Add offline dataset generation (#39)

* add offline dataset gen

Signed-off-by: Varun Khare <varun.khare@nimbleedgehq.ai>

* remove type hints for dynamic classes

Signed-off-by: Varun Khare <varun.khare@nimbleedgehq.ai>

* add evaluation logic for predictors

Signed-off-by: Varun Khare <varun.khare@nimbleedgehq.ai>

* add batch processing in dataset gen

Signed-off-by: Varun Khare <varun.khare@nimbleedgehq.ai>

* random sampling for sequence length

Signed-off-by: Varun Khare <varun.khare@nimbleedgehq.ai>

* add dataset append for memory efficiency

Signed-off-by: Varun Khare <varun.khare@niimbleedgehq.ai>

* add streaming support for training

Signed-off-by: Varun Khare <varun.khare@niimbleedgehq.ai>

* increase defaults for dataset generation

Signed-off-by: Varun Khare <varun.khare@niimbleedgehq.ai>

* rename files and enable full dataset loading

Signed-off-by: Varun Khare <varun.khare@niimbleedgehq.ai>

* add resume checkpointing

Signed-off-by: Varun Khare <varun.khare@niimbleedgehq.ai>

---------

Signed-off-by: Varun Khare <varun.khare@nimbleedgehq.ai>
Signed-off-by: Varun Khare <varun.khare@niimbleedgehq.ai>
Co-authored-by: Varun Khare <varun.khare@niimbleedgehq.ai>

Author: vkkhare

.gitignore

@@ -21,6 +21,7 @@ wheels/
 .installed.cfg
 *.egg
 .cursorrules
+trained_predictors/
 wandb/
 # CUDA
 *.i
@@ -87,6 +88,7 @@ coverage.xml
 # Logs and databases
 *.sqlite
 *.db
+data/
 
 # OS generated files
 .DS_Store

README.md

@@ -70,7 +70,7 @@ Sparse LLaMA 3.2 3B vs Standard LLaMA 3.2 3B CUDA Results (on HuggingFace Implem
 ```bash
 # Run comprehensive benchmark
 
-python run_benchmark.py \
+python benchmark.py \
     --device cpu \                              # Device: 'cpu' or 'cuda'
     --config configs/llama_skip_causal_3b.json \ # Model configuration
     --num_runs 50 \                            # Number of benchmark runs