
|
133 |
-
|
134 |
-
# Cross-thread attention for perspective integration
|
135 |
-
self.cross_thread_attention = nn.MultiheadAttention(
|
136 |
-
embed_dim=self.thread_hidden_size,
|
137 |
-
num_heads=min(4, self.thread_hidden_size // 64),
|
138 |
-
batch_first=True
|
139 |
-
)
|
140 |
-
|
141 |
-
# Thread-specific processing layers
|
142 |
-
self.thread_processors = nn.ModuleList([
|
143 |
-
nn.Sequential(
|
144 |
-
nn.Linear(self.thread_hidden_size, self.thread_hidden_size * 2),
|
145 |
-
nn.GELU(),
|
146 |
-
nn.Linear(self.thread_hidden_size * 2, self.thread_hidden_size),
|
147 |
-
HelpingAIRMSNorm(self.thread_hidden_size, eps=config.rms_norm_eps)
|
148 |
-
)
|
149 |
-
for _ in range(self.perspective_threads)
|
150 |
-
])
|
151 |
-
|
152 |
-
# Output integration
|
153 |
-
self.thread_combiner = nn.Linear(
|
154 |
-
self.thread_hidden_size * self.perspective_threads,
|
155 |
-
self.hidden_size
|
156 |
-
)
|
157 |
-
|
158 |
-
# Thread importance weighting
|
159 |
-
self.thread_weights = nn.Parameter(torch.ones(self.perspective_threads))
|
160 |
-
|
161 |
-
def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, dict]:
|
162 |
-
batch_size, seq_len, _ = hidden_states.shape
|
163 |
-
|
164 |
-
# Process each perspective thread
|
165 |
-
thread_outputs = []
|
166 |
-
thread_metadata = {}
|
167 |
-
|
168 |
-
for i, (projection, processor, thread_name) in enumerate(
|
169 |
-
zip(self.thread_projections, self.thread_processors, self.thread_names)
|
170 |
-
):
|
171 |
-
# Project to thread space
|
172 |
-
thread_input = projection(hidden_states)
|
173 |
-
|
174 |
-
# Process thread-specific perspective
|
175 |
-
thread_output = processor(thread_input)
|
176 |
-
thread_outputs.append(thread_output)
|
177 |
-
|
178 |
-
# Store thread metadata
|
179 |
-
thread_metadata[f"{thread_name}_activation"] = torch.mean(torch.abs(thread_output))
|
180 |
-
|
181 |
-
# Stack threads for cross-thread attention
|
182 |
-
stacked_threads = torch.stack(thread_outputs, dim=2) # [batch, seq_len, num_threads, hidden]
|
183 |
-
stacked_threads = stacked_threads.reshape(batch_size * seq_len, self.perspective_threads, self.thread_hidden_size)
|
184 |
-
|
185 |
-
# Cross-thread attention for perspective integration
|
186 |
-
integrated_threads, cross_attention = self.cross_thread_attention(
|
187 |
-
stacked_threads, stacked_threads, stacked_threads
|
188 |
-
)
|
189 |
-
|
190 |
-
# Apply thread importance weighting
|
191 |
-
thread_weights_normalized = torch.softmax(self.thread_weights, dim=0)
|
192 |
-
weighted_threads = integrated_threads * thread_weights_normalized.unsqueeze(0).unsqueeze(-1)
|
193 |
-
|
194 |
-
# Combine threads - use reshape instead of view for memory layout compatibility
|
195 |
-
combined_threads = weighted_threads.reshape(batch_size, seq_len, -1)
|
196 |
-
final_output = self.thread_combiner(combined_threads)
|
197 |
-
|
198 |
-
# Thread metadata
|
199 |
-
thread_metadata.update({
|
200 |
-
"thread_weights": thread_weights_normalized,
|
201 |
-
"cross_attention": cross_attention,
|
202 |
-
"thread_activations": {
|
203 |
-
name: torch.mean(output) for name, output in zip(self.thread_names, thread_outputs)
|
204 |
-
}
|
205 |
-
})
|
206 |
-
|
207 |
-
return final_output, thread_metadata
|
208 |
-
|
209 |
-
|
210 |
-
class HelpingAIMultiStageThinking(nn.Module):
|
211 |
-
"""
|
212 |
-
Multi-stage thinking module for internal reasoning and reflection processes.
|
213 |
-
Implements cascaded thinking stages with simplified feedback loops.
|
214 |
-
"""
|
215 |
-
def __init__(self, config: HelpingAIConfig):
|
216 |
-
super().__init__()
|
217 |
-
self.config = config
|
218 |
-
self.hidden_size = config.hidden_size
|
219 |
-
self.thinking_stages = config.num_thinking_stages
|
220 |
-
self.thinking_depth = config.thinking_depth
|
221 |
-
|
222 |
-
# Thinking stage processors
|
223 |
-
self.thinking_layers = nn.ModuleList([
|
224 |
-
nn.Sequential(
|
225 |
-
nn.Linear(self.hidden_size, self.hidden_size),
|
226 |
-
nn.GELU(),
|
227 |
-
nn.Linear(self.hidden_size, self.hidden_size),
|
228 |
-
HelpingAIRMSNorm(self.hidden_size, eps=config.rms_norm_eps)
|
229 |
-
)
|
230 |
-
for _ in range(self.thinking_stages)
|
231 |
-
])
|
232 |
-
|
233 |
-
# Simple reflection mechanism without complex attention
|
234 |
-
self.reflection_layers = nn.ModuleList([
|
235 |
-
nn.Linear(self.hidden_size, self.hidden_size)
|
236 |
-
for _ in range(self.thinking_stages - 1)
|
237 |
-
])
|
238 |
-
|
239 |
-
# Stage transition gates
|
240 |
-
self.stage_gates = nn.ModuleList([
|
241 |
-
nn.Linear(self.hidden_size, 1) for _ in range(self.thinking_stages - 1)
|
242 |
-
])
|
243 |
-
|
244 |
-
# Thinking combination weights
|
245 |
-
self.stage_combiner = nn.Linear(self.thinking_stages * self.hidden_size, self.hidden_size)
|
246 |
-
|
247 |
-
def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, dict]:
|
248 |
-
batch_size, seq_len, _ = hidden_states.shape
|
249 |
-
thinking_outputs = []
|
250 |
-
thinking_metadata = {}
|
251 |
-
|
252 |
-
current_thought = hidden_states
|
253 |
-
|
254 |
-
# Multi-stage thinking process
|
255 |
-
for stage_idx, stage_processor in enumerate(self.thinking_layers):
|
256 |
-
# Process current thinking stage
|
257 |
-
current_thought = stage_processor(current_thought)
|
258 |
-
|
259 |
-
# Store stage output
|
260 |
-
thinking_outputs.append(current_thought)
|
261 |
-
thinking_metadata[f"stage_{stage_idx}_activation"] = torch.mean(torch.abs(current_thought)).item()
|
262 |
-
|
263 |
-
# Apply reflection if not the last stage
|
264 |
-
if stage_idx < self.thinking_stages - 1:
|
265 |
-
# Simple reflection mechanism
|
266 |
-
reflection = self.reflection_layers[stage_idx](current_thought)
|
267 |
-
current_thought = current_thought + 0.1 * reflection # Small reflection influence
|
268 |
-
|
269 |
-
# Stage transition gating
|
270 |
-
gate_weight = torch.sigmoid(self.stage_gates[stage_idx](current_thought))
|
271 |
-
current_thought = gate_weight * current_thought + (1 - gate_weight) * hidden_states
|
272 |
-
|
273 |
-
# Combine all thinking stages
|
274 |
-
all_thoughts = torch.cat(thinking_outputs, dim=-1) # Concatenate along hidden dimension
|
275 |
-
final_thought = self.stage_combiner(all_thoughts)
|
276 |
-
|
277 |
-
thinking_metadata["stage_contributions"] = [
|
278 |
-
torch.mean(torch.abs(output)).item() for output in thinking_outputs
|
279 |
-
]
|
280 |
-
|
281 |
-
return final_thought, thinking_metadata
|
282 |
-
|
283 |
-
|
284 |
-
class HelpingAIMLP(nn.Module):
|
285 |
-
def __init__(self, config):
|
286 |
-
super().__init__()
|
287 |
-
self.config = config
|
288 |
-
self.hidden_size = config.hidden_size
|
289 |
-
self.intermediate_size = config.intermediate_size
|
290 |
-
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
291 |
-
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
292 |
-
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
|
293 |
-
self.act_fn = ACT2FN[config.hidden_act]
|
294 |
-
|
295 |
-
# Enhanced MLP with thinking modules
|
296 |
-
if hasattr(config, 'use_emotional_reasoning') and config.use_emotional_reasoning:
|
297 |
-
self.thinking_module = HelpingAIMultiStageThinking(config)
|
298 |
-
self.use_thinking = True
|
299 |
-
else:
|
300 |
-
self.use_thinking = False
|
301 |
-
|
302 |
-
# Reasoning temperature for controlled generation
|
303 |
-
self.reasoning_temperature = getattr(config, 'reasoning_temperature', 1.0)
|
304 |
-
|
305 |
-
def forward(self, x):
|
306 |
-
# Standard MLP forward pass
|
307 |
-
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
|
308 |
-
|
309 |
-
# Apply multi-stage thinking if enabled
|
310 |
-
if self.use_thinking:
|
311 |
-
thinking_output, thinking_metadata = self.thinking_module(down_proj)
|
312 |
-
# Apply reasoning temperature
|
313 |
-
down_proj = down_proj + (thinking_output * self.reasoning_temperature)
|
314 |
-
|
315 |
-
return down_proj
|
316 |
-
|
317 |
-
|
318 |
-
def rotate_half(x):
|
319 |
-
"""Rotates half the hidden dims of the input."""
|
320 |
-
x1 = x[..., : x.shape[-1] // 2]
|
321 |
-
x2 = x[..., x.shape[-1] // 2 :]
|
322 |
-
return torch.cat((-x2, x1), dim=-1)
|
323 |
-
|
324 |
-
|
325 |
-
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
326 |
-
"""Applies Rotary Position Embedding to the query and key tensors.
|
327 |
-
|
328 |
-
Args:
|
329 |
-
q (`torch.Tensor`): The query tensor.
|
330 |
-
k (`torch.Tensor`): The key tensor.
|
331 |
-
cos (`torch.Tensor`): The cosine part of the rotary embedding.
|
332 |
-
sin (`torch.Tensor`): The sine part of the rotary embedding.
|
333 |
-
position_ids (`torch.Tensor`, *optional*):
|
334 |
-
Deprecated and unused.
|
335 |
-
unsqueeze_dim (`int`, *optional*, defaults to 1):
|
336 |
-
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
|
337 |
-
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
|
338 |
-
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
|
339 |
-
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
|
340 |
-
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
|
341 |
-
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
|
342 |
-
Returns:
|
343 |
-
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
344 |
-
"""
|
345 |
-
cos = cos.unsqueeze(unsqueeze_dim)
|
346 |
-
sin = sin.unsqueeze(unsqueeze_dim)
|
347 |
-
q_embed = (q * cos) + (rotate_half(q) * sin)
|
348 |
-
k_embed = (k * cos) + (rotate_half(k) * sin)
|
349 |
-
return q_embed, k_embed
|
350 |
-
|
351 |
-
|
352 |
-
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
353 |
-
"""
|
354 |
-
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
355 |
-
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
356 |
-
"""
|
357 |
-
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
358 |
-
if n_rep == 1:
|
359 |
-
return hidden_states
|
360 |
-
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
361 |
-
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
362 |
-
|
363 |
-
|
364 |
-
def eager_attention_forward(
|
365 |
-
module: nn.Module,
|
366 |
-
query: torch.Tensor,
|
367 |
-
key: torch.Tensor,
|
368 |
-
value: torch.Tensor,
|
369 |
-
attention_mask: Optional[torch.Tensor],
|
370 |
-
scaling: float,
|
371 |
-
dropout: float = 0.0,
|
372 |
-
**kwargs: Unpack[TransformersKwargs],
|
373 |
-
):
|
374 |
-
key_states = repeat_kv(key, module.num_key_value_groups)
|
375 |
-
value_states = repeat_kv(value, module.num_key_value_groups)
|
376 |
-
|
377 |
-
attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
|
378 |
-
if attention_mask is not None:
|
379 |
-
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
380 |
-
attn_weights = attn_weights + causal_mask
|
381 |
-
|
382 |
-
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
|
383 |
-
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|
384 |
-
attn_output = torch.matmul(attn_weights, value_states)
|
385 |
-
attn_output = attn_output.transpose(1, 2).contiguous()
|
386 |
-
|
387 |
-
return attn_output, attn_weights
|
388 |
-
|
389 |
-
|
390 |
-
class HelpingAIAttention(nn.Module):
|
391 |
-
"""Multi-headed attention with specialized emotional and empathetic reasoning capabilities"""
|
392 |
-
|
393 |
-
def __init__(self, config: HelpingAIConfig, layer_idx: int):
|
394 |
-
super().__init__()
|
395 |
-
self.config = config
|
396 |
-
self.layer_idx = layer_idx
|
397 |
-
self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
|
398 |
-
self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
|
399 |
-
self.scaling = self.head_dim**-0.5
|
400 |
-
self.attention_dropout = config.attention_dropout
|
401 |
-
self.is_causal = True
|
402 |
-
|
403 |
-
self.q_proj = nn.Linear(
|
404 |
-
config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
|
405 |
-
)
|
406 |
-
self.k_proj = nn.Linear(
|
407 |
-
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
|
408 |
-
)
|
409 |
-
self.v_proj = nn.Linear(
|
410 |
-
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
|
411 |
-
)
|
412 |
-
self.o_proj = nn.Linear(
|
413 |
-
config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
|
414 |
-
)
|
415 |
-
self.q_norm = HelpingAIRMSNorm(self.head_dim, eps=config.rms_norm_eps)
|
416 |
-
self.k_norm = HelpingAIRMSNorm(self.head_dim, eps=config.rms_norm_eps)
|
417 |
-
self.sliding_window = config.sliding_window if config.layer_types[layer_idx] == "sliding_attention" else None
|
418 |
-
|
419 |
-
# Enhanced emotional and empathetic attention
|
420 |
-
if hasattr(config, 'use_emotional_reasoning') and config.use_emotional_reasoning:
|
421 |
-
self.num_emotion_heads = getattr(config, 'num_emotion_heads', 4)
|
422 |
-
self.empathy_scaling_factor = getattr(config, 'empathy_scaling_factor', 1.2)
|
423 |
-
|
424 |
-
# Specialized emotion attention projections
|
425 |
-
self.emotion_q_proj = nn.Linear(config.hidden_size, self.num_emotion_heads * self.head_dim, bias=False)
|
426 |
-
self.emotion_k_proj = nn.Linear(config.hidden_size, self.num_emotion_heads * self.head_dim, bias=False)
|
427 |
-
self.emotion_v_proj = nn.Linear(config.hidden_size, self.num_emotion_heads * self.head_dim, bias=False)
|
428 |
-
|
429 |
-
# Empathy enhancement layer
|
430 |
-
self.empathy_enhancer = nn.Sequential(
|
431 |
-
nn.Linear(config.hidden_size, config.hidden_size // 2),
|
432 |
-
nn.GELU(),
|
433 |
-
nn.Linear(config.hidden_size // 2, config.num_attention_heads),
|
434 |
-
nn.Softmax(dim=-1)
|
435 |
-
)
|
436 |
-
|
437 |
-
self.use_emotional_attention = True
|
438 |
-
else:
|
439 |
-
self.use_emotional_attention = False
|
440 |
-
|
441 |
-
@deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
|
442 |
-
def forward(
|
443 |
-
self,
|
444 |
-
hidden_states: torch.Tensor,
|
445 |
-
position_embeddings: tuple[torch.Tensor, torch.Tensor],
|
446 |
-
attention_mask: Optional[torch.Tensor],
|
447 |
-
past_key_values: Optional[Cache] = None,
|
448 |
-
cache_position: Optional[torch.LongTensor] = None,
|
449 |
-
**kwargs: Unpack[FlashAttentionKwargs],
|
450 |
-
) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
|
451 |
-
input_shape = hidden_states.shape[:-1]
|
452 |
-
hidden_shape = (*input_shape, -1, self.head_dim)
|
453 |
-
|
454 |
-
# Standard attention processing
|
455 |
-
query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
|
456 |
-
key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
|
457 |
-
value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
458 |
-
|
459 |
-
cos, sin = position_embeddings
|
460 |
-
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
461 |
-
|
462 |
-
if past_key_values is not None:
|
463 |
-
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
464 |
-
key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
465 |
-
|
466 |
-
# Enhanced emotional attention processing
|
467 |
-
if self.use_emotional_attention:
|
468 |
-
# Compute empathy weights
|
469 |
-
empathy_weights = self.empathy_enhancer(hidden_states.mean(dim=1)) # [batch, num_heads]
|
470 |
-
|
471 |
-
# Emotional query, key, value computation
|
472 |
-
emotion_query = self.emotion_q_proj(hidden_states).view(*input_shape, self.num_emotion_heads, self.head_dim).transpose(1, 2)
|
473 |
-
emotion_key = self.emotion_k_proj(hidden_states).view(*input_shape, self.num_emotion_heads, self.head_dim).transpose(1, 2)
|
474 |
-
emotion_value = self.emotion_v_proj(hidden_states).view(*input_shape, self.num_emotion_heads, self.head_dim).transpose(1, 2)
|
475 |
-
|
476 |
-
# Apply rotary embeddings to emotional attention
|
477 |
-
emotion_query, emotion_key = apply_rotary_pos_emb(emotion_query, emotion_key, cos, sin)
|
478 |
-
|
479 |
-
# Emotional attention computation
|
480 |
-
emotion_scaling = (self.head_dim ** -0.5) * self.empathy_scaling_factor
|
481 |
-
emotion_attn_weights = torch.matmul(emotion_query, emotion_key.transpose(2, 3)) * emotion_scaling
|
482 |
-
|
483 |
-
if attention_mask is not None:
|
484 |
-
emotion_causal_mask = attention_mask[:, :, :, :emotion_key.shape[-2]]
|
485 |
-
emotion_attn_weights = emotion_attn_weights + emotion_causal_mask
|
486 |
-
|
487 |
-
emotion_attn_weights = nn.functional.softmax(emotion_attn_weights, dim=-1, dtype=torch.float32).to(emotion_query.dtype)
|
488 |
-
emotion_output = torch.matmul(emotion_attn_weights, emotion_value)
|
489 |
-
|
490 |
-
# Integrate emotional attention with standard attention
|
491 |
-
# Pad or truncate emotional attention to match standard attention heads
|
492 |
-
if self.num_emotion_heads < self.config.num_attention_heads:
|
493 |
-
padding_heads = self.config.num_attention_heads - self.num_emotion_heads
|
494 |
-
emotion_padding = torch.zeros(
|
495 |
-
*emotion_output.shape[:-3], padding_heads, *emotion_output.shape[-2:],
|
496 |
-
device=emotion_output.device, dtype=emotion_output.dtype
|
497 |
-
)
|
498 |
-
emotion_output = torch.cat([emotion_output, emotion_padding], dim=1)
|
499 |
-
|
500 |
-
# Standard attention computation
|
501 |
-
attention_interface: Callable = eager_attention_forward
|
502 |
-
if self.config._attn_implementation != "eager":
|
503 |
-
attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
|
504 |
-
|
505 |
-
attn_output, attn_weights = attention_interface(
|
506 |
-
self,
|
507 |
-
query_states,
|
508 |
-
key_states,
|
509 |
-
value_states,
|
510 |
-
attention_mask,
|
511 |
-
dropout=0.0 if not self.training else self.attention_dropout,
|
512 |
-
scaling=self.scaling,
|
513 |
-
sliding_window=self.sliding_window,
|
514 |
-
**kwargs,
|
515 |
-
)
|
516 |
-
|
517 |
-
# Blend standard and emotional attention if emotional reasoning is enabled
|
518 |
-
if self.use_emotional_attention:
|
519 |
-
# For now, use a simplified approach - just apply empathy scaling
|
520 |
-
# This avoids the complex tensor dimension matching issues
|
521 |
-
batch_size, num_heads, seq_len, head_dim = attn_output.shape
|
522 |
-
|
523 |
-
# Get average empathy weight per batch
|
524 |
-
empathy_scale = torch.mean(empathy_weights, dim=1, keepdim=True) # [batch, 1]
|
525 |
-
empathy_scale = empathy_scale.view(batch_size, 1, 1, 1) # [batch, 1, 1, 1]
|
526 |
-
empathy_scale = empathy_scale.expand(batch_size, num_heads, seq_len, head_dim)
|
527 |
-
|
528 |
-
# Apply empathy scaling to attention output
|
529 |
-
attn_output = attn_output * (1.0 + empathy_scale * 0.1) # Small empathy influence
|
530 |
-
|
531 |
-
attn_output = attn_output.reshape(*input_shape, -1).contiguous()
|
532 |
-
attn_output = self.o_proj(attn_output)
|
533 |
-
return attn_output, attn_weights
|
534 |
-
|
535 |
-
|
536 |
-
class HelpingAIDecoderLayer(GradientCheckpointingLayer):
|
537 |
-
def __init__(self, config: HelpingAIConfig, layer_idx: int):
|
538 |
-
super().__init__()
|
539 |
-
self.hidden_size = config.hidden_size
|
540 |
-
self.layer_idx = layer_idx
|
541 |
-
|
542 |
-
self.self_attn = HelpingAIAttention(config=config, layer_idx=layer_idx)
|
543 |
-
self.mlp = HelpingAIMLP(config)
|
544 |
-
self.input_layernorm = HelpingAIRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
545 |
-
self.post_attention_layernorm = HelpingAIRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
546 |
-
self.attention_type = config.layer_types[layer_idx]
|
547 |
-
|
548 |
-
# Enhanced reasoning layers
|
549 |
-
if hasattr(config, 'use_emotional_reasoning') and config.use_emotional_reasoning:
|
550 |
-
self.ser_layer = HelpingAISemanticEmotionReasoning(config)
|
551 |
-
self.use_ser = True
|
552 |
-
else:
|
553 |
-
self.use_ser = False
|
554 |
-
|
555 |
-
if hasattr(config, 'use_perspective_threading') and config.use_perspective_threading:
|
556 |
-
self.pet_layer = HelpingAIPerspectiveEmotionThreading(config)
|
557 |
-
self.use_pet = True
|
558 |
-
else:
|
559 |
-
self.use_pet = False
|
560 |
-
|
561 |
-
# Reasoning integration layers
|
562 |
-
if self.use_ser or self.use_pet:
|
563 |
-
self.reasoning_norm = HelpingAIRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
564 |
-
self.reasoning_gate = nn.Linear(config.hidden_size, 1)
|
565 |
-
|
566 |
-
@deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
|
567 |
-
def forward(
|
568 |
-
self,
|
569 |
-
hidden_states: torch.Tensor,
|
570 |
-
attention_mask: Optional[torch.Tensor] = None,
|
571 |
-
position_ids: Optional[torch.LongTensor] = None,
|
572 |
-
past_key_values: Optional[Cache] = None,
|
573 |
-
use_cache: Optional[bool] = False,
|
574 |
-
cache_position: Optional[torch.LongTensor] = None,
|
575 |
-
position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
|
576 |
-
**kwargs: Unpack[TransformersKwargs],
|
577 |
-
) -> torch.Tensor:
|
578 |
-
residual = hidden_states
|
579 |
-
hidden_states = self.input_layernorm(hidden_states)
|
580 |
-
|
581 |
-
# Self Attention
|
582 |
-
hidden_states, attention_weights = self.self_attn(
|
583 |
-
hidden_states=hidden_states,
|
584 |
-
attention_mask=attention_mask,
|
585 |
-
position_ids=position_ids,
|
586 |
-
past_key_values=past_key_values,
|
587 |
-
use_cache=use_cache,
|
588 |
-
cache_position=cache_position,
|
589 |
-
position_embeddings=position_embeddings,
|
590 |
-
**kwargs,
|
591 |
-
)
|
592 |
-
hidden_states = residual + hidden_states
|
593 |
-
|
594 |
-
# Enhanced reasoning processing
|
595 |
-
reasoning_outputs = []
|
596 |
-
reasoning_metadata = {}
|
597 |
-
|
598 |
-
if self.use_ser:
|
599 |
-
# Semantic Emotion Reasoning
|
600 |
-
ser_output, ser_meta = self.ser_layer(hidden_states)
|
601 |
-
reasoning_outputs.append(ser_output)
|
602 |
-
reasoning_metadata['ser'] = ser_meta
|
603 |
-
|
604 |
-
if self.use_pet:
|
605 |
-
# Perspective Emotion Threading
|
606 |
-
pet_output, pet_meta = self.pet_layer(hidden_states)
|
607 |
-
reasoning_outputs.append(pet_output)
|
608 |
-
reasoning_metadata['pet'] = pet_meta
|
609 |
-
|
610 |
-
# Integrate reasoning outputs if any
|
611 |
-
if reasoning_outputs:
|
612 |
-
# Combine reasoning outputs
|
613 |
-
combined_reasoning = torch.stack(reasoning_outputs, dim=0).mean(dim=0)
|
614 |
-
combined_reasoning = self.reasoning_norm(combined_reasoning)
|
615 |
-
|
616 |
-
# Apply gating to control reasoning influence
|
617 |
-
reasoning_gate = torch.sigmoid(self.reasoning_gate(hidden_states))
|
618 |
-
hidden_states = hidden_states + (reasoning_gate * combined_reasoning)
|
619 |
-
|
620 |
-
# Fully Connected (MLP)
|
621 |
-
residual = hidden_states
|
622 |
-
hidden_states = self.post_attention_layernorm(hidden_states)
|
623 |
-
hidden_states = self.mlp(hidden_states)
|
624 |
-
hidden_states = residual + hidden_states
|
625 |
-
|
626 |
-
# Store reasoning metadata for analysis (optional)
|
627 |
-
if hasattr(hidden_states, '_reasoning_metadata'):
|
628 |
-
hidden_states._reasoning_metadata = reasoning_metadata
|
629 |
-
|
630 |
-
return hidden_states
|
631 |
-
|
632 |
-
|
633 |
-
@auto_docstring
|
634 |
-
class HelpingAIPreTrainedModel(PreTrainedModel):
|
635 |
-
config: HelpingAIConfig
|
636 |
-
base_model_prefix = "model"
|
637 |
-
supports_gradient_checkpointing = True
|
638 |
-
_no_split_modules = ["HelpingAIDecoderLayer"]
|
639 |
-
_skip_keys_device_placement = ["past_key_values"]
|
640 |
-
_supports_flash_attn = True
|
641 |
-
_supports_sdpa = True
|
642 |
-
_supports_flex_attn = True
|
643 |
-
|
644 |
-
_can_compile_fullgraph = True
|
645 |
-
_supports_attention_backend = True
|
646 |
-
_can_record_outputs = {
|
647 |
-
"hidden_states": HelpingAIDecoderLayer,
|
648 |
-
"attentions": HelpingAIAttention,
|
649 |
-
}
|
650 |
-
|
651 |
-
|
652 |
-
class HelpingAIRotaryEmbedding(nn.Module):
|
653 |
-
inv_freq: torch.Tensor # fix linting for `register_buffer`
|
654 |
-
|
655 |
-
def __init__(self, config: HelpingAIConfig, device=None):
|
656 |
-
super().__init__()
|
657 |
-
# BC: "rope_type" was originally "type"
|
658 |
-
if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
|
659 |
-
self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
|
660 |
-
else:
|
661 |
-
self.rope_type = "default"
|
662 |
-
self.max_seq_len_cached = config.max_position_embeddings
|
663 |
-
self.original_max_seq_len = config.max_position_embeddings
|
664 |
-
|
665 |
-
self.config = config
|
666 |
-
self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
|
667 |
-
|
668 |
-
inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
|
669 |
-
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
670 |
-
self.original_inv_freq = self.inv_freq
|
671 |
-
|
672 |
-
@torch.no_grad()
|
673 |
-
@dynamic_rope_update # power user: used with advanced RoPE types (e.g. dynamic rope)
|
674 |
-
def forward(self, x, position_ids):
|
675 |
-
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
|
676 |
-
position_ids_expanded = position_ids[:, None, :].float()
|
677 |
-
|
678 |
-
device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
|
679 |
-
with torch.autocast(device_type=device_type, enabled=False): # Force float32
|
680 |
-
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
|
681 |
-
emb = torch.cat((freqs, freqs), dim=-1)
|
682 |
-
cos = emb.cos() * self.attention_scaling
|
683 |
-
sin = emb.sin() * self.attention_scaling
|
684 |
-
|
685 |
-
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
686 |
-
|
687 |
-
|
688 |
-
@auto_docstring
|
689 |
-
class HelpingAIModel(HelpingAIPreTrainedModel):
|
690 |
-
def __init__(self, config: HelpingAIConfig):
|
691 |
-
super().__init__(config)
|
692 |
-
self.padding_idx = config.pad_token_id
|
693 |
-
self.vocab_size = config.vocab_size
|
694 |
-
|
695 |
-
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
696 |
-
self.layers = nn.ModuleList(
|
697 |
-
[HelpingAIDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
698 |
-
)
|
699 |
-
self.norm = HelpingAIRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
700 |
-
self.rotary_emb = HelpingAIRotaryEmbedding(config=config)
|
701 |
-
self.gradient_checkpointing = False
|
702 |
-
self.has_sliding_layers = "sliding_attention" in self.config.layer_types
|
703 |
-
|
704 |
-
# Initialize weights and apply final processing
|
705 |
-
self.post_init()
|
706 |
-
|
707 |
-
@check_model_inputs
|
708 |
-
@auto_docstring
|
709 |
-
def forward(
|
710 |
-
self,
|
711 |
-
input_ids: Optional[torch.LongTensor] = None,
|
712 |
-
attention_mask: Optional[torch.Tensor] = None,
|
713 |
-
position_ids: Optional[torch.LongTensor] = None,
|
714 |
-
past_key_values: Optional[Cache] = None,
|
715 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
716 |
-
use_cache: Optional[bool] = None,
|
717 |
-
cache_position: Optional[torch.LongTensor] = None,
|
718 |
-
**kwargs: Unpack[TransformersKwargs],
|
719 |
-
) -> BaseModelOutputWithPast:
|
720 |
-
if (input_ids is None) ^ (inputs_embeds is not None):
|
721 |
-
raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
|
722 |
-
|
723 |
-
if inputs_embeds is None:
|
724 |
-
inputs_embeds = self.embed_tokens(input_ids)
|
725 |
-
|
726 |
-
if use_cache and past_key_values is None:
|
727 |
-
past_key_values = DynamicCache()
|
728 |
-
|
729 |
-
if cache_position is None:
|
730 |
-
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
731 |
-
cache_position = torch.arange(
|
732 |
-
past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
|
733 |
-
)
|
734 |
-
|
735 |
-
if position_ids is None:
|
736 |
-
position_ids = cache_position.unsqueeze(0)
|
737 |
-
|
738 |
-
# It may already have been prepared by e.g. `generate`
|
739 |
-
if not isinstance(causal_mask_mapping := attention_mask, dict):
|
740 |
-
# Prepare mask arguments
|
741 |
-
mask_kwargs = {
|
742 |
-
"config": self.config,
|
743 |
-
"input_embeds": inputs_embeds,
|
744 |
-
"attention_mask": attention_mask,
|
745 |
-
"cache_position": cache_position,
|
746 |
-
"past_key_values": past_key_values,
|
747 |
-
"position_ids": position_ids,
|
748 |
-
}
|
749 |
-
# Create the masks
|
750 |
-
causal_mask_mapping = {
|
751 |
-
"full_attention": create_causal_mask(**mask_kwargs),
|
752 |
-
}
|
753 |
-
# The sliding window alternating layers are not always activated depending on the config
|
754 |
-
if self.has_sliding_layers:
|
755 |
-
causal_mask_mapping["sliding_attention"] = create_sliding_window_causal_mask(**mask_kwargs)
|
756 |
-
|
757 |
-
hidden_states = inputs_embeds
|
758 |
-
|
759 |
-
# create position embeddings to be shared across the decoder layers
|
760 |
-
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
761 |
-
|
762 |
-
for decoder_layer in self.layers[: self.config.num_hidden_layers]:
|
763 |
-
hidden_states = decoder_layer(
|
764 |
-
hidden_states,
|
765 |
-
attention_mask=causal_mask_mapping[decoder_layer.attention_type],
|
766 |
-
position_ids=position_ids,
|
767 |
-
past_key_values=past_key_values,
|
768 |
-
use_cache=use_cache,
|
769 |
-
cache_position=cache_position,
|
770 |
-
position_embeddings=position_embeddings,
|
771 |
-
**kwargs,
|
772 |
-
)
|
773 |
-
|
774 |
-
hidden_states = self.norm(hidden_states)
|
775 |
-
return BaseModelOutputWithPast(
|
776 |
-
last_hidden_state=hidden_states,
|
777 |
-
past_key_values=past_key_values if use_cache else None,
|
778 |
-
)
|
779 |
-
|
780 |
-
|
781 |
-
@auto_docstring
|
782 |
-
class HelpingAIForCausalLM(HelpingAIPreTrainedModel, GenerationMixin):
|
783 |
-
_tied_weights_keys = ["lm_head.weight"]
|
784 |
-
_tp_plan = {"lm_head": "colwise_rep"}
|
785 |
-
_pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
|
786 |
-
|
787 |
-
def __init__(self, config):
|
788 |
-
super().__init__(config)
|
789 |
-
self.model = HelpingAIModel(config)
|
790 |
-
self.vocab_size = config.vocab_size
|
791 |
-
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
792 |
-
|
793 |
-
# Enhanced structured output support
|
794 |
-
if hasattr(config, 'structured_output_vocab_size') and config.structured_output_vocab_size > 0:
|
795 |
-
self.structured_vocab_size = config.structured_output_vocab_size
|
796 |
-
self.structured_lm_head = nn.Linear(config.hidden_size, self.structured_vocab_size, bias=False)
|
797 |
-
self.use_structured_output = True
|
798 |
-
|
799 |
-
# Special token embeddings for structured reasoning
|
800 |
-
self.structured_token_embeddings = nn.Embedding(self.structured_vocab_size, config.hidden_size)
|
801 |
-
|
802 |
-
# Reasoning mode classifier
|
803 |
-
self.reasoning_mode_classifier = nn.Sequential(
|
804 |
-
nn.Linear(config.hidden_size, config.hidden_size // 2),
|
805 |
-
nn.GELU(),
|
806 |
-
nn.Linear(config.hidden_size // 2, 4), # think, ser, pet, normal
|
807 |
-
nn.Softmax(dim=-1)
|
808 |
-
)
|
809 |
-
else:
|
810 |
-
self.use_structured_output = False
|
811 |
-
|
812 |
-
#
|
813 |
-
self.
|
814 |
-
|
815 |
-
|
816 |
-
|
817 |
-
|
818 |
-
|
819 |
-
|
820 |
-
|
821 |
-
|
822 |
-
|
823 |
-
|
824 |
-
|
825 |
-
|
826 |
-
|
827 |
-
|
828 |
-
|
829 |
-
|
830 |
-
|
831 |
-
|
832 |
-
|
833 |
-
|
834 |
-
|
835 |
-
|
836 |
-
|
837 |
-
|
838 |
-
|
839 |
-
|
840 |
-
|
841 |
-
|
842 |
-
|
843 |
-
|
844 |
-
|
845 |
-
|
846 |
-
|
847 |
-
|
848 |
-
|
849 |
-
|
850 |
-
|
851 |
-
|
852 |
-
|
853 |
-
|
854 |
-
|
855 |
-
|
856 |
-
|
857 |
-
|
858 |
-
|
859 |
-
|
860 |
-
|
861 |
-
|
862 |
-
|
863 |
-
|
864 |
-
|
865 |
-
|
866 |
-
|
867 |
-
|
868 |
-
|
869 |
-
|
870 |
-
|
871 |
-
|
872 |
-
`
|
873 |
-
|
874 |
-
|
875 |
-
|
876 |
-
|
877 |
-
|
878 |
-
|
879 |
-
|
880 |
-
|
881 |
-
|
882 |
-
|
883 |
-
|
884 |
-
|
885 |
-
|
886 |
-
|
887 |
-
|
888 |
-
|
889 |
-
|
890 |
-
|
891 |
-
|
892 |
-
|
893 |
-
|
894 |
-
|
895 |
-
|
896 |
-
|
897 |
-
|
898 |
-
|
899 |
-
|
900 |
-
|
901 |
-
|
902 |
-
|
903 |
-
|
904 |
-
|
905 |
-
|
906 |
-
|
907 |
-
|
908 |
-
|
909 |
-
|
910 |
-
|
911 |
-
|
912 |
-
|
913 |
-
|
914 |
-
|
915 |
-
|
916 |
-
|
917 |
-
|
918 |
-
|
919 |
-
|
920 |
-
|
921 |
-
|
922 |
-
|
923 |
-
|
924 |
-
|
925 |
-
|
926 |
-
|
927 |
-
|
928 |
-
|
929 |
-
|
930 |
-
|
931 |
-
|
932 |
-
|
933 |
-
|
934 |
-
output
|
935 |
-
|
936 |
-
|
937 |
-
|
938 |
-
|
939 |
-
|
940 |
-
|
941 |
-
|
942 |
-
|
943 |
-
|
944 |
-
|
945 |
-
|
946 |
-
|
947 |
-
|
948 |
-
|
949 |
-
|
950 |
-
|
951 |
-
|
952 |
-
|
953 |
-
|
954 |
-
|
955 |
-
|
956 |
-
|
957 |
-
|
958 |
-
|
959 |
-
|
960 |
-
|
961 |
-
|
962 |
-
|
963 |
-
|
964 |
-
|
965 |
-
|
966 |
-
|
967 |
-
|
968 |
-
|
969 |
-
]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Callable, Optional, Union
|
2 |
+
|
3 |
+
import torch
|
4 |
+
from torch import nn
|
5 |
+
|
6 |
+
from transformers.activations import ACT2FN
|
7 |
+
from transformers.cache_utils import Cache, DynamicCache
|
8 |
+
from transformers.generation import GenerationMixin
|
9 |
+
from transformers.integrations import use_kernel_forward_from_hub
|
10 |
+
from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask
|
11 |
+
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
|
12 |
+
from transformers.modeling_layers import (
|
13 |
+
GenericForQuestionAnswering,
|
14 |
+
GenericForSequenceClassification,
|
15 |
+
GenericForTokenClassification,
|
16 |
+
GradientCheckpointingLayer,
|
17 |
+
)
|
18 |
+
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
|
19 |
+
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
|
20 |
+
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
|
21 |
+
from transformers.processing_utils import Unpack
|
22 |
+
from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple
|
23 |
+
from transformers.utils.deprecation import deprecate_kwarg
|
24 |
+
from transformers.utils.generic import check_model_inputs
|
25 |
+
from .configuration_helpingai import HelpingAIConfig
|
26 |
+
|
27 |
+
|
28 |
+
@use_kernel_forward_from_hub("RMSNorm")
|
29 |
+
class HelpingAIRMSNorm(nn.Module):
|
30 |
+
def __init__(self, hidden_size, eps=1e-6):
|
31 |
+
"""
|
32 |
+
HelpingAIRMSNorm is equivalent to T5LayerNorm
|
33 |
+
"""
|
34 |
+
super().__init__()
|
35 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
36 |
+
self.variance_epsilon = eps
|
37 |
+
|
38 |
+
def forward(self, hidden_states):
|
39 |
+
input_dtype = hidden_states.dtype
|
40 |
+
hidden_states = hidden_states.to(torch.float32)
|
41 |
+
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
42 |
+
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
43 |
+
return self.weight * hidden_states.to(input_dtype)
|
44 |
+
|
45 |
+
def extra_repr(self):
|
46 |
+
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
47 |
+
|
48 |
+
|
49 |
+
class HelpingAISemanticEmotionReasoning(nn.Module):
|
50 |
+
"""
|
51 |
+
Structured Emotional Reasoning (SER) layer for emotional understanding and processing.
|
52 |
+
Maps emotions to semantic representations and provides contextual emotion analysis.
|
53 |
+
"""
|
54 |
+
def __init__(self, config: HelpingAIConfig):
|
55 |
+
super().__init__()
|
56 |
+
self.config = config
|
57 |
+
self.emotion_hidden_size = config.emotion_hidden_size
|
58 |
+
self.hidden_size = config.hidden_size
|
59 |
+
|
60 |
+
# Emotion detection and mapping
|
61 |
+
self.emotion_detector = nn.Linear(self.hidden_size, self.emotion_hidden_size)
|
62 |
+
self.emotion_mapper = nn.Linear(self.emotion_hidden_size, self.emotion_hidden_size)
|
63 |
+
|
64 |
+
# Contextual emotion analysis
|
65 |
+
self.emotion_context = nn.MultiheadAttention(
|
66 |
+
embed_dim=self.emotion_hidden_size,
|
67 |
+
num_heads=min(8, self.emotion_hidden_size // 64),
|
68 |
+
batch_first=True
|
69 |
+
)
|
70 |
+
|
71 |
+
# Emotion classification heads
|
72 |
+
self.primary_emotion = nn.Linear(self.emotion_hidden_size, 32) # Primary emotions
|
73 |
+
self.emotion_intensity = nn.Linear(self.emotion_hidden_size, 1) # Intensity score
|
74 |
+
self.emotion_valence = nn.Linear(self.emotion_hidden_size, 1) # Positive/negative
|
75 |
+
|
76 |
+
# Output projection
|
77 |
+
self.emotion_output = nn.Linear(self.emotion_hidden_size, self.hidden_size)
|
78 |
+
self.emotion_norm = HelpingAIRMSNorm(self.emotion_hidden_size, eps=config.rms_norm_eps)
|
79 |
+
|
80 |
+
# Activation
|
81 |
+
self.act_fn = ACT2FN[config.hidden_act]
|
82 |
+
|
83 |
+
def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, dict]:
|
84 |
+
# Detect emotional content
|
85 |
+
emotion_features = self.act_fn(self.emotion_detector(hidden_states))
|
86 |
+
emotion_mapped = self.emotion_mapper(emotion_features)
|
87 |
+
emotion_mapped = self.emotion_norm(emotion_mapped)
|
88 |
+
|
89 |
+
# Contextual emotion analysis
|
90 |
+
emotion_context, attention_weights = self.emotion_context(
|
91 |
+
emotion_mapped, emotion_mapped, emotion_mapped
|
92 |
+
)
|
93 |
+
|
94 |
+
# Emotion analysis outputs
|
95 |
+
primary_emotions = self.primary_emotion(emotion_context)
|
96 |
+
emotion_intensity = torch.sigmoid(self.emotion_intensity(emotion_context))
|
97 |
+
emotion_valence = torch.tanh(self.emotion_valence(emotion_context))
|
98 |
+
|
99 |
+
# Project back to hidden size
|
100 |
+
emotion_output = self.emotion_output(emotion_context)
|
101 |
+
|
102 |
+
# Emotion metadata
|
103 |
+
emotion_metadata = {
|
104 |
+
"primary_emotions": primary_emotions,
|
105 |
+
"intensity": emotion_intensity,
|
106 |
+
"valence": emotion_valence,
|
107 |
+
"attention_weights": attention_weights
|
108 |
+
}
|
109 |
+
|
110 |
+
return emotion_output, emotion_metadata
|
111 |
+
|
112 |
+
|
113 |
+
class HelpingAIPerspectiveEmotionThreading(nn.Module):
|
114 |
+
"""
|
115 |
+
Parallel Empathic Threads (PET) layer for multi-threaded emotional reasoning.
|
116 |
+
Processes multiple perspective threads: relatable, supportive, motivational, analytical.
|
117 |
+
"""
|
118 |
+
def __init__(self, config: HelpingAIConfig):
|
119 |
+
super().__init__()
|
120 |
+
self.config = config
|
121 |
+
self.hidden_size = config.hidden_size
|
122 |
+
self.perspective_threads = config.perspective_threads
|
123 |
+
self.thread_hidden_size = config.emotion_hidden_size
|
124 |
+
|
125 |
+
# Thread-specific processors
|
126 |
+
self.thread_projections = nn.ModuleList([
|
127 |
+
nn.Linear(self.hidden_size, self.thread_hidden_size)
|
128 |
+
for _ in range(self.perspective_threads)
|
129 |
+
])
|
130 |
+
|
131 |
+
# Thread names for interpretability
|
132 |
+
self.thread_names = ["relatable", "supportive", "motivational", "analytical"][:self.perspective_threads]
|
133 |
+
|
134 |
+
# Cross-thread attention for perspective integration
|
135 |
+
self.cross_thread_attention = nn.MultiheadAttention(
|
136 |
+
embed_dim=self.thread_hidden_size,
|
137 |
+
num_heads=min(4, self.thread_hidden_size // 64),
|
138 |
+
batch_first=True
|
139 |
+
)
|
140 |
+
|
141 |
+
# Thread-specific processing layers
|
142 |
+
self.thread_processors = nn.ModuleList([
|
143 |
+
nn.Sequential(
|
144 |
+
nn.Linear(self.thread_hidden_size, self.thread_hidden_size * 2),
|
145 |
+
nn.GELU(),
|
146 |
+
nn.Linear(self.thread_hidden_size * 2, self.thread_hidden_size),
|
147 |
+
HelpingAIRMSNorm(self.thread_hidden_size, eps=config.rms_norm_eps)
|
148 |
+
)
|
149 |
+
for _ in range(self.perspective_threads)
|
150 |
+
])
|
151 |
+
|
152 |
+
# Output integration
|
153 |
+
self.thread_combiner = nn.Linear(
|
154 |
+
self.thread_hidden_size * self.perspective_threads,
|
155 |
+
self.hidden_size
|
156 |
+
)
|
157 |
+
|
158 |
+
# Thread importance weighting
|
159 |
+
self.thread_weights = nn.Parameter(torch.ones(self.perspective_threads))
|
160 |
+
|
161 |
+
def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, dict]:
|
162 |
+
batch_size, seq_len, _ = hidden_states.shape
|
163 |
+
|
164 |
+
# Process each perspective thread
|
165 |
+
thread_outputs = []
|
166 |
+
thread_metadata = {}
|
167 |
+
|
168 |
+
for i, (projection, processor, thread_name) in enumerate(
|
169 |
+
zip(self.thread_projections, self.thread_processors, self.thread_names)
|
170 |
+
):
|
171 |
+
# Project to thread space
|
172 |
+
thread_input = projection(hidden_states)
|
173 |
+
|
174 |
+
# Process thread-specific perspective
|
175 |
+
thread_output = processor(thread_input)
|
176 |
+
thread_outputs.append(thread_output)
|
177 |
+
|
178 |
+
# Store thread metadata
|
179 |
+
thread_metadata[f"{thread_name}_activation"] = torch.mean(torch.abs(thread_output))
|
180 |
+
|
181 |
+
# Stack threads for cross-thread attention
|
182 |
+
stacked_threads = torch.stack(thread_outputs, dim=2) # [batch, seq_len, num_threads, hidden]
|
183 |
+
stacked_threads = stacked_threads.reshape(batch_size * seq_len, self.perspective_threads, self.thread_hidden_size)
|
184 |
+
|
185 |
+
# Cross-thread attention for perspective integration
|
186 |
+
integrated_threads, cross_attention = self.cross_thread_attention(
|
187 |
+
stacked_threads, stacked_threads, stacked_threads
|
188 |
+
)
|
189 |
+
|
190 |
+
# Apply thread importance weighting
|
191 |
+
thread_weights_normalized = torch.softmax(self.thread_weights, dim=0)
|
192 |
+
weighted_threads = integrated_threads * thread_weights_normalized.unsqueeze(0).unsqueeze(-1)
|
193 |
+
|
194 |
+
# Combine threads - use reshape instead of view for memory layout compatibility
|
195 |
+
combined_threads = weighted_threads.reshape(batch_size, seq_len, -1)
|
196 |
+
final_output = self.thread_combiner(combined_threads)
|
197 |
+
|
198 |
+
# Thread metadata
|
199 |
+
thread_metadata.update({
|
200 |
+
"thread_weights": thread_weights_normalized,
|
201 |
+
"cross_attention": cross_attention,
|
202 |
+
"thread_activations": {
|
203 |
+
name: torch.mean(output) for name, output in zip(self.thread_names, thread_outputs)
|
204 |
+
}
|
205 |
+
})
|
206 |
+
|
207 |
+
return final_output, thread_metadata
|
208 |
+
|
209 |
+
|
210 |
+
class HelpingAIMultiStageThinking(nn.Module):
|
211 |
+
"""
|
212 |
+
Multi-stage thinking module for internal reasoning and reflection processes.
|
213 |
+
Implements cascaded thinking stages with simplified feedback loops.
|
214 |
+
"""
|
215 |
+
def __init__(self, config: HelpingAIConfig):
|
216 |
+
super().__init__()
|
217 |
+
self.config = config
|
218 |
+
self.hidden_size = config.hidden_size
|
219 |
+
self.thinking_stages = config.num_thinking_stages
|
220 |
+
self.thinking_depth = config.thinking_depth
|
221 |
+
|
222 |
+
# Thinking stage processors
|
223 |
+
self.thinking_layers = nn.ModuleList([
|
224 |
+
nn.Sequential(
|
225 |
+
nn.Linear(self.hidden_size, self.hidden_size),
|
226 |
+
nn.GELU(),
|
227 |
+
nn.Linear(self.hidden_size, self.hidden_size),
|
228 |
+
HelpingAIRMSNorm(self.hidden_size, eps=config.rms_norm_eps)
|
229 |
+
)
|
230 |
+
for _ in range(self.thinking_stages)
|
231 |
+
])
|
232 |
+
|
233 |
+
# Simple reflection mechanism without complex attention
|
234 |
+
self.reflection_layers = nn.ModuleList([
|
235 |
+
nn.Linear(self.hidden_size, self.hidden_size)
|
236 |
+
for _ in range(self.thinking_stages - 1)
|
237 |
+
])
|
238 |
+
|
239 |
+
# Stage transition gates
|
240 |
+
self.stage_gates = nn.ModuleList([
|
241 |
+
nn.Linear(self.hidden_size, 1) for _ in range(self.thinking_stages - 1)
|
242 |
+
])
|
243 |
+
|
244 |
+
# Thinking combination weights
|
245 |
+
self.stage_combiner = nn.Linear(self.thinking_stages * self.hidden_size, self.hidden_size)
|
246 |
+
|
247 |
+
def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, dict]:
|
248 |
+
batch_size, seq_len, _ = hidden_states.shape
|
249 |
+
thinking_outputs = []
|
250 |
+
thinking_metadata = {}
|
251 |
+
|
252 |
+
current_thought = hidden_states
|
253 |
+
|
254 |
+
# Multi-stage thinking process
|
255 |
+
for stage_idx, stage_processor in enumerate(self.thinking_layers):
|
256 |
+
# Process current thinking stage
|
257 |
+
current_thought = stage_processor(current_thought)
|
258 |
+
|
259 |
+
# Store stage output
|
260 |
+
thinking_outputs.append(current_thought)
|
261 |
+
thinking_metadata[f"stage_{stage_idx}_activation"] = torch.mean(torch.abs(current_thought)).item()
|
262 |
+
|
263 |
+
# Apply reflection if not the last stage
|
264 |
+
if stage_idx < self.thinking_stages - 1:
|
265 |
+
# Simple reflection mechanism
|
266 |
+
reflection = self.reflection_layers[stage_idx](current_thought)
|
267 |
+
current_thought = current_thought + 0.1 * reflection # Small reflection influence
|
268 |
+
|
269 |
+
# Stage transition gating
|
270 |
+
gate_weight = torch.sigmoid(self.stage_gates[stage_idx](current_thought))
|
271 |
+
current_thought = gate_weight * current_thought + (1 - gate_weight) * hidden_states
|
272 |
+
|
273 |
+
# Combine all thinking stages
|
274 |
+
all_thoughts = torch.cat(thinking_outputs, dim=-1) # Concatenate along hidden dimension
|
275 |
+
final_thought = self.stage_combiner(all_thoughts)
|
276 |
+
|
277 |
+
thinking_metadata["stage_contributions"] = [
|
278 |
+
torch.mean(torch.abs(output)).item() for output in thinking_outputs
|
279 |
+
]
|
280 |
+
|
281 |
+
return final_thought, thinking_metadata
|
282 |
+
|
283 |
+
|
284 |
+
class HelpingAIMLP(nn.Module):
|
285 |
+
def __init__(self, config):
|
286 |
+
super().__init__()
|
287 |
+
self.config = config
|
288 |
+
self.hidden_size = config.hidden_size
|
289 |
+
self.intermediate_size = config.intermediate_size
|
290 |
+
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
291 |
+
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
292 |
+
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
|
293 |
+
self.act_fn = ACT2FN[config.hidden_act]
|
294 |
+
|
295 |
+
# Enhanced MLP with thinking modules
|
296 |
+
if hasattr(config, 'use_emotional_reasoning') and config.use_emotional_reasoning:
|
297 |
+
self.thinking_module = HelpingAIMultiStageThinking(config)
|
298 |
+
self.use_thinking = True
|
299 |
+
else:
|
300 |
+
self.use_thinking = False
|
301 |
+
|
302 |
+
# Reasoning temperature for controlled generation
|
303 |
+
self.reasoning_temperature = getattr(config, 'reasoning_temperature', 1.0)
|
304 |
+
|
305 |
+
def forward(self, x):
|
306 |
+
# Standard MLP forward pass
|
307 |
+
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
|
308 |
+
|
309 |
+
# Apply multi-stage thinking if enabled
|
310 |
+
if self.use_thinking:
|
311 |
+
thinking_output, thinking_metadata = self.thinking_module(down_proj)
|
312 |
+
# Apply reasoning temperature
|
313 |
+
down_proj = down_proj + (thinking_output * self.reasoning_temperature)
|
314 |
+
|
315 |
+
return down_proj
|
316 |
+
|
317 |
+
|
318 |
+
def rotate_half(x):
|
319 |
+
"""Rotates half the hidden dims of the input."""
|
320 |
+
x1 = x[..., : x.shape[-1] // 2]
|
321 |
+
x2 = x[..., x.shape[-1] // 2 :]
|
322 |
+
return torch.cat((-x2, x1), dim=-1)
|
323 |
+
|
324 |
+
|
325 |
+
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
326 |
+
"""Applies Rotary Position Embedding to the query and key tensors.
|
327 |
+
|
328 |
+
Args:
|
329 |
+
q (`torch.Tensor`): The query tensor.
|
330 |
+
k (`torch.Tensor`): The key tensor.
|
331 |
+
cos (`torch.Tensor`): The cosine part of the rotary embedding.
|
332 |
+
sin (`torch.Tensor`): The sine part of the rotary embedding.
|
333 |
+
position_ids (`torch.Tensor`, *optional*):
|
334 |
+
Deprecated and unused.
|
335 |
+
unsqueeze_dim (`int`, *optional*, defaults to 1):
|
336 |
+
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
|
337 |
+
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
|
338 |
+
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
|
339 |
+
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
|
340 |
+
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
|
341 |
+
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
|
342 |
+
Returns:
|
343 |
+
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
344 |
+
"""
|
345 |
+
cos = cos.unsqueeze(unsqueeze_dim)
|
346 |
+
sin = sin.unsqueeze(unsqueeze_dim)
|
347 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
348 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
349 |
+
return q_embed, k_embed
|
350 |
+
|
351 |
+
|
352 |
+
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
353 |
+
"""
|
354 |
+
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
355 |
+
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
356 |
+
"""
|
357 |
+
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
358 |
+
if n_rep == 1:
|
359 |
+
return hidden_states
|
360 |
+
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
361 |
+
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
362 |
+
|
363 |
+
|
364 |
+
def eager_attention_forward(
|
365 |
+
module: nn.Module,
|
366 |
+
query: torch.Tensor,
|
367 |
+
key: torch.Tensor,
|
368 |
+
value: torch.Tensor,
|
369 |
+
attention_mask: Optional[torch.Tensor],
|
370 |
+
scaling: float,
|
371 |
+
dropout: float = 0.0,
|
372 |
+
**kwargs: Unpack[TransformersKwargs],
|
373 |
+
):
|
374 |
+
key_states = repeat_kv(key, module.num_key_value_groups)
|
375 |
+
value_states = repeat_kv(value, module.num_key_value_groups)
|
376 |
+
|
377 |
+
attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
|
378 |
+
if attention_mask is not None:
|
379 |
+
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
380 |
+
attn_weights = attn_weights + causal_mask
|
381 |
+
|
382 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
|
383 |
+
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|
384 |
+
attn_output = torch.matmul(attn_weights, value_states)
|
385 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
386 |
+
|
387 |
+
return attn_output, attn_weights
|
388 |
+
|
389 |
+
|
390 |
+
class HelpingAIAttention(nn.Module):
|
391 |
+
"""Multi-headed attention with specialized emotional and empathetic reasoning capabilities"""
|
392 |
+
|
393 |
+
def __init__(self, config: HelpingAIConfig, layer_idx: int):
|
394 |
+
super().__init__()
|
395 |
+
self.config = config
|
396 |
+
self.layer_idx = layer_idx
|
397 |
+
self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
|
398 |
+
self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
|
399 |
+
self.scaling = self.head_dim**-0.5
|
400 |
+
self.attention_dropout = config.attention_dropout
|
401 |
+
self.is_causal = True
|
402 |
+
|
403 |
+
self.q_proj = nn.Linear(
|
404 |
+
config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
|
405 |
+
)
|
406 |
+
self.k_proj = nn.Linear(
|
407 |
+
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
|
408 |
+
)
|
409 |
+
self.v_proj = nn.Linear(
|
410 |
+
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
|
411 |
+
)
|
412 |
+
self.o_proj = nn.Linear(
|
413 |
+
config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
|
414 |
+
)
|
415 |
+
self.q_norm = HelpingAIRMSNorm(self.head_dim, eps=config.rms_norm_eps)
|
416 |
+
self.k_norm = HelpingAIRMSNorm(self.head_dim, eps=config.rms_norm_eps)
|
417 |
+
self.sliding_window = config.sliding_window if config.layer_types[layer_idx] == "sliding_attention" else None
|
418 |
+
|
419 |
+
# Enhanced emotional and empathetic attention
|
420 |
+
if hasattr(config, 'use_emotional_reasoning') and config.use_emotional_reasoning:
|
421 |
+
self.num_emotion_heads = getattr(config, 'num_emotion_heads', 4)
|
422 |
+
self.empathy_scaling_factor = getattr(config, 'empathy_scaling_factor', 1.2)
|
423 |
+
|
424 |
+
# Specialized emotion attention projections
|
425 |
+
self.emotion_q_proj = nn.Linear(config.hidden_size, self.num_emotion_heads * self.head_dim, bias=False)
|
426 |
+
self.emotion_k_proj = nn.Linear(config.hidden_size, self.num_emotion_heads * self.head_dim, bias=False)
|
427 |
+
self.emotion_v_proj = nn.Linear(config.hidden_size, self.num_emotion_heads * self.head_dim, bias=False)
|
428 |
+
|
429 |
+
# Empathy enhancement layer
|
430 |
+
self.empathy_enhancer = nn.Sequential(
|
431 |
+
nn.Linear(config.hidden_size, config.hidden_size // 2),
|
432 |
+
nn.GELU(),
|
433 |
+
nn.Linear(config.hidden_size // 2, config.num_attention_heads),
|
434 |
+
nn.Softmax(dim=-1)
|
435 |
+
)
|
436 |
+
|
437 |
+
self.use_emotional_attention = True
|
438 |
+
else:
|
439 |
+
self.use_emotional_attention = False
|
440 |
+
|
441 |
+
@deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
|
442 |
+
def forward(
|
443 |
+
self,
|
444 |
+
hidden_states: torch.Tensor,
|
445 |
+
position_embeddings: tuple[torch.Tensor, torch.Tensor],
|
446 |
+
attention_mask: Optional[torch.Tensor],
|
447 |
+
past_key_values: Optional[Cache] = None,
|
448 |
+
cache_position: Optional[torch.LongTensor] = None,
|
449 |
+
**kwargs: Unpack[FlashAttentionKwargs],
|
450 |
+
) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
|
451 |
+
input_shape = hidden_states.shape[:-1]
|
452 |
+
hidden_shape = (*input_shape, -1, self.head_dim)
|
453 |
+
|
454 |
+
# Standard attention processing
|
455 |
+
query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
|
456 |
+
key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
|
457 |
+
value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
458 |
+
|
459 |
+
cos, sin = position_embeddings
|
460 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
461 |
+
|
462 |
+
if past_key_values is not None:
|
463 |
+
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
464 |
+
key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
465 |
+
|
466 |
+
# Enhanced emotional attention processing
|
467 |
+
if self.use_emotional_attention:
|
468 |
+
# Compute empathy weights
|
469 |
+
empathy_weights = self.empathy_enhancer(hidden_states.mean(dim=1)) # [batch, num_heads]
|
470 |
+
|
471 |
+
# Emotional query, key, value computation
|
472 |
+
emotion_query = self.emotion_q_proj(hidden_states).view(*input_shape, self.num_emotion_heads, self.head_dim).transpose(1, 2)
|
473 |
+
emotion_key = self.emotion_k_proj(hidden_states).view(*input_shape, self.num_emotion_heads, self.head_dim).transpose(1, 2)
|
474 |
+
emotion_value = self.emotion_v_proj(hidden_states).view(*input_shape, self.num_emotion_heads, self.head_dim).transpose(1, 2)
|
475 |
+
|
476 |
+
# Apply rotary embeddings to emotional attention
|
477 |
+
emotion_query, emotion_key = apply_rotary_pos_emb(emotion_query, emotion_key, cos, sin)
|
478 |
+
|
479 |
+
# Emotional attention computation
|
480 |
+
emotion_scaling = (self.head_dim ** -0.5) * self.empathy_scaling_factor
|
481 |
+
emotion_attn_weights = torch.matmul(emotion_query, emotion_key.transpose(2, 3)) * emotion_scaling
|
482 |
+
|
483 |
+
if attention_mask is not None:
|
484 |
+
emotion_causal_mask = attention_mask[:, :, :, :emotion_key.shape[-2]]
|
485 |
+
emotion_attn_weights = emotion_attn_weights + emotion_causal_mask
|
486 |
+
|
487 |
+
emotion_attn_weights = nn.functional.softmax(emotion_attn_weights, dim=-1, dtype=torch.float32).to(emotion_query.dtype)
|
488 |
+
emotion_output = torch.matmul(emotion_attn_weights, emotion_value)
|
489 |
+
|
490 |
+
# Integrate emotional attention with standard attention
|
491 |
+
# Pad or truncate emotional attention to match standard attention heads
|
492 |
+
if self.num_emotion_heads < self.config.num_attention_heads:
|
493 |
+
padding_heads = self.config.num_attention_heads - self.num_emotion_heads
|
494 |
+
emotion_padding = torch.zeros(
|
495 |
+
*emotion_output.shape[:-3], padding_heads, *emotion_output.shape[-2:],
|
496 |
+
device=emotion_output.device, dtype=emotion_output.dtype
|
497 |
+
)
|
498 |
+
emotion_output = torch.cat([emotion_output, emotion_padding], dim=1)
|
499 |
+
|
500 |
+
# Standard attention computation
|
501 |
+
attention_interface: Callable = eager_attention_forward
|
502 |
+
if self.config._attn_implementation != "eager":
|
503 |
+
attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
|
504 |
+
|
505 |
+
attn_output, attn_weights = attention_interface(
|
506 |
+
self,
|
507 |
+
query_states,
|
508 |
+
key_states,
|
509 |
+
value_states,
|
510 |
+
attention_mask,
|
511 |
+
dropout=0.0 if not self.training else self.attention_dropout,
|
512 |
+
scaling=self.scaling,
|
513 |
+
sliding_window=self.sliding_window,
|
514 |
+
**kwargs,
|
515 |
+
)
|
516 |
+
|
517 |
+
# Blend standard and emotional attention if emotional reasoning is enabled
|
518 |
+
if self.use_emotional_attention:
|
519 |
+
# For now, use a simplified approach - just apply empathy scaling
|
520 |
+
# This avoids the complex tensor dimension matching issues
|
521 |
+
batch_size, num_heads, seq_len, head_dim = attn_output.shape
|
522 |
+
|
523 |
+
# Get average empathy weight per batch
|
524 |
+
empathy_scale = torch.mean(empathy_weights, dim=1, keepdim=True) # [batch, 1]
|
525 |
+
empathy_scale = empathy_scale.view(batch_size, 1, 1, 1) # [batch, 1, 1, 1]
|
526 |
+
empathy_scale = empathy_scale.expand(batch_size, num_heads, seq_len, head_dim)
|
527 |
+
|
528 |
+
# Apply empathy scaling to attention output
|
529 |
+
attn_output = attn_output * (1.0 + empathy_scale * 0.1) # Small empathy influence
|
530 |
+
|
531 |
+
attn_output = attn_output.reshape(*input_shape, -1).contiguous()
|
532 |
+
attn_output = self.o_proj(attn_output)
|
533 |
+
return attn_output, attn_weights
|
534 |
+
|
535 |
+
|
536 |
+
class HelpingAIDecoderLayer(GradientCheckpointingLayer):
|
537 |
+
def __init__(self, config: HelpingAIConfig, layer_idx: int):
|
538 |
+
super().__init__()
|
539 |
+
self.hidden_size = config.hidden_size
|
540 |
+
self.layer_idx = layer_idx
|
541 |
+
|
542 |
+
self.self_attn = HelpingAIAttention(config=config, layer_idx=layer_idx)
|
543 |
+
self.mlp = HelpingAIMLP(config)
|
544 |
+
self.input_layernorm = HelpingAIRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
545 |
+
self.post_attention_layernorm = HelpingAIRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
546 |
+
self.attention_type = config.layer_types[layer_idx]
|
547 |
+
|
548 |
+
# Enhanced reasoning layers
|
549 |
+
if hasattr(config, 'use_emotional_reasoning') and config.use_emotional_reasoning:
|
550 |
+
self.ser_layer = HelpingAISemanticEmotionReasoning(config)
|
551 |
+
self.use_ser = True
|
552 |
+
else:
|
553 |
+
self.use_ser = False
|
554 |
+
|
555 |
+
if hasattr(config, 'use_perspective_threading') and config.use_perspective_threading:
|
556 |
+
self.pet_layer = HelpingAIPerspectiveEmotionThreading(config)
|
557 |
+
self.use_pet = True
|
558 |
+
else:
|
559 |
+
self.use_pet = False
|
560 |
+
|
561 |
+
# Reasoning integration layers
|
562 |
+
if self.use_ser or self.use_pet:
|
563 |
+
self.reasoning_norm = HelpingAIRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
564 |
+
self.reasoning_gate = nn.Linear(config.hidden_size, 1)
|
565 |
+
|
566 |
+
@deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
|
567 |
+
def forward(
|
568 |
+
self,
|
569 |
+
hidden_states: torch.Tensor,
|
570 |
+
attention_mask: Optional[torch.Tensor] = None,
|
571 |
+
position_ids: Optional[torch.LongTensor] = None,
|
572 |
+
past_key_values: Optional[Cache] = None,
|
573 |
+
use_cache: Optional[bool] = False,
|
574 |
+
cache_position: Optional[torch.LongTensor] = None,
|
575 |
+
position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
|
576 |
+
**kwargs: Unpack[TransformersKwargs],
|
577 |
+
) -> torch.Tensor:
|
578 |
+
residual = hidden_states
|
579 |
+
hidden_states = self.input_layernorm(hidden_states)
|
580 |
+
|
581 |
+
# Self Attention
|
582 |
+
hidden_states, attention_weights = self.self_attn(
|
583 |
+
hidden_states=hidden_states,
|
584 |
+
attention_mask=attention_mask,
|
585 |
+
position_ids=position_ids,
|
586 |
+
past_key_values=past_key_values,
|
587 |
+
use_cache=use_cache,
|
588 |
+
cache_position=cache_position,
|
589 |
+
position_embeddings=position_embeddings,
|
590 |
+
**kwargs,
|
591 |
+
)
|
592 |
+
hidden_states = residual + hidden_states
|
593 |
+
|
594 |
+
# Enhanced reasoning processing
|
595 |
+
reasoning_outputs = []
|
596 |
+
reasoning_metadata = {}
|
597 |
+
|
598 |
+
if self.use_ser:
|
599 |
+
# Semantic Emotion Reasoning
|
600 |
+
ser_output, ser_meta = self.ser_layer(hidden_states)
|
601 |
+
reasoning_outputs.append(ser_output)
|
602 |
+
reasoning_metadata['ser'] = ser_meta
|
603 |
+
|
604 |
+
if self.use_pet:
|
605 |
+
# Perspective Emotion Threading
|
606 |
+
pet_output, pet_meta = self.pet_layer(hidden_states)
|
607 |
+
reasoning_outputs.append(pet_output)
|
608 |
+
reasoning_metadata['pet'] = pet_meta
|
609 |
+
|
610 |
+
# Integrate reasoning outputs if any
|
611 |
+
if reasoning_outputs:
|
612 |
+
# Combine reasoning outputs
|
613 |
+
combined_reasoning = torch.stack(reasoning_outputs, dim=0).mean(dim=0)
|
614 |
+
combined_reasoning = self.reasoning_norm(combined_reasoning)
|
615 |
+
|
616 |
+
# Apply gating to control reasoning influence
|
617 |
+
reasoning_gate = torch.sigmoid(self.reasoning_gate(hidden_states))
|
618 |
+
hidden_states = hidden_states + (reasoning_gate * combined_reasoning)
|
619 |
+
|
620 |
+
# Fully Connected (MLP)
|
621 |
+
residual = hidden_states
|
622 |
+
hidden_states = self.post_attention_layernorm(hidden_states)
|
623 |
+
hidden_states = self.mlp(hidden_states)
|
624 |
+
hidden_states = residual + hidden_states
|
625 |
+
|
626 |
+
# Store reasoning metadata for analysis (optional)
|
627 |
+
if hasattr(hidden_states, '_reasoning_metadata'):
|
628 |
+
hidden_states._reasoning_metadata = reasoning_metadata
|
629 |
+
|
630 |
+
return hidden_states
|
631 |
+
|
632 |
+
|
633 |
+
@auto_docstring
|
634 |
+
class HelpingAIPreTrainedModel(PreTrainedModel):
|
635 |
+
config: HelpingAIConfig
|
636 |
+
base_model_prefix = "model"
|
637 |
+
supports_gradient_checkpointing = True
|
638 |
+
_no_split_modules = ["HelpingAIDecoderLayer"]
|
639 |
+
_skip_keys_device_placement = ["past_key_values"]
|
640 |
+
_supports_flash_attn = True
|
641 |
+
_supports_sdpa = True
|
642 |
+
_supports_flex_attn = True
|
643 |
+
|
644 |
+
_can_compile_fullgraph = True
|
645 |
+
_supports_attention_backend = True
|
646 |
+
_can_record_outputs = {
|
647 |
+
"hidden_states": HelpingAIDecoderLayer,
|
648 |
+
"attentions": HelpingAIAttention,
|
649 |
+
}
|
650 |
+
|
651 |
+
|
652 |
+
class HelpingAIRotaryEmbedding(nn.Module):
|
653 |
+
inv_freq: torch.Tensor # fix linting for `register_buffer`
|
654 |
+
|
655 |
+
def __init__(self, config: HelpingAIConfig, device=None):
|
656 |
+
super().__init__()
|
657 |
+
# BC: "rope_type" was originally "type"
|
658 |
+
if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
|
659 |
+
self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
|
660 |
+
else:
|
661 |
+
self.rope_type = "default"
|
662 |
+
self.max_seq_len_cached = config.max_position_embeddings
|
663 |
+
self.original_max_seq_len = config.max_position_embeddings
|
664 |
+
|
665 |
+
self.config = config
|
666 |
+
self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
|
667 |
+
|
668 |
+
inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
|
669 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
670 |
+
self.original_inv_freq = self.inv_freq
|
671 |
+
|
672 |
+
@torch.no_grad()
|
673 |
+
@dynamic_rope_update # power user: used with advanced RoPE types (e.g. dynamic rope)
|
674 |
+
def forward(self, x, position_ids):
|
675 |
+
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
|
676 |
+
position_ids_expanded = position_ids[:, None, :].float()
|
677 |
+
|
678 |
+
device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
|
679 |
+
with torch.autocast(device_type=device_type, enabled=False): # Force float32
|
680 |
+
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
|
681 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
682 |
+
cos = emb.cos() * self.attention_scaling
|
683 |
+
sin = emb.sin() * self.attention_scaling
|
684 |
+
|
685 |
+
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
686 |
+
|
687 |
+
|
688 |
+
@auto_docstring
|
689 |
+
class HelpingAIModel(HelpingAIPreTrainedModel):
|
690 |
+
def __init__(self, config: HelpingAIConfig):
|
691 |
+
super().__init__(config)
|
692 |
+
self.padding_idx = config.pad_token_id
|
693 |
+
self.vocab_size = config.vocab_size
|
694 |
+
|
695 |
+
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
696 |
+
self.layers = nn.ModuleList(
|
697 |
+
[HelpingAIDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
698 |
+
)
|
699 |
+
self.norm = HelpingAIRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
700 |
+
self.rotary_emb = HelpingAIRotaryEmbedding(config=config)
|
701 |
+
self.gradient_checkpointing = False
|
702 |
+
self.has_sliding_layers = "sliding_attention" in self.config.layer_types
|
703 |
+
|
704 |
+
# Initialize weights and apply final processing
|
705 |
+
self.post_init()
|
706 |
+
|
707 |
+
@check_model_inputs
|
708 |
+
@auto_docstring
|
709 |
+
def forward(
|
710 |
+
self,
|
711 |
+
input_ids: Optional[torch.LongTensor] = None,
|
712 |
+
attention_mask: Optional[torch.Tensor] = None,
|
713 |
+
position_ids: Optional[torch.LongTensor] = None,
|
714 |
+
past_key_values: Optional[Cache] = None,
|
715 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
716 |
+
use_cache: Optional[bool] = None,
|
717 |
+
cache_position: Optional[torch.LongTensor] = None,
|
718 |
+
**kwargs: Unpack[TransformersKwargs],
|
719 |
+
) -> BaseModelOutputWithPast:
|
720 |
+
if (input_ids is None) ^ (inputs_embeds is not None):
|
721 |
+
raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
|
722 |
+
|
723 |
+
if inputs_embeds is None:
|
724 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
725 |
+
|
726 |
+
if use_cache and past_key_values is None:
|
727 |
+
past_key_values = DynamicCache()
|
728 |
+
|
729 |
+
if cache_position is None:
|
730 |
+
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
731 |
+
cache_position = torch.arange(
|
732 |
+
past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
|
733 |
+
)
|
734 |
+
|
735 |
+
if position_ids is None:
|
736 |
+
position_ids = cache_position.unsqueeze(0)
|
737 |
+
|
738 |
+
# It may already have been prepared by e.g. `generate`
|
739 |
+
if not isinstance(causal_mask_mapping := attention_mask, dict):
|
740 |
+
# Prepare mask arguments
|
741 |
+
mask_kwargs = {
|
742 |
+
"config": self.config,
|
743 |
+
"input_embeds": inputs_embeds,
|
744 |
+
"attention_mask": attention_mask,
|
745 |
+
"cache_position": cache_position,
|
746 |
+
"past_key_values": past_key_values,
|
747 |
+
"position_ids": position_ids,
|
748 |
+
}
|
749 |
+
# Create the masks
|
750 |
+
causal_mask_mapping = {
|
751 |
+
"full_attention": create_causal_mask(**mask_kwargs),
|
752 |
+
}
|
753 |
+
# The sliding window alternating layers are not always activated depending on the config
|
754 |
+
if self.has_sliding_layers:
|
755 |
+
causal_mask_mapping["sliding_attention"] = create_sliding_window_causal_mask(**mask_kwargs)
|
756 |
+
|
757 |
+
hidden_states = inputs_embeds
|
758 |
+
|
759 |
+
# create position embeddings to be shared across the decoder layers
|
760 |
+
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
761 |
+
|
762 |
+
for decoder_layer in self.layers[: self.config.num_hidden_layers]:
|
763 |
+
hidden_states = decoder_layer(
|
764 |
+
hidden_states,
|
765 |
+
attention_mask=causal_mask_mapping[decoder_layer.attention_type],
|
766 |
+
position_ids=position_ids,
|
767 |
+
past_key_values=past_key_values,
|
768 |
+
use_cache=use_cache,
|
769 |
+
cache_position=cache_position,
|
770 |
+
position_embeddings=position_embeddings,
|
771 |
+
**kwargs,
|
772 |
+
)
|
773 |
+
|
774 |
+
hidden_states = self.norm(hidden_states)
|
775 |
+
return BaseModelOutputWithPast(
|
776 |
+
last_hidden_state=hidden_states,
|
777 |
+
past_key_values=past_key_values if use_cache else None,
|
778 |
+
)
|
779 |
+
|
780 |
+
|
781 |
+
@auto_docstring
|
782 |
+
class HelpingAIForCausalLM(HelpingAIPreTrainedModel, GenerationMixin):
|
783 |
+
_tied_weights_keys = ["lm_head.weight"]
|
784 |
+
_tp_plan = {"lm_head": "colwise_rep"}
|
785 |
+
_pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
|
786 |
+
|
787 |
+
def __init__(self, config):
|
788 |
+
super().__init__(config)
|
789 |
+
self.model = HelpingAIModel(config)
|
790 |
+
self.vocab_size = config.vocab_size
|
791 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
792 |
+
|
793 |
+
# Enhanced structured output support
|
794 |
+
if hasattr(config, 'structured_output_vocab_size') and config.structured_output_vocab_size > 0:
|
795 |
+
self.structured_vocab_size = config.structured_output_vocab_size
|
796 |
+
self.structured_lm_head = nn.Linear(config.hidden_size, self.structured_vocab_size, bias=False)
|
797 |
+
self.use_structured_output = True
|
798 |
+
|
799 |
+
# Special token embeddings for structured reasoning
|
800 |
+
self.structured_token_embeddings = nn.Embedding(self.structured_vocab_size, config.hidden_size)
|
801 |
+
|
802 |
+
# Reasoning mode classifier
|
803 |
+
self.reasoning_mode_classifier = nn.Sequential(
|
804 |
+
nn.Linear(config.hidden_size, config.hidden_size // 2),
|
805 |
+
nn.GELU(),
|
806 |
+
nn.Linear(config.hidden_size // 2, 4), # think, ser, pet, normal
|
807 |
+
nn.Softmax(dim=-1)
|
808 |
+
)
|
809 |
+
else:
|
810 |
+
self.use_structured_output = False
|
811 |
+
|
812 |
+
# Optional speech output head (predict mel-spectrogram frames)
|
813 |
+
self.use_speech_output = getattr(config, "use_speech_output", False)
|
814 |
+
if self.use_speech_output:
|
815 |
+
self.speech_num_mels = getattr(config, "speech_num_mels", 80)
|
816 |
+
self.speech_upsample_factor = getattr(config, "speech_upsample_factor", 1)
|
817 |
+
hidden_dim = getattr(config, "speech_head_hidden_dim", None)
|
818 |
+
if hidden_dim is None:
|
819 |
+
hidden_dim = config.hidden_size // 2
|
820 |
+
# Projector from hidden_size -> hidden_dim -> mel bins
|
821 |
+
self.speech_proj = nn.Sequential(
|
822 |
+
nn.Linear(config.hidden_size, hidden_dim),
|
823 |
+
nn.GELU(),
|
824 |
+
nn.Linear(hidden_dim, self.speech_num_mels),
|
825 |
+
)
|
826 |
+
self.speech_loss_type = getattr(config, "speech_loss_type", "l1")
|
827 |
+
|
828 |
+
# Initialize weights and apply final processing
|
829 |
+
self.post_init()
|
830 |
+
|
831 |
+
def set_decoder(self, decoder):
|
832 |
+
self.model = decoder
|
833 |
+
|
834 |
+
def get_decoder(self):
|
835 |
+
return self.model
|
836 |
+
|
837 |
+
def get_reasoning_mode_probabilities(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
838 |
+
"""Get probabilities for different reasoning modes: think, ser, pet, normal"""
|
839 |
+
if self.use_structured_output:
|
840 |
+
# Use the last token's hidden state for mode classification
|
841 |
+
last_hidden = hidden_states[:, -1, :] # [batch_size, hidden_size]
|
842 |
+
mode_probs = self.reasoning_mode_classifier(last_hidden)
|
843 |
+
return mode_probs
|
844 |
+
return None
|
845 |
+
|
846 |
+
@can_return_tuple
|
847 |
+
@auto_docstring
|
848 |
+
def forward(
|
849 |
+
self,
|
850 |
+
input_ids: Optional[torch.LongTensor] = None,
|
851 |
+
attention_mask: Optional[torch.Tensor] = None,
|
852 |
+
position_ids: Optional[torch.LongTensor] = None,
|
853 |
+
past_key_values: Optional[Cache] = None,
|
854 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
855 |
+
labels: Optional[torch.LongTensor] = None,
|
856 |
+
# Optional supervision for speech frames: float tensor [B, T_frames, n_mels]
|
857 |
+
speech_targets: Optional[torch.FloatTensor] = None,
|
858 |
+
use_cache: Optional[bool] = None,
|
859 |
+
cache_position: Optional[torch.LongTensor] = None,
|
860 |
+
logits_to_keep: Union[int, torch.Tensor] = 0,
|
861 |
+
return_reasoning_metadata: Optional[bool] = False,
|
862 |
+
**kwargs: Unpack[TransformersKwargs],
|
863 |
+
) -> CausalLMOutputWithPast:
|
864 |
+
r"""
|
865 |
+
Enhanced HelpingAI forward pass with structured reasoning support.
|
866 |
+
|
867 |
+
Args:
|
868 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
869 |
+
Indices of input sequence tokens in the vocabulary.
|
870 |
+
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
871 |
+
Mask to avoid performing attention on padding token indices.
|
872 |
+
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
873 |
+
Indices of positions of each input sequence tokens in the position embeddings.
|
874 |
+
past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
|
875 |
+
Pre-computed hidden-states that can be used to speed up autoregressive decoding.
|
876 |
+
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
877 |
+
Embedded representation of the input tokens. Can be used instead of `input_ids`.
|
878 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
879 |
+
Labels for computing the masked language modeling loss.
|
880 |
+
use_cache (`bool`, *optional*):
|
881 |
+
If set to `True`, past key values are returned and can be used to speed up decoding.
|
882 |
+
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
883 |
+
Indices depicting the position of the input tokens in the sequence.
|
884 |
+
logits_to_keep (`Union[int, torch.Tensor]`, *optional*, defaults to 0):
|
885 |
+
Number of logits to keep from the end of the sequence.
|
886 |
+
return_reasoning_metadata (`bool`, *optional*, defaults to `False`):
|
887 |
+
Whether to return reasoning metadata including SER and PET analysis for structured reasoning.
|
888 |
+
|
889 |
+
Returns:
|
890 |
+
`CausalLMOutputWithPast`: Model output containing logits, past key values, and optional reasoning metadata.
|
891 |
+
|
892 |
+
Example:
|
893 |
+
|
894 |
+
```python
|
895 |
+
>>> from transformers import AutoTokenizer, HelpingAIForCausalLM
|
896 |
+
|
897 |
+
>>> model = HelpingAIForCausalLM.from_pretrained("HelpingAI/HelpingAI-8B")
|
898 |
+
>>> tokenizer = AutoTokenizer.from_pretrained("HelpingAI/HelpingAI-8B")
|
899 |
+
|
900 |
+
>>> # Standard generation
|
901 |
+
>>> prompt = "Hey, are you conscious? Can you talk to me?"
|
902 |
+
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
903 |
+
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
904 |
+
>>> response = tokenizer.batch_decode(generate_ids, skip_special_tokens=True)[0]
|
905 |
+
|
906 |
+
>>> # Structured reasoning generation
|
907 |
+
>>> outputs = model(inputs.input_ids, return_reasoning_metadata=True)
|
908 |
+
>>> reasoning_modes = model.get_reasoning_mode_probabilities(outputs.hidden_states)
|
909 |
+
```"""
|
910 |
+
outputs: BaseModelOutputWithPast = self.model(
|
911 |
+
input_ids=input_ids,
|
912 |
+
attention_mask=attention_mask,
|
913 |
+
position_ids=position_ids,
|
914 |
+
past_key_values=past_key_values,
|
915 |
+
inputs_embeds=inputs_embeds,
|
916 |
+
use_cache=use_cache,
|
917 |
+
cache_position=cache_position,
|
918 |
+
**kwargs,
|
919 |
+
)
|
920 |
+
|
921 |
+
hidden_states = outputs.last_hidden_state
|
922 |
+
|
923 |
+
# Standard language modeling head
|
924 |
+
slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
|
925 |
+
logits = self.lm_head(hidden_states[:, slice_indices, :])
|
926 |
+
|
927 |
+
# Enhanced structured output logits
|
928 |
+
structured_logits = None
|
929 |
+
reasoning_mode_probs = None
|
930 |
+
if self.use_structured_output:
|
931 |
+
structured_logits = self.structured_lm_head(hidden_states[:, slice_indices, :])
|
932 |
+
reasoning_mode_probs = self.get_reasoning_mode_probabilities(hidden_states)
|
933 |
+
|
934 |
+
# Speech output prediction
|
935 |
+
speech_mels = None
|
936 |
+
if self.use_speech_output:
|
937 |
+
token_level = hidden_states # [B, T_tok, H]
|
938 |
+
# Simple temporal upsampling by repetition to approximate frame rate
|
939 |
+
if getattr(self, "speech_upsample_factor", 1) > 1:
|
940 |
+
token_level = token_level.repeat_interleave(self.speech_upsample_factor, dim=1)
|
941 |
+
# Project to mel bins per (upsampled) time-step
|
942 |
+
speech_mels = self.speech_proj(token_level) # [B, T_frames, n_mels]
|
943 |
+
|
944 |
+
loss = None
|
945 |
+
if labels is not None:
|
946 |
+
# Standard loss computation
|
947 |
+
loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
|
948 |
+
|
949 |
+
# Add structured output loss if applicable
|
950 |
+
if self.use_structured_output and structured_logits is not None:
|
951 |
+
# Additional loss term for structured reasoning (if labels include structured tokens)
|
952 |
+
structured_loss_weight = 0.1 # Weight for structured output loss
|
953 |
+
structured_loss = self.loss_function(
|
954 |
+
logits=structured_logits,
|
955 |
+
labels=labels,
|
956 |
+
vocab_size=self.structured_vocab_size,
|
957 |
+
**kwargs
|
958 |
+
)
|
959 |
+
loss = loss + (structured_loss_weight * structured_loss)
|
960 |
+
|
961 |
+
# Add speech supervision if provided
|
962 |
+
if self.use_speech_output and speech_targets is not None:
|
963 |
+
# Ensure time dimension alignment by trimming or padding speech_mels to targets
|
964 |
+
B, T_pred, M = speech_mels.shape
|
965 |
+
B2, T_tgt, M2 = speech_targets.shape
|
966 |
+
if B != B2 or M != M2:
|
967 |
+
raise ValueError("speech_targets shape mismatch. Expected [B, T, n_mels] with same B and n_mels as model output.")
|
968 |
+
if T_pred > T_tgt:
|
969 |
+
speech_mels_aligned = speech_mels[:, :T_tgt, :]
|
970 |
+
elif T_pred < T_tgt:
|
971 |
+
pad = torch.zeros(B, T_tgt - T_pred, M, device=speech_mels.device, dtype=speech_mels.dtype)
|
972 |
+
speech_mels_aligned = torch.cat([speech_mels, pad], dim=1)
|
973 |
+
else:
|
974 |
+
speech_mels_aligned = speech_mels
|
975 |
+
|
976 |
+
if self.speech_loss_type == "mse":
|
977 |
+
speech_loss = nn.functional.mse_loss(speech_mels_aligned, speech_targets)
|
978 |
+
else:
|
979 |
+
speech_loss = nn.functional.l1_loss(speech_mels_aligned, speech_targets)
|
980 |
+
loss = speech_loss if loss is None else (loss + speech_loss)
|
981 |
+
|
982 |
+
# Prepare output with enhanced reasoning metadata
|
983 |
+
output = CausalLMOutputWithPast(
|
984 |
+
loss=loss,
|
985 |
+
logits=logits,
|
986 |
+
past_key_values=outputs.past_key_values,
|
987 |
+
hidden_states=outputs.hidden_states,
|
988 |
+
attentions=outputs.attentions,
|
989 |
+
)
|
990 |
+
|
991 |
+
# Add custom attributes for reasoning
|
992 |
+
if return_reasoning_metadata and self.use_structured_output:
|
993 |
+
output.structured_logits = structured_logits
|
994 |
+
output.reasoning_mode_probabilities = reasoning_mode_probs
|
995 |
+
if self.use_speech_output:
|
996 |
+
output.speech_mels = speech_mels
|
997 |
+
|
998 |
+
return output
|
999 |
+
|
1000 |
+
|
1001 |
+
class HelpingAIForSequenceClassification(GenericForSequenceClassification, HelpingAIPreTrainedModel):
|
1002 |
+
pass
|
1003 |
+
|
1004 |
+
|
1005 |
+
class HelpingAIForTokenClassification(GenericForTokenClassification, HelpingAIPreTrainedModel):
|
1006 |
+
pass
|
1007 |
+
|
1008 |
+
|
1009 |
+
class HelpingAIForQuestionAnswering(GenericForQuestionAnswering, HelpingAIPreTrainedModel):
|
1010 |
+
base_model_prefix = "transformer" # For BC, where `transformer` was used instead of `model`
|
1011 |
+
|
1012 |
+
|
1013 |
+
__all__ = [
|
1014 |
+
"HelpingAIForCausalLM",
|
1015 |
+
"HelpingAIForQuestionAnswering",
|
1016 |
+
"HelpingAIPreTrainedModel",
|
1017 |
+
"HelpingAIModel",
|
1018 |
+
"HelpingAIForSequenceClassification",
|
1019 |
+
"HelpingAIForTokenClassification",
|
1020 |
+
]
|