Cristian Cezar Moisés
ebbbf84d35
Update generate.py
...
Distributed Training Enhancements:
Proper NCCL/Gloo backend selection
Distributed timeout handling
Rank-aware input broadcasting
Graceful process group cleanup
Error Handling & Validation
Comprehensive path validation
Config schema validation
Tokenization error handling
Batch processing safeguards
CUDA OOM fallback handling
Generation Improvements:
Top-k sampling support
Repetition penalty
Dynamic sequence length management
Progress tracking with tqdm
Sequence truncation warnings
Performance Optimizations:
Device-aware tensor placement
Batch tokenization
Memory-efficient generation loop
Model parallelism support
User Experience:
Interactive mode enhancements:
Command history
Input validation
Graceful exit handling
Batch processing:
Progress tracking
Error resilience
Clean output formatting
Code Quality:
Type hints throughout
Configurable constants
Modular architecture
Docstrings with examples
Logging integration
Safety Features:
Tokenizer trust_remote_code handling
Config validation
Input sanitization
Resource cleanup guarantees
2025-01-27 23:16:21 -03:00
Cristian Cezar Moisés
eee820cc36
Update fp8_cast_bf16.py
...
Type Hints & Path Management:
Added comprehensive type annotations
Used pathlib.Path for safer path handling
Enhanced Error Handling:
Structured exception handling throughout
Clear error messages with context
Safe resource cleanup
Memory Management:
LRU cache implementation with OrderedDict
Configurable cache size
Explicit GPU memory cleanup
Logging System:
Configurable logging levels
Detailed progress tracking
Structured error reporting
Code Organization:
Split into focused, testable functions
Clear separation of concerns
Documented public methods
Validation & Safety:
Input path validation
Weight type checking
Clone tensors to prevent reference issues
Performance:
Optimized file loading with LRU cache
Batched tensor processing
Asynchronous CUDA operations
Metadata & Traceability:
Added conversion metadata to output files
Preserved original index structure
Enhanced output index information
Configuration:
Centralized constants
Device-aware execution (CUDA/CPU)
Progress Tracking:
Nested progress bars
Detailed file processing status
2025-01-27 23:13:11 -03:00
Cristian Cezar Moisés
a26fca4a41
Update convert.py
2025-01-27 23:10:08 -03:00
enoch kan
bc77f22afc
Updated model.py docstrings
2025-01-05 18:24:31 +00:00
enoch kan
a1296f099e
Enhance documentation and update .gitignore for model conversion scripts
2025-01-05 18:18:18 +00:00
GeeeekExplorer
fd011c11aa
torch rmsnorm
2025-01-05 14:33:48 +08:00
Xingkai Yu
8710ec2ecb
require model-parallel in convert.py
2024-12-31 18:05:55 +08:00
Yang Wang
8f1c9488b5
handle missing scale_inv_name ( #2 )
...
* handle missing scale_inv_name
Fixed an issue where `weight` and `weight_scale_inv` (e.g. `model.layers.39.mlp.experts.92.gate_proj.weight` and `model.layers.39.mlp.experts.92.gate_proj.weight_scale_inv`) were not in the same SafeTensor, causing an assertion error due to scale_inv_name not being in the state_dict.
* sort filename to reduce memory costs
* Add CUDA cache clearing in memory management
Added torch.cuda.empty_cache() to free up unused memory on the GPU,
2024-12-27 09:34:38 +08:00
stack-heap-overflow
4c2fdb8f55
Release DeepSeek-V3
2024-12-26 19:01:57 +08:00
