fix: incorrect reference to quantization file

Fixes issue #4 pilancilab/matrix-compressor.
pilancilab · Nov 21, 2023 · 2d53a4f · 2d53a4f
1 parent d3fa5c1
commit 2d53a4f
Show file tree

Hide file tree

Showing 2 changed files with 9 additions and 24 deletions.
diff --git a/README.md b/README.md
@@ -57,31 +57,15 @@ You can try out the image compression experiments from the paper using `example.
 
 ## 📊 Replicating experiments in paper
 
-### LlaMa Layer Analysis
+### LlaMa Layer wise Analysis
 
 Replace the MODEL_DIRECTORY by the location of the unquantized model, and adjust parameters `b1`, `b2` and `cr`.
 
 ```bash
 export OUTPUT_DIRECTORY="./artifacts/llama-quantized"
 export MODEL_DIRECTORY="./artifacts/llama-7b-hf/"
 export LOGURU_LEVEL=INFO 
-stdbuf -oL python scripts/llama/layer_wise_quantization.py --model-directory $MODEL_DIRECTORY --output-directory $OUTPUT_DIRECTORY --b1 8 --b2 8 --cr 1 --map-location "cuda:1" 2>&1 | stdbuf -oL tee -i $OUTPUT_DIRECTORY/quantization-$(date +%m%d%H%M%S).log
+stdbuf -oL python scripts/llama/per_layer_naive_quantization_comparison/lplr_vanilla.py --model-directory $MODEL_DIRECTORY --output-directory $OUTPUT_DIRECTORY --b1 8 --b2 8 --cr 1 --map-location "cuda:1" 2>&1 | stdbuf -oL tee -i $OUTPUT_DIRECTORY/quantization-$(date +%m%d%H%M%S).log
 ```
 
-### Quantization and Evaluation
-
-```bash
-export RF=0.8
-export B1=16
-export B2=16
-export CUDA_VISIBLE_DEVICES=1
-export LOGURU_LEVEL=TRACE
-export INPUT_DIR=artifacts/llama-7b-hf
-export OUTPUT_DIR=artifacts/llama-quantized-svd-r{$RF}-{$B1}-{$B2}
-
-stdbuf --output=L python scripts/lplr/quantize_torch_splits_svd.py --in-path $INPUT_DIR --out-path $OUTPUT_DIR --map-device 'cuda:0' --rank-fraction $RF --b1 $B1 --b2 $B2 2>&1 | stdbuf --output=L tee $OUTPUT_DIR/quantization.log
-
-cp artifacts/llama-7b-hf/tokenizer.model artifacts/llama-7b-hf/*.json $INPUT_DIR
-
-stdbuf --output=L python repos/lm-evaluation-harness/main.py --model hf-causal --model_args pretrained=/home/rsaha/varun/matrix-compressor/$INPUT_DIR --tasks boolq,hellaswag,piqa 2>&1 | stdbuf --output=L tee $INPUT_DIR/evaluation.log
-```
+Replace the file `lplr_vanilla.py` by the corresponding variant in the folder `scripts/llama/per_layer_naive_quantization_comparison` to use LSVD/DSVD.
diff --git a/scripts/llama/per_layer_naive_quantization_comparison/lplr_vanilla.py b/scripts/llama/per_layer_naive_quantization_comparison/lplr_vanilla.py
@@ -7,10 +7,10 @@
 
 
 # Gaussian
-stdbuf -oL python scripts/llama/layer_wise_lplr_quantization.py --model-directory $MODEL_DIRECTORY --output-directory $OUTPUT_DIRECTORY --b1 8 --b2 8 --b_nq 8 --cr 1 --map-location "cuda:1" 2>&1 | stdbuf -oL tee -i $OUTPUT_DIRECTORY/quantization-$(date +%m%d%H%M%S).log
+stdbuf -oL python scripts/llama/per_layer_naive_quantization_comparison/lplr_vanilla.py --model-directory $MODEL_DIRECTORY --output-directory $OUTPUT_DIRECTORY --b1 8 --b2 8 --b_nq 8 --cr 1 --map-location "cuda:1" 2>&1 | stdbuf -oL tee -i $OUTPUT_DIRECTORY/quantization-$(date +%m%d%H%M%S).log
 
 # Sparse JL
-stdbuf -oL python scripts/llama/layer_wise_lplr_quantization.py --model-directory $MODEL_DIRECTORY --output-directory $OUTPUT_DIRECTORY --b1 8 --b2 8 --b_nq 8 --cr 1 --map-location "cuda:2" 2>&1 --sketch SparseJL --sparse-jl-s 1 | stdbuf -oL tee -i $OUTPUT_DIRECTORY/quantization-$(date +%m%d%H%M%S).log
+stdbuf -oL python scripts/llama/per_layer_naive_quantization_comparison/lplr_vanilla.py --model-directory $MODEL_DIRECTORY --output-directory $OUTPUT_DIRECTORY --b1 8 --b2 8 --b_nq 8 --cr 1 --map-location "cuda:2" 2>&1 --sketch SparseJL --sparse-jl-s 1 | stdbuf -oL tee -i $OUTPUT_DIRECTORY/quantization-$(date +%m%d%H%M%S).log
 """
 import json
 import pathlib
@@ -74,15 +74,16 @@ def main(
 
                 col_output_rank = maximum_output_rank(cr, b1, b2, b_nq, col_input.shape)
                 row_output_rank = maximum_output_rank(cr, b1, b2, b_nq, row_input.shape)
-                logger.debug(f"Name: {name} "
+                logger.debug(
+                    f"Name: {name} "
                     f"Shape: {weights.shape} "
                     f"Row Output Rank: {row_output_rank} "
                     f"Col Output Rank: {col_output_rank}"
-                    )
+                )
 
                 if b1 == b2:
                     assert col_output_rank == row_output_rank
-                
+
                 col_sketch = lplr(
                     col_input, col_output_rank, b1, b2, sketch=sketch, **kwargs
                 )