Spaces:
Sleeping
Sleeping
Upload 4 files
Browse files- system/inference.py +138 -0
- system/mambaSwarm.py +816 -0
- system/memory_manager.py +306 -0
- system/weight_manager.py +168 -0
system/inference.py
ADDED
@@ -0,0 +1,138 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# =============================================================================
|
2 |
+
# system/inference.py
|
3 |
+
# =============================================================================
|
4 |
+
import torch
|
5 |
+
from typing import Dict, List, Optional, Union
|
6 |
+
import time
|
7 |
+
|
8 |
+
class MambaInferenceEngine:
|
9 |
+
"""Optimized inference engine for Mamba swarm"""
|
10 |
+
|
11 |
+
def __init__(self, swarm_engine):
|
12 |
+
self.swarm_engine = swarm_engine
|
13 |
+
self.config = swarm_engine.config
|
14 |
+
|
15 |
+
# Inference optimizations
|
16 |
+
self.use_half_precision = True
|
17 |
+
self.use_torch_compile = hasattr(torch, 'compile')
|
18 |
+
|
19 |
+
# Apply optimizations
|
20 |
+
self._optimize_models()
|
21 |
+
|
22 |
+
def _optimize_models(self):
|
23 |
+
"""Apply inference optimizations"""
|
24 |
+
if self.use_half_precision and self.config.device != 'cpu':
|
25 |
+
# Convert to half precision for faster inference
|
26 |
+
for specialist in self.swarm_engine.tlm_manager.specialists.values():
|
27 |
+
specialist.model = specialist.model.half()
|
28 |
+
self.swarm_engine.aggregator = self.swarm_engine.aggregator.half()
|
29 |
+
|
30 |
+
if self.use_torch_compile:
|
31 |
+
try:
|
32 |
+
# Compile models for faster inference (PyTorch 2.0+)
|
33 |
+
for specialist in self.swarm_engine.tlm_manager.specialists.values():
|
34 |
+
specialist.model = torch.compile(specialist.model)
|
35 |
+
self.swarm_engine.aggregator = torch.compile(self.swarm_engine.aggregator)
|
36 |
+
print("Models compiled for faster inference")
|
37 |
+
except Exception as e:
|
38 |
+
print(f"Could not compile models: {e}")
|
39 |
+
|
40 |
+
def generate(self, prompt: str, max_tokens: int = 100,
|
41 |
+
temperature: float = 0.7, top_k: int = 50) -> Dict:
|
42 |
+
"""
|
43 |
+
Generate text response with advanced sampling
|
44 |
+
|
45 |
+
Args:
|
46 |
+
prompt: Input text prompt
|
47 |
+
max_tokens: Maximum tokens to generate
|
48 |
+
temperature: Sampling temperature
|
49 |
+
top_k: Top-k sampling parameter
|
50 |
+
|
51 |
+
Returns:
|
52 |
+
Dict with generated text and metadata
|
53 |
+
"""
|
54 |
+
start_time = time.time()
|
55 |
+
|
56 |
+
# Process through swarm
|
57 |
+
result = self.swarm_engine.process_request(prompt, max_tokens)
|
58 |
+
|
59 |
+
if not result['success']:
|
60 |
+
return result
|
61 |
+
|
62 |
+
# Add inference metadata
|
63 |
+
result.update({
|
64 |
+
'temperature': temperature,
|
65 |
+
'top_k': top_k,
|
66 |
+
'inference_time': time.time() - start_time,
|
67 |
+
'tokens_per_second': max_tokens / (time.time() - start_time)
|
68 |
+
})
|
69 |
+
|
70 |
+
return result
|
71 |
+
|
72 |
+
def stream_generate(self, prompt: str, max_tokens: int = 100):
|
73 |
+
"""
|
74 |
+
Stream generation token by token (placeholder implementation)
|
75 |
+
"""
|
76 |
+
# This would implement streaming generation
|
77 |
+
# For now, return the full response
|
78 |
+
result = self.generate(prompt, max_tokens)
|
79 |
+
yield result['response']
|
80 |
+
|
81 |
+
def chat_completion(self, messages: List[Dict], max_tokens: int = 100) -> Dict:
|
82 |
+
"""
|
83 |
+
Chat completion interface similar to OpenAI API
|
84 |
+
|
85 |
+
Args:
|
86 |
+
messages: List of message dicts with 'role' and 'content'
|
87 |
+
max_tokens: Maximum tokens to generate
|
88 |
+
|
89 |
+
Returns:
|
90 |
+
Chat completion response
|
91 |
+
"""
|
92 |
+
# Convert messages to single prompt
|
93 |
+
prompt = self._format_chat_prompt(messages)
|
94 |
+
|
95 |
+
# Generate response
|
96 |
+
result = self.generate(prompt, max_tokens)
|
97 |
+
|
98 |
+
if result['success']:
|
99 |
+
# Format as chat completion
|
100 |
+
return {
|
101 |
+
'choices': [{
|
102 |
+
'message': {
|
103 |
+
'role': 'assistant',
|
104 |
+
'content': result['response']
|
105 |
+
},
|
106 |
+
'finish_reason': 'stop'
|
107 |
+
}],
|
108 |
+
'usage': {
|
109 |
+
'prompt_tokens': len(prompt.split()),
|
110 |
+
'completion_tokens': len(result['response'].split()),
|
111 |
+
'total_tokens': len(prompt.split()) + len(result['response'].split())
|
112 |
+
},
|
113 |
+
'model': 'mamba-swarm-70m',
|
114 |
+
'inference_time': result.get('inference_time', 0)
|
115 |
+
}
|
116 |
+
else:
|
117 |
+
return {
|
118 |
+
'error': result.get('error', 'Unknown error'),
|
119 |
+
'success': False
|
120 |
+
}
|
121 |
+
|
122 |
+
def _format_chat_prompt(self, messages: List[Dict]) -> str:
|
123 |
+
"""Format chat messages into a single prompt"""
|
124 |
+
formatted = ""
|
125 |
+
|
126 |
+
for message in messages:
|
127 |
+
role = message.get('role', 'user')
|
128 |
+
content = message.get('content', '')
|
129 |
+
|
130 |
+
if role == 'system':
|
131 |
+
formatted += f"System: {content}\n"
|
132 |
+
elif role == 'user':
|
133 |
+
formatted += f"User: {content}\n"
|
134 |
+
elif role == 'assistant':
|
135 |
+
formatted += f"Assistant: {content}\n"
|
136 |
+
|
137 |
+
formatted += "Assistant: "
|
138 |
+
return formatted
|
system/mambaSwarm.py
ADDED
@@ -0,0 +1,816 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# =============================================================================
|
2 |
+
# system/mambaSwarm.py - Unified Scalable Mamba Swarm Engine
|
3 |
+
# =============================================================================
|
4 |
+
import torch
|
5 |
+
import time
|
6 |
+
import os
|
7 |
+
import asyncio
|
8 |
+
from typing import Dict, List, Tuple, Optional, Union
|
9 |
+
from concurrent.futures import ThreadPoolExecutor
|
10 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
11 |
+
|
12 |
+
# Core imports
|
13 |
+
from core.config import MambaConfig, MambaSwarmConfig, auto_detect_tier
|
14 |
+
from core.tokenizer import MambaTokenizer
|
15 |
+
from core.preprocess import TextPreprocessor
|
16 |
+
from core.model import MambaModel
|
17 |
+
from core.mamba_swarm_integration import MambaEncoderSwarmModel, create_swarm_from_existing_config
|
18 |
+
|
19 |
+
# Routing imports
|
20 |
+
from routing.router import TopicRouter, ContentBasedRouter
|
21 |
+
from routing.tlm_manager import TLMManager
|
22 |
+
from routing.aggregator import AttentionAggregator, WeightedAggregator
|
23 |
+
from utils.domain_configs import DomainConfigs
|
24 |
+
|
25 |
+
|
26 |
+
class UnifiedMambaSwarm:
|
27 |
+
"""
|
28 |
+
Unified Mamba Swarm Engine combining the best of both architectures:
|
29 |
+
- Scalable tier-based system with auto-detection
|
30 |
+
- Production-ready async processing and monitoring
|
31 |
+
- Graceful fallback to simulation mode
|
32 |
+
- Support for both custom and pre-trained models
|
33 |
+
"""
|
34 |
+
|
35 |
+
def __init__(self,
|
36 |
+
tier: Optional[str] = None,
|
37 |
+
config: Optional[Union[MambaConfig, MambaSwarmConfig]] = None,
|
38 |
+
use_pretrained: bool = True,
|
39 |
+
config_override: Optional[Dict] = None):
|
40 |
+
"""
|
41 |
+
Initialize the unified swarm engine
|
42 |
+
|
43 |
+
Args:
|
44 |
+
tier: Scaling tier (demo/small/medium/large/full) or None for auto-detect
|
45 |
+
config: Either MambaConfig for custom models or MambaSwarmConfig for scaling
|
46 |
+
use_pretrained: Whether to use HuggingFace pretrained models
|
47 |
+
config_override: Dictionary to override config settings
|
48 |
+
"""
|
49 |
+
# Auto-detect tier if not specified
|
50 |
+
if tier is None:
|
51 |
+
tier = auto_detect_tier()
|
52 |
+
print(f"Auto-detected tier: {tier}")
|
53 |
+
|
54 |
+
self.tier = tier
|
55 |
+
self.use_pretrained = use_pretrained
|
56 |
+
|
57 |
+
# Initialize configuration
|
58 |
+
if config is None:
|
59 |
+
if use_pretrained:
|
60 |
+
self.swarm_config = MambaSwarmConfig(tier=tier)
|
61 |
+
if config_override:
|
62 |
+
self.swarm_config.config.update(config_override)
|
63 |
+
self.config = self._create_legacy_config()
|
64 |
+
else:
|
65 |
+
# Use custom config for legacy components
|
66 |
+
self.config = MambaConfig() # Default config
|
67 |
+
self.swarm_config = None
|
68 |
+
else:
|
69 |
+
if isinstance(config, MambaSwarmConfig):
|
70 |
+
self.swarm_config = config
|
71 |
+
self.config = self._create_legacy_config()
|
72 |
+
else:
|
73 |
+
self.config = config
|
74 |
+
self.swarm_config = None
|
75 |
+
|
76 |
+
self.device = getattr(self.config, 'device', 'cuda' if torch.cuda.is_available() else 'cpu')
|
77 |
+
|
78 |
+
# System properties
|
79 |
+
if self.swarm_config:
|
80 |
+
self.num_encoders = self.swarm_config.config["num_encoders"]
|
81 |
+
self.encoder_size = self.swarm_config.config["encoder_size"]
|
82 |
+
else:
|
83 |
+
self.num_encoders = getattr(self.config, 'num_specialists', 5)
|
84 |
+
self.encoder_size = "130M"
|
85 |
+
|
86 |
+
# Initialize components
|
87 |
+
self.encoders = []
|
88 |
+
self.tokenizer = None
|
89 |
+
self.preprocessor = None
|
90 |
+
self.router = None
|
91 |
+
self.aggregator = None
|
92 |
+
self.tlm_manager = None
|
93 |
+
|
94 |
+
# Performance tracking
|
95 |
+
self.stats = {
|
96 |
+
'total_requests': 0,
|
97 |
+
'total_tokens_processed': 0,
|
98 |
+
'avg_response_time': 0.0,
|
99 |
+
'specialist_usage': {i: 0 for i in range(self.num_encoders)},
|
100 |
+
'simulation_mode': False,
|
101 |
+
'model_load_errors': 0
|
102 |
+
}
|
103 |
+
|
104 |
+
# Initialize system
|
105 |
+
self._initialize_system()
|
106 |
+
|
107 |
+
print(f"✅ Unified Mamba Swarm initialized: {self.tier} tier, {self.num_encoders} encoders")
|
108 |
+
|
109 |
+
def _create_legacy_config(self) -> MambaConfig:
|
110 |
+
"""Create legacy MambaConfig from SwarmConfig for compatibility"""
|
111 |
+
legacy_config = MambaConfig()
|
112 |
+
if self.swarm_config:
|
113 |
+
legacy_config.num_specialists = self.swarm_config.config["num_encoders"]
|
114 |
+
legacy_config.device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
115 |
+
return legacy_config
|
116 |
+
|
117 |
+
def _initialize_system(self):
|
118 |
+
"""Initialize the complete swarm system"""
|
119 |
+
try:
|
120 |
+
# Initialize tokenizer and preprocessor
|
121 |
+
self._initialize_tokenizer()
|
122 |
+
self._initialize_preprocessor()
|
123 |
+
|
124 |
+
# Initialize encoders/specialists
|
125 |
+
if self.use_pretrained:
|
126 |
+
self._initialize_pretrained_encoders()
|
127 |
+
else:
|
128 |
+
self._initialize_custom_specialists()
|
129 |
+
|
130 |
+
# Initialize routing system
|
131 |
+
self._initialize_routing()
|
132 |
+
|
133 |
+
# Initialize aggregation system
|
134 |
+
self._initialize_aggregation()
|
135 |
+
|
136 |
+
print(f"🚀 System initialization complete!")
|
137 |
+
|
138 |
+
except Exception as e:
|
139 |
+
print(f"⚠️ Error during initialization: {e}")
|
140 |
+
self._fallback_to_simulation()
|
141 |
+
|
142 |
+
def _initialize_tokenizer(self):
|
143 |
+
"""Initialize tokenizer based on mode"""
|
144 |
+
if self.use_pretrained:
|
145 |
+
base_model_name = self._get_base_model_name()
|
146 |
+
try:
|
147 |
+
self.tokenizer = AutoTokenizer.from_pretrained(base_model_name)
|
148 |
+
if self.tokenizer.pad_token is None:
|
149 |
+
self.tokenizer.pad_token = self.tokenizer.eos_token
|
150 |
+
print(f"📝 Loaded HuggingFace tokenizer: {base_model_name}")
|
151 |
+
except:
|
152 |
+
print("⚠️ HuggingFace tokenizer failed, using custom tokenizer")
|
153 |
+
self.tokenizer = MambaTokenizer(self.config)
|
154 |
+
else:
|
155 |
+
self.tokenizer = MambaTokenizer(self.config)
|
156 |
+
|
157 |
+
def _initialize_preprocessor(self):
|
158 |
+
"""Initialize text preprocessor"""
|
159 |
+
self.preprocessor = TextPreprocessor(self.config)
|
160 |
+
|
161 |
+
def _get_base_model_name(self):
|
162 |
+
"""Get the appropriate base model for current tier"""
|
163 |
+
model_mapping = {
|
164 |
+
"130M": "state-spaces/mamba-130m",
|
165 |
+
"370M": "state-spaces/mamba-370m",
|
166 |
+
"790M": "state-spaces/mamba-790m",
|
167 |
+
"1.4B": "state-spaces/mamba-1.4b",
|
168 |
+
"2.8B": "state-spaces/mamba-2.8b"
|
169 |
+
}
|
170 |
+
return model_mapping.get(self.encoder_size, "state-spaces/mamba-130m")
|
171 |
+
|
172 |
+
def _initialize_pretrained_encoders(self):
|
173 |
+
"""Initialize pretrained encoder swarm"""
|
174 |
+
print(f"🔄 Loading {self.num_encoders} pretrained encoders...")
|
175 |
+
|
176 |
+
base_model_name = self._get_base_model_name()
|
177 |
+
|
178 |
+
try:
|
179 |
+
# Load base model
|
180 |
+
base_model = AutoModelForCausalLM.from_pretrained(
|
181 |
+
base_model_name,
|
182 |
+
torch_dtype=torch.float16 if self.num_encoders > 5 else torch.float32,
|
183 |
+
device_map="auto" if torch.cuda.is_available() else "cpu"
|
184 |
+
)
|
185 |
+
|
186 |
+
# Create encoder instances
|
187 |
+
for i in range(self.num_encoders):
|
188 |
+
domain_info = self.swarm_config.domain_assignments[i] if self.swarm_config else {
|
189 |
+
"domain": f"general_{i}", "specialty": "general"
|
190 |
+
}
|
191 |
+
|
192 |
+
if self.tier == "demo" or self.num_encoders <= 5:
|
193 |
+
# Share model instance for smaller configurations
|
194 |
+
encoder = {
|
195 |
+
"id": i,
|
196 |
+
"model": base_model,
|
197 |
+
"domain": domain_info["domain"],
|
198 |
+
"specialty": domain_info["specialty"],
|
199 |
+
"shared": True
|
200 |
+
}
|
201 |
+
else:
|
202 |
+
# Separate instances for larger configurations
|
203 |
+
encoder = {
|
204 |
+
"id": i,
|
205 |
+
"model": AutoModelForCausalLM.from_pretrained(
|
206 |
+
base_model_name,
|
207 |
+
torch_dtype=torch.float16,
|
208 |
+
device_map="auto"
|
209 |
+
),
|
210 |
+
"domain": domain_info["domain"],
|
211 |
+
"specialty": domain_info["specialty"],
|
212 |
+
"shared": False
|
213 |
+
}
|
214 |
+
|
215 |
+
self.encoders.append(encoder)
|
216 |
+
print(f" ✓ Encoder {i}: {encoder['domain']} specialist")
|
217 |
+
|
218 |
+
except Exception as e:
|
219 |
+
print(f"❌ Failed to load pretrained models: {e}")
|
220 |
+
self.stats['model_load_errors'] += 1
|
221 |
+
self._create_simulated_encoders()
|
222 |
+
|
223 |
+
def _initialize_custom_specialists(self):
|
224 |
+
"""Initialize custom TLM specialists or native Mamba swarm"""
|
225 |
+
try:
|
226 |
+
if hasattr(self, 'use_native_swarm') and self.use_native_swarm:
|
227 |
+
# Use the native Mamba swarm integration
|
228 |
+
self.native_swarm_model = create_swarm_from_existing_config(
|
229 |
+
self.config, num_encoders=self.num_encoders
|
230 |
+
)
|
231 |
+
print(f"✓ Initialized native Mamba swarm with {self.num_encoders} encoders")
|
232 |
+
else:
|
233 |
+
# Use TLM manager (legacy approach)
|
234 |
+
self.tlm_manager = TLMManager(self.config)
|
235 |
+
print(f"✓ Initialized {self.num_encoders} custom specialists")
|
236 |
+
except Exception as e:
|
237 |
+
print(f"⚠️ Custom specialists failed: {e}")
|
238 |
+
self._create_simulated_encoders()
|
239 |
+
|
240 |
+
def _create_simulated_encoders(self):
|
241 |
+
"""Create simulated encoders for demonstration/fallback"""
|
242 |
+
print("🎭 Creating simulated encoders...")
|
243 |
+
self.stats['simulation_mode'] = True
|
244 |
+
|
245 |
+
for i in range(self.num_encoders):
|
246 |
+
domain_info = self.swarm_config.domain_assignments[i] if self.swarm_config else {
|
247 |
+
"domain": f"general_{i}", "specialty": "general"
|
248 |
+
}
|
249 |
+
|
250 |
+
encoder = {
|
251 |
+
"id": i,
|
252 |
+
"model": None,
|
253 |
+
"domain": domain_info["domain"],
|
254 |
+
"specialty": domain_info["specialty"],
|
255 |
+
"simulated": True
|
256 |
+
}
|
257 |
+
self.encoders.append(encoder)
|
258 |
+
|
259 |
+
def _initialize_routing(self):
|
260 |
+
"""Initialize routing system"""
|
261 |
+
try:
|
262 |
+
if self.use_pretrained and self.swarm_config:
|
263 |
+
# Use content-based router for pretrained models
|
264 |
+
router_config = self.swarm_config.get_router_config()
|
265 |
+
self.router = ContentBasedRouter(
|
266 |
+
num_encoders=self.num_encoders,
|
267 |
+
domain_assignments=self.swarm_config.domain_assignments,
|
268 |
+
config=router_config
|
269 |
+
)
|
270 |
+
else:
|
271 |
+
# Use topic router for custom models
|
272 |
+
domain_configs = DomainConfigs.get_domain_configs(self.num_encoders)
|
273 |
+
self.router = TopicRouter(self.config, domain_configs)
|
274 |
+
if hasattr(self.router, 'to'):
|
275 |
+
self.router.to(self.device)
|
276 |
+
|
277 |
+
print("🧭 Router initialized")
|
278 |
+
|
279 |
+
except Exception as e:
|
280 |
+
print(f"⚠️ Router initialization failed: {e}")
|
281 |
+
# Create basic fallback router
|
282 |
+
self.router = self._create_fallback_router()
|
283 |
+
|
284 |
+
def _initialize_aggregation(self):
|
285 |
+
"""Initialize aggregation system"""
|
286 |
+
try:
|
287 |
+
if self.use_pretrained:
|
288 |
+
self.aggregator = WeightedAggregator(
|
289 |
+
num_encoders=self.num_encoders,
|
290 |
+
hidden_dim=768
|
291 |
+
)
|
292 |
+
else:
|
293 |
+
self.aggregator = AttentionAggregator(self.config)
|
294 |
+
if hasattr(self.aggregator, 'to'):
|
295 |
+
self.aggregator.to(self.device)
|
296 |
+
|
297 |
+
print("🔄 Aggregator initialized")
|
298 |
+
|
299 |
+
except Exception as e:
|
300 |
+
print(f"⚠️ Aggregator initialization failed: {e}")
|
301 |
+
self.aggregator = None
|
302 |
+
|
303 |
+
def _create_fallback_router(self):
|
304 |
+
"""Create a simple fallback router"""
|
305 |
+
class FallbackRouter:
|
306 |
+
def __init__(self, num_encoders):
|
307 |
+
self.num_encoders = num_encoders
|
308 |
+
|
309 |
+
def route(self, text):
|
310 |
+
# Simple round-robin routing
|
311 |
+
import random
|
312 |
+
num_selected = min(3, self.num_encoders)
|
313 |
+
return {
|
314 |
+
"selected_encoders": random.sample(range(self.num_encoders), num_selected)
|
315 |
+
}
|
316 |
+
|
317 |
+
def chunk_and_route(self, text):
|
318 |
+
return [{"specialists": [(0, 1.0)], "chunk": text}]
|
319 |
+
|
320 |
+
return FallbackRouter(self.num_encoders)
|
321 |
+
|
322 |
+
def _fallback_to_simulation(self):
|
323 |
+
"""Complete fallback to simulation mode"""
|
324 |
+
print("🎭 Entering full simulation mode")
|
325 |
+
self.stats['simulation_mode'] = True
|
326 |
+
self._create_simulated_encoders()
|
327 |
+
if not self.router:
|
328 |
+
self.router = self._create_fallback_router()
|
329 |
+
|
330 |
+
# =============================================================================
|
331 |
+
# MAIN PROCESSING METHODS
|
332 |
+
# =============================================================================
|
333 |
+
|
334 |
+
def generate(self, prompt: str, max_length: int = 100, temperature: float = 0.7,
|
335 |
+
show_routing: bool = True) -> Dict:
|
336 |
+
"""
|
337 |
+
Generate response using the swarm (from swarmEngine2 style)
|
338 |
+
|
339 |
+
Args:
|
340 |
+
prompt: Input text prompt
|
341 |
+
max_length: Maximum tokens to generate
|
342 |
+
temperature: Sampling temperature
|
343 |
+
show_routing: Whether to display routing information
|
344 |
+
|
345 |
+
Returns:
|
346 |
+
Dict with response and metadata
|
347 |
+
"""
|
348 |
+
start_time = time.time()
|
349 |
+
|
350 |
+
try:
|
351 |
+
# Route to appropriate encoders
|
352 |
+
if hasattr(self.router, 'route'):
|
353 |
+
routing_decision = self.router.route(prompt)
|
354 |
+
selected_encoders = routing_decision.get("selected_encoders", [0])
|
355 |
+
else:
|
356 |
+
# Fallback routing
|
357 |
+
selected_encoders = [0]
|
358 |
+
|
359 |
+
if show_routing:
|
360 |
+
print(f"🔀 Routing: Selected {len(selected_encoders)} encoders")
|
361 |
+
for enc_id in selected_encoders[:3]:
|
362 |
+
if enc_id < len(self.encoders):
|
363 |
+
domain = self.encoders[enc_id]["domain"]
|
364 |
+
print(f" Encoder {enc_id}: {domain}")
|
365 |
+
|
366 |
+
# Generate response
|
367 |
+
if self.stats['simulation_mode'] or any(enc.get("simulated") for enc in self.encoders):
|
368 |
+
response = self._simulate_generation(prompt, selected_encoders, max_length)
|
369 |
+
else:
|
370 |
+
response = self._real_generation(prompt, selected_encoders, max_length, temperature)
|
371 |
+
|
372 |
+
# Update statistics
|
373 |
+
processing_time = time.time() - start_time
|
374 |
+
self._update_stats_simple(prompt, selected_encoders, processing_time)
|
375 |
+
|
376 |
+
return {
|
377 |
+
"response": response,
|
378 |
+
"processing_time": processing_time,
|
379 |
+
"routing_info": {
|
380 |
+
"selected_encoders": selected_encoders,
|
381 |
+
"num_active": len(selected_encoders),
|
382 |
+
"total_encoders": self.num_encoders,
|
383 |
+
"domains": [self.encoders[i]["domain"] for i in selected_encoders
|
384 |
+
if i < len(self.encoders)]
|
385 |
+
},
|
386 |
+
"success": True
|
387 |
+
}
|
388 |
+
|
389 |
+
except Exception as e:
|
390 |
+
return {
|
391 |
+
"response": f"Error generating response: {str(e)}",
|
392 |
+
"processing_time": time.time() - start_time,
|
393 |
+
"success": False,
|
394 |
+
"error": str(e)
|
395 |
+
}
|
396 |
+
|
397 |
+
def process_request(self, text: str, max_new_tokens: int = 100) -> Dict:
|
398 |
+
"""
|
399 |
+
Process request using traditional pipeline (from swarm_engine style)
|
400 |
+
|
401 |
+
Args:
|
402 |
+
text: Input text to process
|
403 |
+
max_new_tokens: Maximum tokens to generate
|
404 |
+
|
405 |
+
Returns:
|
406 |
+
Dict with response and metadata
|
407 |
+
"""
|
408 |
+
start_time = time.time()
|
409 |
+
|
410 |
+
try:
|
411 |
+
# Step 1: Preprocess input
|
412 |
+
if self.preprocessor:
|
413 |
+
clean_text = self.preprocessor.clean_text(text)
|
414 |
+
else:
|
415 |
+
clean_text = text
|
416 |
+
|
417 |
+
# Step 2: Route to specialists
|
418 |
+
if hasattr(self.router, 'chunk_and_route'):
|
419 |
+
routing_results = self.router.chunk_and_route(clean_text)
|
420 |
+
else:
|
421 |
+
# Fallback for content-based router
|
422 |
+
routing_decision = self.router.route(clean_text)
|
423 |
+
routing_results = [{"specialists": [(enc_id, 1.0) for enc_id in routing_decision["selected_encoders"]],
|
424 |
+
"chunk": clean_text}]
|
425 |
+
|
426 |
+
# Step 3: Process chunks
|
427 |
+
if self.tlm_manager and not self.stats['simulation_mode']:
|
428 |
+
specialist_outputs = self.tlm_manager.encode_parallel(routing_results)
|
429 |
+
else:
|
430 |
+
# Simulate processing
|
431 |
+
specialist_outputs = [{"response": f"Processed chunk: {res['chunk'][:50]}..."}
|
432 |
+
for res in routing_results]
|
433 |
+
|
434 |
+
# Step 4: Aggregate results
|
435 |
+
if self.aggregator and not self.stats['simulation_mode']:
|
436 |
+
response = self.aggregator.generate_response(specialist_outputs, max_new_tokens)
|
437 |
+
else:
|
438 |
+
# Simple aggregation fallback
|
439 |
+
response = " ".join([out.get("response", "") for out in specialist_outputs])
|
440 |
+
|
441 |
+
# Update stats
|
442 |
+
processing_time = time.time() - start_time
|
443 |
+
self._update_stats(text, routing_results, processing_time)
|
444 |
+
|
445 |
+
return {
|
446 |
+
'response': response,
|
447 |
+
'processing_time': processing_time,
|
448 |
+
'chunks_processed': len(routing_results),
|
449 |
+
'specialists_used': self._get_specialists_used(routing_results),
|
450 |
+
'success': True
|
451 |
+
}
|
452 |
+
|
453 |
+
except Exception as e:
|
454 |
+
return {
|
455 |
+
'response': f"Error processing request: {str(e)}",
|
456 |
+
'processing_time': time.time() - start_time,
|
457 |
+
'success': False,
|
458 |
+
'error': str(e)
|
459 |
+
}
|
460 |
+
|
461 |
+
# =============================================================================
|
462 |
+
# ASYNC AND BATCH PROCESSING
|
463 |
+
# =============================================================================
|
464 |
+
|
465 |
+
async def process_request_async(self, text: str, max_new_tokens: int = 100) -> Dict:
|
466 |
+
"""Async version of process_request"""
|
467 |
+
loop = asyncio.get_event_loop()
|
468 |
+
|
469 |
+
with ThreadPoolExecutor() as executor:
|
470 |
+
result = await loop.run_in_executor(
|
471 |
+
executor, self.process_request, text, max_new_tokens
|
472 |
+
)
|
473 |
+
|
474 |
+
return result
|
475 |
+
|
476 |
+
async def generate_async(self, prompt: str, max_length: int = 100,
|
477 |
+
temperature: float = 0.7) -> Dict:
|
478 |
+
"""Async version of generate"""
|
479 |
+
loop = asyncio.get_event_loop()
|
480 |
+
|
481 |
+
with ThreadPoolExecutor() as executor:
|
482 |
+
result = await loop.run_in_executor(
|
483 |
+
executor, self.generate, prompt, max_length, temperature, False
|
484 |
+
)
|
485 |
+
|
486 |
+
return result
|
487 |
+
|
488 |
+
def batch_process(self, texts: List[str], max_new_tokens: int = 100,
|
489 |
+
method: str = "process") -> List[Dict]:
|
490 |
+
"""
|
491 |
+
Process multiple texts in batch
|
492 |
+
|
493 |
+
Args:
|
494 |
+
texts: List of input texts
|
495 |
+
max_new_tokens: Maximum tokens to generate
|
496 |
+
method: "process" or "generate" for processing method
|
497 |
+
"""
|
498 |
+
results = []
|
499 |
+
|
500 |
+
for text in texts:
|
501 |
+
if method == "generate":
|
502 |
+
result = self.generate(text, max_new_tokens, show_routing=False)
|
503 |
+
else:
|
504 |
+
result = self.process_request(text, max_new_tokens)
|
505 |
+
results.append(result)
|
506 |
+
|
507 |
+
return results
|
508 |
+
|
509 |
+
# =============================================================================
|
510 |
+
# GENERATION METHODS
|
511 |
+
# =============================================================================
|
512 |
+
|
513 |
+
def _simulate_generation(self, prompt: str, selected_encoders: List[int], max_length: int) -> str:
|
514 |
+
"""Simulate generation for demo/fallback purposes"""
|
515 |
+
import random
|
516 |
+
|
517 |
+
# Determine response type based on selected encoder domains
|
518 |
+
domains = [self.encoders[i]["domain"] for i in selected_encoders if i < len(self.encoders)]
|
519 |
+
|
520 |
+
if any("code" in domain.lower() for domain in domains):
|
521 |
+
return f"Here's a solution for '{prompt[:30]}...':\n\n```python\ndef solution():\n # Implementation here\n return result\n```"
|
522 |
+
elif any("medical" in domain.lower() for domain in domains):
|
523 |
+
return f"Regarding '{prompt[:30]}...': This medical topic requires careful consideration. Please consult healthcare professionals."
|
524 |
+
elif any("science" in domain.lower() for domain in domains):
|
525 |
+
return f"From a scientific perspective on '{prompt[:30]}...': Current research indicates several key factors..."
|
526 |
+
else:
|
527 |
+
return f"Thank you for asking about '{prompt[:30]}...'. Based on expertise from {len(selected_encoders)} specialized domains, here's a comprehensive response..."
|
528 |
+
|
529 |
+
def _real_generation(self, prompt: str, selected_encoders: List[int],
|
530 |
+
max_length: int, temperature: float) -> str:
|
531 |
+
"""Real generation using loaded models"""
|
532 |
+
if not selected_encoders or selected_encoders[0] >= len(self.encoders):
|
533 |
+
return "No valid encoders available for generation."
|
534 |
+
|
535 |
+
try:
|
536 |
+
# Use primary encoder for generation
|
537 |
+
primary_encoder = self.encoders[selected_encoders[0]]
|
538 |
+
|
539 |
+
if primary_encoder.get("simulated") or not primary_encoder["model"]:
|
540 |
+
return self._simulate_generation(prompt, selected_encoders, max_length)
|
541 |
+
|
542 |
+
# Tokenize input
|
543 |
+
if hasattr(self.tokenizer, 'encode'):
|
544 |
+
inputs = self.tokenizer(prompt, return_tensors="pt")
|
545 |
+
else:
|
546 |
+
# Fallback tokenization
|
547 |
+
return self._simulate_generation(prompt, selected_encoders, max_length)
|
548 |
+
|
549 |
+
# Generate with model
|
550 |
+
with torch.no_grad():
|
551 |
+
outputs = primary_encoder["model"].generate(
|
552 |
+
**inputs,
|
553 |
+
max_length=max_length,
|
554 |
+
temperature=temperature,
|
555 |
+
do_sample=True,
|
556 |
+
pad_token_id=self.tokenizer.eos_token_id if hasattr(self.tokenizer, 'eos_token_id') else 0
|
557 |
+
)
|
558 |
+
|
559 |
+
response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
|
560 |
+
# Remove original prompt from response
|
561 |
+
response = response[len(prompt):].strip()
|
562 |
+
|
563 |
+
return response if response else "Generated response was empty."
|
564 |
+
|
565 |
+
except Exception as e:
|
566 |
+
print(f"⚠️ Real generation failed: {e}")
|
567 |
+
return self._simulate_generation(prompt, selected_encoders, max_length)
|
568 |
+
|
569 |
+
# =============================================================================
|
570 |
+
# UTILITY METHODS
|
571 |
+
# =============================================================================
|
572 |
+
|
573 |
+
def _get_specialists_used(self, routing_results: List[Dict]) -> List[int]:
|
574 |
+
"""Extract specialist IDs used in routing"""
|
575 |
+
specialists_used = set()
|
576 |
+
|
577 |
+
for chunk_info in routing_results:
|
578 |
+
if 'specialists' in chunk_info:
|
579 |
+
for specialist_id, _ in chunk_info['specialists']:
|
580 |
+
specialists_used.add(specialist_id)
|
581 |
+
|
582 |
+
return list(specialists_used)
|
583 |
+
|
584 |
+
def _update_stats(self, text: str, routing_results: List[Dict], processing_time: float):
|
585 |
+
"""Update detailed performance statistics"""
|
586 |
+
self.stats['total_requests'] += 1
|
587 |
+
self.stats['total_tokens_processed'] += len(text.split())
|
588 |
+
|
589 |
+
# Update average response time
|
590 |
+
prev_avg = self.stats['avg_response_time']
|
591 |
+
n = self.stats['total_requests']
|
592 |
+
self.stats['avg_response_time'] = (prev_avg * (n-1) + processing_time) / n
|
593 |
+
|
594 |
+
# Update specialist usage
|
595 |
+
specialists_used = self._get_specialists_used(routing_results)
|
596 |
+
for specialist_id in specialists_used:
|
597 |
+
if specialist_id in self.stats['specialist_usage']:
|
598 |
+
self.stats['specialist_usage'][specialist_id] += 1
|
599 |
+
|
600 |
+
def _update_stats_simple(self, text: str, selected_encoders: List[int], processing_time: float):
|
601 |
+
"""Update simple statistics for generate method"""
|
602 |
+
self.stats['total_requests'] += 1
|
603 |
+
self.stats['total_tokens_processed'] += len(text.split())
|
604 |
+
|
605 |
+
# Update average response time
|
606 |
+
prev_avg = self.stats['avg_response_time']
|
607 |
+
n = self.stats['total_requests']
|
608 |
+
self.stats['avg_response_time'] = (prev_avg * (n-1) + processing_time) / n
|
609 |
+
|
610 |
+
# Update encoder usage
|
611 |
+
for enc_id in selected_encoders:
|
612 |
+
if enc_id in self.stats['specialist_usage']:
|
613 |
+
self.stats['specialist_usage'][enc_id] += 1
|
614 |
+
|
615 |
+
# =============================================================================
|
616 |
+
# SCALING AND MANAGEMENT
|
617 |
+
# =============================================================================
|
618 |
+
|
619 |
+
def scale_up(self, new_tier: str):
|
620 |
+
"""Scale up to a higher tier"""
|
621 |
+
if new_tier not in ["demo", "small", "medium", "large", "full"]:
|
622 |
+
raise ValueError(f"Invalid tier: {new_tier}")
|
623 |
+
|
624 |
+
print(f"🚀 Scaling from {self.tier} to {new_tier}")
|
625 |
+
|
626 |
+
# Preserve current stats
|
627 |
+
old_stats = self.stats.copy()
|
628 |
+
|
629 |
+
# Reinitialize with new tier
|
630 |
+
self.__init__(tier=new_tier, use_pretrained=self.use_pretrained)
|
631 |
+
|
632 |
+
# Restore relevant stats
|
633 |
+
self.stats['total_requests'] = old_stats['total_requests']
|
634 |
+
self.stats['total_tokens_processed'] = old_stats['total_tokens_processed']
|
635 |
+
self.stats['avg_response_time'] = old_stats['avg_response_time']
|
636 |
+
|
637 |
+
def get_system_info(self) -> Dict:
|
638 |
+
"""Get comprehensive system information"""
|
639 |
+
info = {
|
640 |
+
"tier": self.tier,
|
641 |
+
"num_encoders": self.num_encoders,
|
642 |
+
"encoder_size": self.encoder_size,
|
643 |
+
"use_pretrained": self.use_pretrained,
|
644 |
+
"simulation_mode": self.stats['simulation_mode'],
|
645 |
+
"device": self.device,
|
646 |
+
"domains": list(set(enc["domain"] for enc in self.encoders)),
|
647 |
+
}
|
648 |
+
|
649 |
+
if self.swarm_config:
|
650 |
+
info.update({
|
651 |
+
"total_parameters": self.swarm_config.config["total_params"],
|
652 |
+
"memory_estimate": self.swarm_config.config["memory_estimate"],
|
653 |
+
"hardware_recommendation": self.swarm_config.config["hardware"]
|
654 |
+
})
|
655 |
+
|
656 |
+
return info
|
657 |
+
|
658 |
+
def get_stats(self) -> Dict:
|
659 |
+
"""Get current performance statistics"""
|
660 |
+
return self.stats.copy()
|
661 |
+
|
662 |
+
def load_models(self, checkpoint_path: str):
|
663 |
+
"""Load trained models from checkpoint"""
|
664 |
+
if not os.path.exists(checkpoint_path):
|
665 |
+
print(f"❌ Checkpoint not found: {checkpoint_path}")
|
666 |
+
return
|
667 |
+
|
668 |
+
try:
|
669 |
+
checkpoint = torch.load(checkpoint_path, map_location=self.device)
|
670 |
+
|
671 |
+
# Load aggregator
|
672 |
+
if self.aggregator and 'aggregator_state' in checkpoint:
|
673 |
+
self.aggregator.load_state_dict(checkpoint['aggregator_state'])
|
674 |
+
|
675 |
+
# Load specialists (if using custom models)
|
676 |
+
if self.tlm_manager and 'specialist_states' in checkpoint:
|
677 |
+
for specialist_id, state_dict in checkpoint['specialist_states'].items():
|
678 |
+
if specialist_id in self.tlm_manager.specialists:
|
679 |
+
self.tlm_manager.specialists[specialist_id].model.load_state_dict(state_dict)
|
680 |
+
|
681 |
+
print(f"✅ Models loaded from {checkpoint_path}")
|
682 |
+
|
683 |
+
except Exception as e:
|
684 |
+
print(f"❌ Error loading models: {e}")
|
685 |
+
|
686 |
+
def set_eval_mode(self):
|
687 |
+
"""Set all models to evaluation mode"""
|
688 |
+
if self.tlm_manager:
|
689 |
+
for specialist in self.tlm_manager.specialists.values():
|
690 |
+
if hasattr(specialist, 'model'):
|
691 |
+
specialist.model.eval()
|
692 |
+
|
693 |
+
if self.aggregator and hasattr(self.aggregator, 'eval'):
|
694 |
+
self.aggregator.eval()
|
695 |
+
|
696 |
+
if self.router and hasattr(self.router, 'eval'):
|
697 |
+
self.router.eval()
|
698 |
+
|
699 |
+
# Set pretrained encoders to eval mode
|
700 |
+
for encoder in self.encoders:
|
701 |
+
if encoder.get("model") and hasattr(encoder["model"], 'eval'):
|
702 |
+
encoder["model"].eval()
|
703 |
+
|
704 |
+
def set_train_mode(self):
|
705 |
+
"""Set all models to training mode"""
|
706 |
+
if self.tlm_manager:
|
707 |
+
for specialist in self.tlm_manager.specialists.values():
|
708 |
+
if hasattr(specialist, 'model'):
|
709 |
+
specialist.model.train()
|
710 |
+
|
711 |
+
if self.aggregator and hasattr(self.aggregator, 'train'):
|
712 |
+
self.aggregator.train()
|
713 |
+
|
714 |
+
if self.router and hasattr(self.router, 'train'):
|
715 |
+
self.router.train()
|
716 |
+
|
717 |
+
|
718 |
+
# =============================================================================
|
719 |
+
# FACTORY FUNCTIONS
|
720 |
+
# =============================================================================
|
721 |
+
|
722 |
+
def create_mamba_swarm(tier: str = "auto", use_pretrained: bool = True,
|
723 |
+
config_override: Optional[Dict] = None) -> UnifiedMambaSwarm:
|
724 |
+
"""
|
725 |
+
Factory function to create appropriately configured swarm
|
726 |
+
|
727 |
+
Args:
|
728 |
+
tier: Scaling tier or "auto" for auto-detection
|
729 |
+
use_pretrained: Whether to use pretrained HuggingFace models
|
730 |
+
config_override: Dictionary to override default config
|
731 |
+
|
732 |
+
Returns:
|
733 |
+
Configured UnifiedMambaSwarm instance
|
734 |
+
"""
|
735 |
+
if tier == "auto":
|
736 |
+
tier = auto_detect_tier()
|
737 |
+
|
738 |
+
return UnifiedMambaSwarm(
|
739 |
+
tier=tier,
|
740 |
+
use_pretrained=use_pretrained,
|
741 |
+
config_override=config_override
|
742 |
+
)
|
743 |
+
|
744 |
+
|
745 |
+
def create_production_swarm(tier: str = "medium") -> UnifiedMambaSwarm:
|
746 |
+
"""Create production-ready swarm with optimal settings"""
|
747 |
+
return UnifiedMambaSwarm(
|
748 |
+
tier=tier,
|
749 |
+
use_pretrained=True,
|
750 |
+
config_override={
|
751 |
+
"batch_size": 32,
|
752 |
+
"max_sequence_length": 2048
|
753 |
+
}
|
754 |
+
)
|
755 |
+
|
756 |
+
|
757 |
+
def create_development_swarm() -> UnifiedMambaSwarm:
|
758 |
+
"""Create development swarm with simulation fallback"""
|
759 |
+
return UnifiedMambaSwarm(
|
760 |
+
tier="demo",
|
761 |
+
use_pretrained=True,
|
762 |
+
config_override={
|
763 |
+
"simulation_fallback": True
|
764 |
+
}
|
765 |
+
)
|
766 |
+
|
767 |
+
|
768 |
+
# =============================================================================
|
769 |
+
# MAIN EXECUTION
|
770 |
+
# =============================================================================
|
771 |
+
|
772 |
+
if __name__ == "__main__":
|
773 |
+
print("🧪 Testing Unified Mamba Swarm...")
|
774 |
+
|
775 |
+
# Create swarm instance
|
776 |
+
swarm = create_mamba_swarm(tier="demo")
|
777 |
+
|
778 |
+
# Display system info
|
779 |
+
print("\n📊 System Information:")
|
780 |
+
info = swarm.get_system_info()
|
781 |
+
for key, value in info.items():
|
782 |
+
print(f" {key}: {value}")
|
783 |
+
|
784 |
+
# Test both processing methods
|
785 |
+
test_prompts = [
|
786 |
+
"Write a Python function to calculate fibonacci numbers",
|
787 |
+
"Explain the process of photosynthesis",
|
788 |
+
"What are the symptoms of diabetes?"
|
789 |
+
]
|
790 |
+
|
791 |
+
print("\n🧪 Testing generate method:")
|
792 |
+
for prompt in test_prompts[:2]:
|
793 |
+
result = swarm.generate(prompt, max_length=150)
|
794 |
+
print(f"\nPrompt: {prompt}")
|
795 |
+
print(f"Response: {result['response'][:100]}...")
|
796 |
+
print(f"Processing time: {result['processing_time']:.3f}s")
|
797 |
+
print(f"Routing: {result['routing_info']['domains']}")
|
798 |
+
|
799 |
+
print("\n🧪 Testing process_request method:")
|
800 |
+
result = swarm.process_request(test_prompts[2])
|
801 |
+
print(f"Response: {result['response'][:100]}...")
|
802 |
+
print(f"Success: {result['success']}")
|
803 |
+
|
804 |
+
# Test batch processing
|
805 |
+
print("\n🧪 Testing batch processing:")
|
806 |
+
batch_results = swarm.batch_process(test_prompts, method="generate")
|
807 |
+
print(f"Processed {len(batch_results)} requests in batch")
|
808 |
+
|
809 |
+
# Display final stats
|
810 |
+
print("\n📈 Final Statistics:")
|
811 |
+
stats = swarm.get_stats()
|
812 |
+
for key, value in stats.items():
|
813 |
+
if key != 'specialist_usage':
|
814 |
+
print(f" {key}: {value}")
|
815 |
+
|
816 |
+
print("\n✅ Testing complete!")
|
system/memory_manager.py
ADDED
@@ -0,0 +1,306 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Memory Manager for Mamba Swarm
|
3 |
+
Handles memory optimization, caching, and distributed memory management
|
4 |
+
"""
|
5 |
+
|
6 |
+
import torch
|
7 |
+
import torch.nn as nn
|
8 |
+
import gc
|
9 |
+
import psutil
|
10 |
+
import threading
|
11 |
+
from typing import Dict, Any, Optional, List, Tuple
|
12 |
+
from dataclasses import dataclass
|
13 |
+
from collections import OrderedDict
|
14 |
+
import numpy as np
|
15 |
+
import logging
|
16 |
+
|
17 |
+
@dataclass
|
18 |
+
class MemoryStats:
|
19 |
+
total_memory: float
|
20 |
+
used_memory: float
|
21 |
+
free_memory: float
|
22 |
+
gpu_memory: float
|
23 |
+
gpu_free: float
|
24 |
+
cache_size: float
|
25 |
+
|
26 |
+
class LRUCache:
|
27 |
+
"""Least Recently Used cache for model states and activations"""
|
28 |
+
|
29 |
+
def __init__(self, max_size: int = 1000):
|
30 |
+
self.max_size = max_size
|
31 |
+
self.cache = OrderedDict()
|
32 |
+
self.lock = threading.Lock()
|
33 |
+
|
34 |
+
def get(self, key: str) -> Optional[torch.Tensor]:
|
35 |
+
with self.lock:
|
36 |
+
if key in self.cache:
|
37 |
+
# Move to end (most recently used)
|
38 |
+
value = self.cache.pop(key)
|
39 |
+
self.cache[key] = value
|
40 |
+
return value
|
41 |
+
return None
|
42 |
+
|
43 |
+
def put(self, key: str, value: torch.Tensor):
|
44 |
+
with self.lock:
|
45 |
+
if key in self.cache:
|
46 |
+
self.cache.pop(key)
|
47 |
+
elif len(self.cache) >= self.max_size:
|
48 |
+
# Remove least recently used
|
49 |
+
oldest_key = next(iter(self.cache))
|
50 |
+
old_value = self.cache.pop(oldest_key)
|
51 |
+
del old_value
|
52 |
+
|
53 |
+
self.cache[key] = value.clone() if isinstance(value, torch.Tensor) else value
|
54 |
+
|
55 |
+
def clear(self):
|
56 |
+
with self.lock:
|
57 |
+
self.cache.clear()
|
58 |
+
gc.collect()
|
59 |
+
|
60 |
+
class GradientAccumulator:
|
61 |
+
"""Manages gradient accumulation across multiple steps"""
|
62 |
+
|
63 |
+
def __init__(self, accumulation_steps: int = 8):
|
64 |
+
self.accumulation_steps = accumulation_steps
|
65 |
+
self.current_step = 0
|
66 |
+
self.accumulated_gradients = {}
|
67 |
+
|
68 |
+
def accumulate(self, model: nn.Module):
|
69 |
+
"""Accumulate gradients from current backward pass"""
|
70 |
+
for name, param in model.named_parameters():
|
71 |
+
if param.grad is not None:
|
72 |
+
if name not in self.accumulated_gradients:
|
73 |
+
self.accumulated_gradients[name] = param.grad.clone()
|
74 |
+
else:
|
75 |
+
self.accumulated_gradients[name] += param.grad
|
76 |
+
|
77 |
+
self.current_step += 1
|
78 |
+
|
79 |
+
def should_update(self) -> bool:
|
80 |
+
"""Check if we should perform optimizer step"""
|
81 |
+
return self.current_step % self.accumulation_steps == 0
|
82 |
+
|
83 |
+
def get_averaged_gradients(self) -> Dict[str, torch.Tensor]:
|
84 |
+
"""Get accumulated gradients averaged over accumulation steps"""
|
85 |
+
averaged = {}
|
86 |
+
for name, grad in self.accumulated_gradients.items():
|
87 |
+
averaged[name] = grad / self.accumulation_steps
|
88 |
+
return averaged
|
89 |
+
|
90 |
+
def reset(self):
|
91 |
+
"""Reset accumulator"""
|
92 |
+
self.accumulated_gradients.clear()
|
93 |
+
self.current_step = 0
|
94 |
+
|
95 |
+
class MemoryManager:
|
96 |
+
"""Comprehensive memory management for Mamba Swarm"""
|
97 |
+
|
98 |
+
def __init__(self,
|
99 |
+
max_cache_size: int = 2000,
|
100 |
+
gradient_accumulation_steps: int = 8,
|
101 |
+
auto_cleanup: bool = True,
|
102 |
+
memory_threshold: float = 0.85):
|
103 |
+
|
104 |
+
self.logger = logging.getLogger(__name__)
|
105 |
+
self.max_cache_size = max_cache_size
|
106 |
+
self.gradient_accumulation_steps = gradient_accumulation_steps
|
107 |
+
self.auto_cleanup = auto_cleanup
|
108 |
+
self.memory_threshold = memory_threshold
|
109 |
+
|
110 |
+
# Initialize components
|
111 |
+
self.activation_cache = LRUCache(max_cache_size)
|
112 |
+
self.state_cache = LRUCache(max_cache_size // 2)
|
113 |
+
self.gradient_accumulator = GradientAccumulator(gradient_accumulation_steps)
|
114 |
+
|
115 |
+
# Memory tracking
|
116 |
+
self.peak_memory_usage = 0.0
|
117 |
+
self.memory_history = []
|
118 |
+
self.cleanup_threshold = memory_threshold
|
119 |
+
|
120 |
+
# Device management
|
121 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
122 |
+
self.setup_memory_optimization()
|
123 |
+
|
124 |
+
def setup_memory_optimization(self):
|
125 |
+
"""Setup memory optimization settings"""
|
126 |
+
if torch.cuda.is_available():
|
127 |
+
# Enable memory mapping for large tensors
|
128 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
129 |
+
torch.backends.cudnn.allow_tf32 = True
|
130 |
+
|
131 |
+
# Set memory fraction
|
132 |
+
if hasattr(torch.cuda, 'set_per_process_memory_fraction'):
|
133 |
+
torch.cuda.set_per_process_memory_fraction(0.9)
|
134 |
+
|
135 |
+
def get_memory_stats(self) -> MemoryStats:
|
136 |
+
"""Get current memory statistics"""
|
137 |
+
# System memory
|
138 |
+
memory = psutil.virtual_memory()
|
139 |
+
total_memory = memory.total / (1024**3) # GB
|
140 |
+
used_memory = memory.used / (1024**3)
|
141 |
+
free_memory = memory.available / (1024**3)
|
142 |
+
|
143 |
+
# GPU memory
|
144 |
+
gpu_memory = 0.0
|
145 |
+
gpu_free = 0.0
|
146 |
+
if torch.cuda.is_available():
|
147 |
+
gpu_memory = torch.cuda.memory_allocated() / (1024**3)
|
148 |
+
gpu_free = (torch.cuda.memory_reserved() - torch.cuda.memory_allocated()) / (1024**3)
|
149 |
+
|
150 |
+
# Cache size estimation
|
151 |
+
cache_size = (len(self.activation_cache.cache) + len(self.state_cache.cache)) * 0.001 # Rough estimate
|
152 |
+
|
153 |
+
stats = MemoryStats(
|
154 |
+
total_memory=total_memory,
|
155 |
+
used_memory=used_memory,
|
156 |
+
free_memory=free_memory,
|
157 |
+
gpu_memory=gpu_memory,
|
158 |
+
gpu_free=gpu_free,
|
159 |
+
cache_size=cache_size
|
160 |
+
)
|
161 |
+
|
162 |
+
# Update peak usage
|
163 |
+
current_usage = used_memory + gpu_memory
|
164 |
+
if current_usage > self.peak_memory_usage:
|
165 |
+
self.peak_memory_usage = current_usage
|
166 |
+
|
167 |
+
return stats
|
168 |
+
|
169 |
+
def check_memory_pressure(self) -> bool:
|
170 |
+
"""Check if system is under memory pressure"""
|
171 |
+
stats = self.get_memory_stats()
|
172 |
+
memory_usage_ratio = stats.used_memory / stats.total_memory
|
173 |
+
|
174 |
+
if torch.cuda.is_available():
|
175 |
+
gpu_usage_ratio = stats.gpu_memory / (stats.gpu_memory + stats.gpu_free + 1e-6)
|
176 |
+
return memory_usage_ratio > self.cleanup_threshold or gpu_usage_ratio > self.cleanup_threshold
|
177 |
+
|
178 |
+
return memory_usage_ratio > self.cleanup_threshold
|
179 |
+
|
180 |
+
def cleanup_memory(self, aggressive: bool = False):
|
181 |
+
"""Perform memory cleanup"""
|
182 |
+
if aggressive:
|
183 |
+
self.activation_cache.clear()
|
184 |
+
self.state_cache.clear()
|
185 |
+
self.gradient_accumulator.reset()
|
186 |
+
|
187 |
+
# Python garbage collection
|
188 |
+
gc.collect()
|
189 |
+
|
190 |
+
# GPU memory cleanup
|
191 |
+
if torch.cuda.is_available():
|
192 |
+
torch.cuda.empty_cache()
|
193 |
+
torch.cuda.synchronize()
|
194 |
+
|
195 |
+
self.logger.info(f"Memory cleanup completed. Aggressive: {aggressive}")
|
196 |
+
|
197 |
+
def cache_activation(self, key: str, activation: torch.Tensor):
|
198 |
+
"""Cache activation with memory pressure check"""
|
199 |
+
if self.auto_cleanup and self.check_memory_pressure():
|
200 |
+
self.cleanup_memory()
|
201 |
+
|
202 |
+
self.activation_cache.put(key, activation)
|
203 |
+
|
204 |
+
def get_cached_activation(self, key: str) -> Optional[torch.Tensor]:
|
205 |
+
"""Retrieve cached activation"""
|
206 |
+
return self.activation_cache.get(key)
|
207 |
+
|
208 |
+
def cache_hidden_state(self, key: str, state: torch.Tensor):
|
209 |
+
"""Cache hidden state"""
|
210 |
+
self.state_cache.put(key, state)
|
211 |
+
|
212 |
+
def get_cached_state(self, key: str) -> Optional[torch.Tensor]:
|
213 |
+
"""Retrieve cached hidden state"""
|
214 |
+
return self.state_cache.get(key)
|
215 |
+
|
216 |
+
def manage_gradient_accumulation(self, model: nn.Module) -> bool:
|
217 |
+
"""Manage gradient accumulation and return if optimizer step should be taken"""
|
218 |
+
self.gradient_accumulator.accumulate(model)
|
219 |
+
|
220 |
+
if self.gradient_accumulator.should_update():
|
221 |
+
# Apply accumulated gradients
|
222 |
+
averaged_grads = self.gradient_accumulator.get_averaged_gradients()
|
223 |
+
|
224 |
+
for name, param in model.named_parameters():
|
225 |
+
if name in averaged_grads:
|
226 |
+
param.grad = averaged_grads[name]
|
227 |
+
|
228 |
+
self.gradient_accumulator.reset()
|
229 |
+
return True
|
230 |
+
|
231 |
+
return False
|
232 |
+
|
233 |
+
def optimize_model_memory(self, model: nn.Module):
|
234 |
+
"""Optimize model memory usage"""
|
235 |
+
# Enable gradient checkpointing for large models
|
236 |
+
for module in model.modules():
|
237 |
+
if hasattr(module, 'gradient_checkpointing'):
|
238 |
+
module.gradient_checkpointing = True
|
239 |
+
|
240 |
+
# Convert to half precision if possible
|
241 |
+
if torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 7:
|
242 |
+
model = model.half()
|
243 |
+
|
244 |
+
return model
|
245 |
+
|
246 |
+
def create_memory_efficient_dataloader(self, dataset, batch_size: int, **kwargs):
|
247 |
+
"""Create memory-efficient dataloader"""
|
248 |
+
# Adjust batch size based on available memory
|
249 |
+
stats = self.get_memory_stats()
|
250 |
+
|
251 |
+
if stats.free_memory < 2.0: # Less than 2GB free
|
252 |
+
batch_size = max(1, batch_size // 2)
|
253 |
+
self.logger.warning(f"Reduced batch size to {batch_size} due to low memory")
|
254 |
+
|
255 |
+
return torch.utils.data.DataLoader(
|
256 |
+
dataset,
|
257 |
+
batch_size=batch_size,
|
258 |
+
num_workers=min(4, psutil.cpu_count()),
|
259 |
+
pin_memory=torch.cuda.is_available(),
|
260 |
+
prefetch_factor=2,
|
261 |
+
**kwargs
|
262 |
+
)
|
263 |
+
|
264 |
+
def monitor_memory_usage(self):
|
265 |
+
"""Monitor and log memory usage"""
|
266 |
+
stats = self.get_memory_stats()
|
267 |
+
self.memory_history.append({
|
268 |
+
'timestamp': torch.cuda.Event(enable_timing=True) if torch.cuda.is_available() else None,
|
269 |
+
'stats': stats
|
270 |
+
})
|
271 |
+
|
272 |
+
# Keep only recent history
|
273 |
+
if len(self.memory_history) > 100:
|
274 |
+
self.memory_history = self.memory_history[-50:]
|
275 |
+
|
276 |
+
self.logger.debug(f"Memory - System: {stats.used_memory:.2f}GB/{stats.total_memory:.2f}GB, "
|
277 |
+
f"GPU: {stats.gpu_memory:.2f}GB, Cache: {stats.cache_size:.2f}GB")
|
278 |
+
|
279 |
+
def get_memory_report(self) -> Dict[str, Any]:
|
280 |
+
"""Generate comprehensive memory report"""
|
281 |
+
stats = self.get_memory_stats()
|
282 |
+
|
283 |
+
return {
|
284 |
+
'current_stats': stats.__dict__,
|
285 |
+
'peak_usage': self.peak_memory_usage,
|
286 |
+
'cache_stats': {
|
287 |
+
'activation_cache_size': len(self.activation_cache.cache),
|
288 |
+
'state_cache_size': len(self.state_cache.cache),
|
289 |
+
'max_cache_size': self.max_cache_size
|
290 |
+
},
|
291 |
+
'gradient_accumulation': {
|
292 |
+
'current_step': self.gradient_accumulator.current_step,
|
293 |
+
'accumulation_steps': self.gradient_accumulation_steps,
|
294 |
+
'accumulated_params': len(self.gradient_accumulator.accumulated_gradients)
|
295 |
+
},
|
296 |
+
'memory_pressure': self.check_memory_pressure(),
|
297 |
+
'device': str(self.device)
|
298 |
+
}
|
299 |
+
|
300 |
+
def __enter__(self):
|
301 |
+
"""Context manager entry"""
|
302 |
+
return self
|
303 |
+
|
304 |
+
def __exit__(self, exc_type, exc_val, exc_tb):
|
305 |
+
"""Context manager exit with cleanup"""
|
306 |
+
self.cleanup_memory(aggressive=True)
|
system/weight_manager.py
ADDED
@@ -0,0 +1,168 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# =============================================================================
|
2 |
+
# system/weight_manager.py
|
3 |
+
# =============================================================================
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
from typing import Dict, List, Optional
|
7 |
+
import os
|
8 |
+
from pathlib import Path
|
9 |
+
|
10 |
+
class WeightManager:
|
11 |
+
"""Manages hierarchical weight sharing and loading/saving"""
|
12 |
+
|
13 |
+
def __init__(self, config, tlm_manager):
|
14 |
+
self.config = config
|
15 |
+
self.tlm_manager = tlm_manager
|
16 |
+
|
17 |
+
# Track shared weights
|
18 |
+
self.shared_embeddings = None
|
19 |
+
self.shared_foundation_layers = {}
|
20 |
+
|
21 |
+
def setup_hierarchical_sharing(self):
|
22 |
+
"""Setup hierarchical weight sharing between specialists"""
|
23 |
+
print("Setting up hierarchical weight sharing...")
|
24 |
+
|
25 |
+
# Create shared embedding if enabled
|
26 |
+
if self.config.shared_embedding:
|
27 |
+
self.shared_embeddings = nn.Embedding(
|
28 |
+
self.config.vocab_size,
|
29 |
+
self.config.d_model
|
30 |
+
).to(self.config.device)
|
31 |
+
|
32 |
+
# Share embedding across all specialists
|
33 |
+
for specialist in self.tlm_manager.specialists.values():
|
34 |
+
specialist.model.embedding.token_embedding = self.shared_embeddings
|
35 |
+
|
36 |
+
# Setup foundation layer sharing
|
37 |
+
self._setup_foundation_sharing()
|
38 |
+
|
39 |
+
print("Hierarchical weight sharing setup complete!")
|
40 |
+
|
41 |
+
def _setup_foundation_sharing(self):
|
42 |
+
"""Setup sharing of foundation layers"""
|
43 |
+
num_shared_layers = self.config.n_layers // 2
|
44 |
+
|
45 |
+
# Group specialists by domain similarity
|
46 |
+
domain_groups = self._group_specialists_by_domain()
|
47 |
+
|
48 |
+
for group_name, specialist_ids in domain_groups.items():
|
49 |
+
if len(specialist_ids) > 1:
|
50 |
+
# Create shared foundation layers for this group
|
51 |
+
reference_specialist = self.tlm_manager.specialists[specialist_ids[0]]
|
52 |
+
shared_layers = reference_specialist.model.layers[:num_shared_layers]
|
53 |
+
|
54 |
+
# Share with other specialists in the group
|
55 |
+
for specialist_id in specialist_ids[1:]:
|
56 |
+
specialist = self.tlm_manager.specialists[specialist_id]
|
57 |
+
for i in range(num_shared_layers):
|
58 |
+
specialist.model.layers[i] = shared_layers[i]
|
59 |
+
|
60 |
+
self.shared_foundation_layers[group_name] = shared_layers
|
61 |
+
|
62 |
+
def _group_specialists_by_domain(self) -> Dict[str, List[int]]:
|
63 |
+
"""Group specialists by domain for weight sharing"""
|
64 |
+
domain_groups = {
|
65 |
+
'stem': [],
|
66 |
+
'programming': [],
|
67 |
+
'language': [],
|
68 |
+
'business': [],
|
69 |
+
'general': []
|
70 |
+
}
|
71 |
+
|
72 |
+
for specialist_id, specialist in self.tlm_manager.specialists.items():
|
73 |
+
domain_name = specialist.domain_info['name'].lower()
|
74 |
+
|
75 |
+
if any(x in domain_name for x in ['math', 'physics', 'chemistry', 'biology']):
|
76 |
+
domain_groups['stem'].append(specialist_id)
|
77 |
+
elif any(x in domain_name for x in ['python', 'javascript', 'systems']):
|
78 |
+
domain_groups['programming'].append(specialist_id)
|
79 |
+
elif any(x in domain_name for x in ['writing', 'translation']):
|
80 |
+
domain_groups['language'].append(specialist_id)
|
81 |
+
elif any(x in domain_name for x in ['business', 'legal']):
|
82 |
+
domain_groups['business'].append(specialist_id)
|
83 |
+
else:
|
84 |
+
domain_groups['general'].append(specialist_id)
|
85 |
+
|
86 |
+
return {k: v for k, v in domain_groups.items() if len(v) > 1}
|
87 |
+
|
88 |
+
def save_weights(self, save_path: str):
|
89 |
+
"""Save all weights with hierarchical structure"""
|
90 |
+
save_path = Path(save_path)
|
91 |
+
save_path.mkdir(parents=True, exist_ok=True)
|
92 |
+
|
93 |
+
# Save shared embeddings
|
94 |
+
if self.shared_embeddings is not None:
|
95 |
+
torch.save(
|
96 |
+
self.shared_embeddings.state_dict(),
|
97 |
+
save_path / "shared_embeddings.pt"
|
98 |
+
)
|
99 |
+
|
100 |
+
# Save shared foundation layers
|
101 |
+
for group_name, layers in self.shared_foundation_layers.items():
|
102 |
+
group_state = {}
|
103 |
+
for i, layer in enumerate(layers):
|
104 |
+
group_state[f"layer_{i}"] = layer.state_dict()
|
105 |
+
torch.save(group_state, save_path / f"shared_foundation_{group_name}.pt")
|
106 |
+
|
107 |
+
# Save specialist-specific weights
|
108 |
+
specialists_path = save_path / "specialists"
|
109 |
+
specialists_path.mkdir(exist_ok=True)
|
110 |
+
|
111 |
+
for specialist_id, specialist in self.tlm_manager.specialists.items():
|
112 |
+
torch.save(
|
113 |
+
specialist.model.state_dict(),
|
114 |
+
specialists_path / f"specialist_{specialist_id}.pt"
|
115 |
+
)
|
116 |
+
|
117 |
+
print(f"Weights saved to {save_path}")
|
118 |
+
|
119 |
+
def load_weights(self, load_path: str):
|
120 |
+
"""Load weights with hierarchical structure"""
|
121 |
+
load_path = Path(load_path)
|
122 |
+
|
123 |
+
if not load_path.exists():
|
124 |
+
raise FileNotFoundError(f"Weight path {load_path} not found")
|
125 |
+
|
126 |
+
# Load shared embeddings
|
127 |
+
embeddings_path = load_path / "shared_embeddings.pt"
|
128 |
+
if embeddings_path.exists() and self.shared_embeddings is not None:
|
129 |
+
self.shared_embeddings.load_state_dict(torch.load(embeddings_path))
|
130 |
+
|
131 |
+
# Load shared foundation layers
|
132 |
+
for group_name in self.shared_foundation_layers.keys():
|
133 |
+
foundation_path = load_path / f"shared_foundation_{group_name}.pt"
|
134 |
+
if foundation_path.exists():
|
135 |
+
group_state = torch.load(foundation_path)
|
136 |
+
for i, layer in enumerate(self.shared_foundation_layers[group_name]):
|
137 |
+
if f"layer_{i}" in group_state:
|
138 |
+
layer.load_state_dict(group_state[f"layer_{i}"])
|
139 |
+
|
140 |
+
# Load specialist weights
|
141 |
+
specialists_path = load_path / "specialists"
|
142 |
+
if specialists_path.exists():
|
143 |
+
for specialist_id, specialist in self.tlm_manager.specialists.items():
|
144 |
+
specialist_path = specialists_path / f"specialist_{specialist_id}.pt"
|
145 |
+
if specialist_path.exists():
|
146 |
+
specialist.model.load_state_dict(torch.load(specialist_path))
|
147 |
+
|
148 |
+
print(f"Weights loaded from {load_path}")
|
149 |
+
|
150 |
+
def get_memory_usage(self) -> Dict[str, int]:
|
151 |
+
"""Get memory usage breakdown"""
|
152 |
+
usage = {}
|
153 |
+
|
154 |
+
# Shared embedding memory
|
155 |
+
if self.shared_embeddings is not None:
|
156 |
+
usage['shared_embeddings'] = sum(
|
157 |
+
p.numel() * p.element_size()
|
158 |
+
for p in self.shared_embeddings.parameters()
|
159 |
+
)
|
160 |
+
|
161 |
+
# Shared foundation layer memory
|
162 |
+
total_foundation = 0
|
163 |
+
for layers in self.shared_foundation_layers.values():
|
164 |
+
for layer in layers:
|
165 |
+
total_foundation += sum(
|
166 |
+
p.numel() * p.element_size()
|
167 |
+
for p in layer.parameters()
|
168 |
+
)
|