phi3-M3-V2 / architecture.py
moelanoby's picture
Upload model weights and config
bda169a verified
# --- START OF FILE architecture.py ---
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import Phi3Config, Phi3ForCausalLM
from typing import Optional, Dict, List
# --- BUILDING BLOCK 1: Hierarchical VectorMemoryHead ---
# This version is improved with a hierarchical memory system (L1/L2 cache)
# to handle much longer contexts and a gated update mechanism for stability.
class VectorMemoryHead(nn.Module):
def __init__(self, hidden_dim: int, num_memory_slots: int, num_heads: int, ff_dim: int,
num_long_term_memory_slots: int = 0, # <-- NEW: Size of the L2 memory cache
device=None, dtype=None):
super().__init__()
self.hidden_dim = hidden_dim
self.num_memory_slots = num_memory_slots # L1 cache size
self.num_long_term_memory_slots = num_long_term_memory_slots # L2 cache size
# --- L1 Working Memory Components (same as before) ---
encoder_layer = nn.TransformerEncoderLayer(
d_model=hidden_dim, nhead=num_heads, dim_feedforward=ff_dim, dropout=0.1, batch_first=True,
device=device, dtype=dtype
)
self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=1)
self.memory_queries = nn.Parameter(torch.randn(1, num_memory_slots, hidden_dim, device=device, dtype=dtype))
self.memory_attention = nn.MultiheadAttention(
embed_dim=hidden_dim, num_heads=num_heads, dropout=0.1, batch_first=True,
device=device, dtype=dtype
)
self.memory_layernorm = nn.LayerNorm(hidden_dim, device=device, dtype=dtype)
self.decoder_attention = nn.MultiheadAttention(
embed_dim=hidden_dim, num_heads=num_heads, dropout=0.1, batch_first=True,
device=device, dtype=dtype
)
self.decoder_layernorm = nn.LayerNorm(hidden_dim, device=device, dtype=dtype)
self.decoder_ffn = nn.Sequential(
nn.Linear(hidden_dim, ff_dim, device=device, dtype=dtype),
nn.ReLU(),
nn.Linear(ff_dim, hidden_dim, device=device, dtype=dtype)
)
# --- NEW: L2 Long-Term Memory Components ---
self.use_long_term_memory = self.num_long_term_memory_slots > 0
if self.use_long_term_memory:
self.long_term_memory = nn.Parameter(
torch.zeros(1, self.num_long_term_memory_slots, hidden_dim, device=device, dtype=dtype)
)
# Gate for updating long-term memory (similar to GRU/LSTM gates)
self.memory_update_gate = nn.Sequential(
nn.Linear(2 * hidden_dim, hidden_dim, device=device, dtype=dtype),
nn.Sigmoid()
)
# Attention to read from L2 memory
self.ltm_retrieval_attention = nn.MultiheadAttention(
embed_dim=hidden_dim, num_heads=num_heads, dropout=0.1, batch_first=True,
device=device, dtype=dtype
)
def forward(self, memory_input_sequence: torch.Tensor):
batch_size = memory_input_sequence.shape[0]
# 1. Encode input sequence
encoded_vectors = self.encoder(memory_input_sequence)
# 2. Compress into L1 working memory
queries = self.memory_queries.expand(batch_size, -1, -1)
compressed_memory, _ = self.memory_attention(query=queries, key=encoded_vectors, value=encoded_vectors)
compressed_memory = self.memory_layernorm(compressed_memory + queries) # (B, num_memory_slots, D)
final_memory_context = compressed_memory
# --- NEW: Interact with L2 Long-Term Memory ---
if self.use_long_term_memory and self.long_term_memory.shape[0] == batch_size:
# 3a. Retrieve relevant context from L2 memory using L1 as query
retrieved_ltm, _ = self.ltm_retrieval_attention(
query=compressed_memory,
key=self.long_term_memory,
value=self.long_term_memory
)
# 3b. Gated update of the Long-Term Memory
# Average the L1 memory to get a summary vector for the update
l1_summary = compressed_memory.mean(dim=1)
ltm_summary = self.long_term_memory.mean(dim=1)
gate_input = torch.cat([l1_summary, ltm_summary], dim=-1)
update_gate = self.memory_update_gate(gate_input).unsqueeze(1) # (B, 1, D)
# Update LTM by blending new info from L1
self.long_term_memory.data = (update_gate * l1_summary.unsqueeze(1)) + ((1 - update_gate) * self.long_term_memory.data)
# Combine L1 and retrieved L2 context for the final output
final_memory_context = final_memory_context + retrieved_ltm
# 4. Decode from the final memory context to reconstruct original sequence
reconstructed, _ = self.decoder_attention(query=encoded_vectors, key=final_memory_context, value=final_memory_context)
reconstructed_vectors = self.decoder_layernorm(reconstructed + encoded_vectors)
reconstructed_vectors = self.decoder_ffn(reconstructed_vectors)
return compressed_memory, reconstructed_vectors
# --- BUILDING BLOCK 2: The Custom Layer (With Per-Dataset Parameters and Refinement) ---
class GCVectorMemoryLayer(nn.Module):
def __init__(self, original_layer: nn.Linear, global_input_dim: int,
memory_dim: int, num_memory_slots: int, memory_num_heads: int,
global_state_storage: Dict, dataset_keys: List[str]):
super().__init__()
self.input_dim = original_layer.in_features
self.output_dim = original_layer.out_features
self.memory_dim = memory_dim
self.global_state_storage = global_state_storage
self.dataset_keys = dataset_keys
self.linear = original_layer # Shared linear layer
device, dtype = self.linear.weight.device, self.linear.weight.dtype
# --- NEW: Per-dataset specialized parameters ---
self.local_state_projs = nn.ModuleDict()
self.global_state_projs = nn.ModuleDict()
self.memory_heads = nn.ModuleDict()
self.correction_heads = nn.ModuleDict()
for key in self.dataset_keys:
self.local_state_projs[key] = nn.Linear(self.input_dim, memory_dim, device=device, dtype=dtype)
self.global_state_projs[key] = nn.Linear(global_input_dim, memory_dim, device=device, dtype=dtype)
self.memory_heads[key] = VectorMemoryHead(
hidden_dim=memory_dim, num_memory_slots=num_memory_slots,
num_heads=memory_num_heads, ff_dim=memory_dim * 2,
num_long_term_memory_slots=32, # Enable L2 Cache
device=device, dtype=dtype
)
self.correction_heads[key] = nn.Linear(memory_dim, 2 * self.output_dim, device=device, dtype=dtype)
self.refinement_passes: int = 2 # Default to 2 passes for deeper refinement
self.last_corrected_activation: Optional[torch.Tensor] = None
self.last_additive_correction: Optional[torch.Tensor] = None
self.last_memory_input: Optional[torch.Tensor] = None
self.last_reconstructed_from_memory: Optional[torch.Tensor] = None
def forward(self, x: torch.Tensor):
base_output = self.linear(x)
# Determine which set of specialized parameters to use
dataset_key = self.global_state_storage.get('dataset_key')
if not dataset_key or 'embeds' not in self.global_state_storage or self.refinement_passes < 1:
return base_output
# Select the correct modules for the current context
local_state_proj = self.local_state_projs[dataset_key]
global_state_proj = self.global_state_projs[dataset_key]
memory_head = self.memory_heads[dataset_key]
correction_head = self.correction_heads[dataset_key]
global_embeds = self.global_state_storage['embeds']
if global_embeds.shape[1] != x.shape[1]: global_embeds = global_embeds[:, -x.shape[1]:, :]
B, S, _ = x.shape
# Ensure LTM is initialized with correct batch size for the specific memory head
if memory_head.use_long_term_memory and memory_head.long_term_memory.shape[0] != B:
memory_head.long_term_memory.data = memory_head.long_term_memory.data.expand(B, -1, -1)
with torch.no_grad(): # Use no_grad for the refinement loop as it's an inference-like process
proj_local = local_state_proj(x.detach())
proj_global = global_state_proj(global_embeds.detach())
memory_input = torch.stack([proj_global, proj_local], dim=2)
memory_input_flat = memory_input.view(B * S, 2, self.memory_dim)
compressed_mem_flat, _ = memory_head(memory_input_flat)
aggregated_thought = compressed_mem_flat.mean(dim=1).view(B, S, self.memory_dim)
# Iteratively refine the output using state-feedback
corrected_activation = base_output
current_thought = aggregated_thought
for _ in range(self.refinement_passes):
raw_correction = correction_head(current_thought)
gate, value = torch.chunk(raw_correction, 2, dim=-1)
corrected_activation = corrected_activation * torch.sigmoid(gate) + value
current_thought_flat = current_thought.view(B * S, self.memory_dim)
refined_thought, _ = memory_head.decoder_attention(
query=current_thought_flat.unsqueeze(1), key=compressed_mem_flat, value=compressed_mem_flat
)
refined_thought = memory_head.decoder_layernorm(refined_thought.squeeze(1) + current_thought_flat)
current_thought = refined_thought.view(B, S, self.memory_dim)
if self.training:
with torch.enable_grad():
proj_local_grad = local_state_proj(x)
proj_global_grad = global_state_proj(global_embeds)
memory_input_grad = torch.stack([proj_global_grad, proj_local_grad], dim=2)
memory_input_flat_grad = memory_input_grad.view(B * S, 2, self.memory_dim)
compressed_mem_flat_grad, recon_flat_grad = memory_head(memory_input_flat_grad)
aggregated_thought_grad = compressed_mem_flat_grad.mean(dim=1).view(B, S, self.memory_dim)
raw_correction_grad = correction_head(aggregated_thought_grad)
gate_grad, value_grad = torch.chunk(raw_correction_grad, 2, dim=-1)
final_activation = base_output * torch.sigmoid(gate_grad.to(x.dtype)) + value_grad.to(x.dtype)
self.last_corrected_activation = final_activation
self.last_additive_correction = value_grad
self.last_memory_input = memory_input_flat_grad
self.last_reconstructed_from_memory = recon_flat_grad
return final_activation
else:
return corrected_activation.to(x.dtype)
# --- BUILDING BLOCK 3: The Full Custom Model Wrapper (for saving/loading) ---
class Phi3WithVectorMemoryForCausalLM(Phi3ForCausalLM):
def __init__(self, config):
super().__init__(config)
self.global_state_storage = {}
# Target a central layer in the network for maximum impact
self.target_layer_path = "model.layers.15.mlp.gate_up_proj"
self.model.embed_tokens.register_forward_hook(
lambda module, input, output: self.global_state_storage.update({'embeds': output.detach()})
)
# This logic is primarily for loading a pre-trained model.
# The training script handles the initial creation.
if hasattr(config, "dataset_keys") and config.dataset_keys:
try:
print(f"Re-initializing GCVectorMemoryLayer with dataset keys: {config.dataset_keys}")
original_layer = self.get_submodule(self.target_layer_path)
custom_layer = GCVectorMemoryLayer(
original_layer=original_layer, global_input_dim=config.hidden_size,
memory_dim=64,
num_memory_slots=8,
memory_num_heads=4,
global_state_storage=self.global_state_storage,
dataset_keys=config.dataset_keys # Use keys from config
)
parent_path = ".".join(self.target_layer_path.split('.')[:-1])
child_name = self.target_layer_path.split('.')[-1]
setattr(self.get_submodule(parent_path), child_name, custom_layer)
print(f"Successfully reloaded and replaced '{self.target_layer_path}' with specialized GCVectorMemoryLayer.")
except AttributeError:
print(f"Could not find target layer '{self.target_layer_path}' during reload. Model remains unmodified.")
else:
print("No 'dataset_keys' found in config. The custom layer will not be initialized.")