-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy path.env.example
More file actions
83 lines (76 loc) · 6.19 KB
/
Copy path.env.example
File metadata and controls
83 lines (76 loc) · 6.19 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
# =============================================================================
# API Keys (Required to enable respective provider)
# =============================================================================
ANTHROPIC_API_KEY="your_anthropic_api_key_here" # Required: Format: sk-ant-api03-...
PERPLEXITY_API_KEY="your_perplexity_api_key_here" # Optional: Format: pplx-...
OPENAI_API_KEY="your_openai_api_key_here" # Optional, for OpenAI models. Format: sk-proj-...
GOOGLE_API_KEY="your_google_api_key_here" # Optional, for Google Gemini models.
MISTRAL_API_KEY="your_mistral_key_here" # Optional, for Mistral AI models.
XAI_API_KEY="YOUR_XAI_KEY_HERE" # Optional, for xAI AI models.
GROQ_API_KEY="YOUR_GROQ_KEY_HERE" # Optional, for Groq models.
OPENROUTER_API_KEY="YOUR_OPENROUTER_KEY_HERE" # Optional, for OpenRouter models.
AZURE_OPENAI_API_KEY="your_azure_key_here" # Optional, for Azure OpenAI models (requires endpoint in .taskmaster/config.json).
OLLAMA_API_KEY="your_ollama_api_key_here" # Optional: For remote Ollama servers that require authentication.
GITHUB_API_KEY="your_github_api_key_here" # Optional: For GitHub import/export features. Format: ghp_... or github_pat_...
# =============================================================================
# MLX Configuration (Apple Silicon Optimization)
# =============================================================================
MLX_ENABLE_UNIFIED_MEMORY="1" # Enable unified memory optimization for M1/M2/M3
MLX_DEFAULT_CACHE_SIZE="4096" # Default KV cache size in tokens
MLX_COMPUTE_UNITS="all" # Use all available GPU cores (all/8/4/2)
MLX_FLASH_ATTENTION="1" # Enable flash attention for memory efficiency
MLX_MMAP_MODEL_LOADING="1" # Memory-map models for lower RAM usage
# =============================================================================
# QLoRA Quantization Settings (4-bit Model Compression)
# =============================================================================
QLORA_BITS="4" # Quantization bits (4 or 8)
QLORA_BLOCK_SIZE="64" # Block size for quantization
QLORA.compute_dtype="float16" # Computation dtype (float16/bfloat16/float32)
QLORA_QUANT_TYPE="qmm" # Quantization method (qmm/nf4)
QLORA_USE_DOUBLE_QUANT="1" # Enable double quantization for further compression
QLORA_LORA_RANK="64" # LoRA attention rank
QLORA_LORA_ALPHA="16" # LoRA alpha parameter
QLORA_LORA_DROPOUT="0.05" # LoRA dropout probability
QLORA_TARGET_MODULES="attn,mlp" # Modules to apply LoRA to
# =============================================================================
# TurboQuant KV Cache (3-bit Compression for 6x Memory Reduction)
# =============================================================================
TURBOQUANT_ENABLE="1" # Enable TurboQuant KV cache compression
TURBOQUANT_BITS="3" # KV cache quantization bits (2, 3, or 4)
TURBOQUANT_ENTROPY_CODING="1" # Use entropy coding for compression
TURBOQUANT_SCALE_METHOD="abs_max" # Scale method (abs_max/percentile/entropy)
TURBOQUANT_SPARSITY_THRESHOLD="0.1" # Sparsity threshold for additional compression
# =============================================================================
# Memory Management
# =============================================================================
ENGRAM_ENABLE="1" # Enable Engram O(1) SSD offloading
ENGRAM_SSD_CACHE_DIR="./ssd_cache" # Directory for SSD cache
ENGRAM_MEMORY_BUDGET="6144" # Memory budget in MB (6GB target for 8GB RAM)
ENGRAM_OFFLOAD_THRESHOLD="0.8" # Offload when memory usage exceeds this fraction
# =============================================================================
# Deterministic Seed Policy (Reproducibility)
# =============================================================================
PYTHONHASHSEED="42" # Python hash seed
TORCH_DETERMINISTIC="1" # Enable deterministic operations in PyTorch
TORCH_USE_RD quadruplets="1" # Use deterministic cuBLAS
MLX_DETERMINISTIC="1" # Enable deterministic MLX operations
RANDOM_SEED="42" # Global random seed
NP_RANDOM_SEED="42" # NumPy random seed
TF_DETERMINISTIC="1" # TensorFlow deterministic mode (if used)
# =============================================================================
# Training Configuration
# =============================================================================
MAX_SEQ_LENGTH="2048" # Maximum sequence length
GRADIENT_ACCUMULATION_STEPS="4" # Gradient accumulation for larger batch size
GRADIENT_CHECKPOINTING="1" # Enable gradient checkpointing
TRAIN_BATCH_SIZE="2" # Per-device batch size (constrained by 8GB RAM)
NUM_WORKERS="4" # DataLoader workers
PREFETCH_FACTOR="2" # Prefetch factor for DataLoader
# =============================================================================
# Experiment Tracking
# =============================================================================
WANDB_MODE="disabled" # Weights & Biases mode (disabled/online/offline)
WANDB_PROJECT="autonomous_laughter_prediction" # W&B project name
EXPERIMENT_NAME="" # Set dynamically per experiment run
HF_DATASETS_CACHE="./datasets_cache" # HuggingFace datasets cache directory
MODEL_CACHE="./model_cache" # Model cache directory