Add custom model class with proper latent attention architecture
Browse files- modeling_llm2vec4cxr.py +55 -0
modeling_llm2vec4cxr.py
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Custom model class for LLM2Vec4CXR that properly handles latent attention pooling.
|
| 3 |
+
"""
|
| 4 |
+
|
| 5 |
+
from llm2vec.models.bidirectional_llama import LlamaBiModel
|
| 6 |
+
from llm2vec.pooling import LatentAttentionPooling
|
| 7 |
+
import torch
|
| 8 |
+
import torch.nn as nn
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
class LLM2Vec4CXRModel(LlamaBiModel):
|
| 12 |
+
"""
|
| 13 |
+
Custom LlamaBiModel that includes latent attention pooling by default.
|
| 14 |
+
This prevents the warning about unused latent attention weights.
|
| 15 |
+
"""
|
| 16 |
+
|
| 17 |
+
def __init__(self, config, **kwargs):
|
| 18 |
+
super().__init__(config, **kwargs)
|
| 19 |
+
|
| 20 |
+
# Initialize latent attention pooling
|
| 21 |
+
self.latent_attn = LatentAttentionPooling(
|
| 22 |
+
d_model=config.hidden_size,
|
| 23 |
+
num_heads=8, # Standard for this model size
|
| 24 |
+
num_latents=512 # Standard for LLM2Vec
|
| 25 |
+
)
|
| 26 |
+
|
| 27 |
+
# Move to the same device/dtype as the base model
|
| 28 |
+
if hasattr(self, 'model') and hasattr(self.model, 'embed_tokens'):
|
| 29 |
+
device = self.model.embed_tokens.weight.device
|
| 30 |
+
dtype = self.model.embed_tokens.weight.dtype
|
| 31 |
+
self.latent_attn = self.latent_attn.to(device=device, dtype=dtype)
|
| 32 |
+
|
| 33 |
+
def forward(self, input_ids, attention_mask=None, embed_mask=None, **kwargs):
|
| 34 |
+
"""
|
| 35 |
+
Forward pass that properly handles latent attention pooling.
|
| 36 |
+
"""
|
| 37 |
+
# Get base model output
|
| 38 |
+
outputs = super().forward(input_ids, attention_mask=attention_mask, **kwargs)
|
| 39 |
+
|
| 40 |
+
# If we have latent attention pooling, apply it
|
| 41 |
+
if hasattr(self, 'latent_attn') and self.latent_attn is not None:
|
| 42 |
+
if embed_mask is not None:
|
| 43 |
+
# Use embed_mask for instruction-following tasks
|
| 44 |
+
pooled_output = self.latent_attn(outputs.last_hidden_state, embed_mask)
|
| 45 |
+
else:
|
| 46 |
+
# Use attention_mask for simple encoding
|
| 47 |
+
pooled_output = self.latent_attn(outputs.last_hidden_state, attention_mask)
|
| 48 |
+
return pooled_output
|
| 49 |
+
|
| 50 |
+
return outputs.last_hidden_state
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
# Register the model for auto loading
|
| 54 |
+
from transformers import AutoModel
|
| 55 |
+
AutoModel.register(LLM2Vec4CXRModel.__name__, LLM2Vec4CXRModel)
|