Open
Conversation
…artifacts untracked
…sues - Updated alignment.sh: Added conda activation, set CUDA paths to 12.1, configured for GPU 7 only - Modified stage_1_alignment_llava_ov_4b.sh: Changed TP from 2 to 1, updated checkpoint path to tp1_pp1 - Fixed Makefile: Added check to skip recompilation if helpers_cpp .so files already exist - Updated utils.py: Added pre-compilation check to avoid unnecessary builds
- Added print statements for debugging in dataloader_provider.py - Added print statements in qwen2vl_task_encoder.py for tracing preprocessing - Added debugging prints in llavaov_1_5_provider.py - Added print statements in train.py, megatron_trainer.py, and other training files - These changes were made during codebase exploration and understanding
- Add FastViT model implementation (mobileclip_l_384) in aiak_training_llm/models/fastvit/ - Update LlavaOnevision1_5 model to use FastViT encoder - Add FastViT preprocessing in qwen2vl_task_encoder.py - Add --use-fastvit and related command-line arguments - Add checkpoint conversion scripts for FastVLM - Update training configs for 2-GPU setup (TP=2) - Add .gitignore entries for checkpoints and training outputs
There was a problem hiding this comment.
Pull request overview
This PR integrates FastVit functionality into the Apex library by adding extensive CUDA/cuDNN-accelerated operations for convolutional neural networks, batch normalization, gradient clipping, and RNN implementations. The changes introduce new modules for fused operations, optimized convolution paths, and automatic mixed precision training support.
- Adds cuDNN-accelerated batch normalization and convolution-bias-relu fusion operations
- Implements gradient clipping utilities with fused CUDA kernels
- Introduces bottleneck layer implementations with spatial parallelism support
- Adds comprehensive AMP (Automatic Mixed Precision) framework with optimizer integration
Reviewed changes
Copilot reviewed 99 out of 499 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| apex/apex/contrib/csrc/cudnn_gbn/cudnn_gbn.cpp | Implements group batch normalization forward/backward with NCCL peer communication |
| apex/apex/contrib/csrc/conv_bias_relu/conv_bias_relu.cpp | Provides fused convolution-bias-relu operations using cuDNN frontend API |
| apex/apex/contrib/conv_bias_relu/conv_bias_relu.py | Python wrapper for fused conv-bias-relu autograd functions |
| apex/apex/contrib/conv_bias_relu/init.py | Exports conv-bias-relu function variants |
| apex/apex/contrib/clip_grad/clip_grad.py | Implements optimized gradient norm clipping with multi-tensor operations |
| apex/apex/contrib/clip_grad/init.py | Exports gradient clipping function |
| apex/apex/contrib/bottleneck/test.py | Test suite for bottleneck layer implementation |
| apex/apex/contrib/bottleneck/halo_exchangers.py | Implements halo exchange patterns for spatial parallelism |
| apex/apex/contrib/bottleneck/bottleneck.py | Bottleneck and SpatialBottleneck layer implementations with frozen batch norm |
| apex/apex/contrib/bottleneck/init.py | Exports bottleneck classes and halo exchangers |
| apex/apex/amp/wrap.py | Provides function wrapping utilities for automatic casting |
| apex/apex/amp/utils.py | Utility functions for tensor type checking and casting |
| apex/apex/amp/scaler.py | Loss scaling implementation for mixed precision training |
| apex/apex/amp/rnn_compat.py | RNN compatibility layer for different PyTorch versions |
| apex/apex/amp/opt.py | Optimizer wrapper for AMP integration |
| apex/apex/amp/lists/torch_overrides.py | Function lists for torch module automatic casting |
| apex/apex/amp/lists/tensor_overrides.py | Function lists for tensor method automatic casting |
| apex/apex/amp/lists/functional_overrides.py | Function lists for torch.nn.functional automatic casting |
| apex/apex/amp/handle.py | AMP handle implementation for managing mixed precision state |
| apex/apex/amp/frontend.py | User-facing API for AMP initialization and configuration |
| apex/apex/amp/compat.py | PyTorch version compatibility utilities |
| apex/apex/amp/amp.py | Core AMP implementation with function patching |
| apex/apex/amp/_process_optimizer.py | Optimizer processing for master weights and gradient scaling |
| apex/apex/amp/_initialize.py | Model and optimizer initialization for AMP |
| apex/apex/amp/_amp_state.py | Global state management for AMP |
| apex/apex/amp/version.py | Version information |
| apex/apex/amp/init.py | AMP module exports |
| apex/apex/amp/README.md | Documentation for AMP user annotations |
| apex/apex/_autocast_utils.py | Utilities for PyTorch autocast integration |
| apex/apex/init.py | Main package initialization with logging and deprecation warnings |
| apex/apex/RNN/models.py | RNN model factory functions (LSTM, GRU, etc.) |
| apex/apex/RNN/cells.py | Custom RNN cell implementations including mLSTM |
| apex/apex/RNN/init.py | RNN module exports |
| apex/apex/RNN/RNNBackend.py | Backend implementation for bidirectional and stacked RNNs |
| apex/apex/RNN/README.md | Deprecation notice for RNN module |
| apex/README.md | Comprehensive documentation for Apex library features |
| apex/LICENSE | BSD 3-Clause license text |
| apex/.gitmodules | Git submodule configuration for cutlass and cudnn-frontend |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| DEBUG_CUDNN_MSG(log_buf, knobs.begin()->describe()); | ||
| } | ||
|
|
||
| // Createmplacee the requisite engine config |
There was a problem hiding this comment.
Corrected spelling of 'Createmplacee' to 'Create'.
Suggested change
| // Createmplacee the requisite engine config | |
| // Create the requisite engine config |
- Add debug prints in FastViT forward pass (mci.py, mobileclip_encoder.py, fastvit_vision_model.py) - Update GPU configuration to use 2 GPUs (TP=2) in alignment.sh - Add comprehensive code documentation and comments - Update training configuration for FastViT image processing - Add FastViT preprocessing path in qwen2vl_task_encoder.py
- Update megatron_core SOURCES.txt - Update stage 1 alignment script configuration
…izer/processor loading
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.