.

MaxText/benchmarks/decode_paged_attention-7b.sh

@@ -3,17 +3,17 @@ python MaxText/decode.py \
     MaxText/configs/base.yml \
     model_name=llama2-7b \
     tokenizer_path=assets/tokenizer.llama2 \
-    per_device_batch_size=8 \
-    max_prefill_predict_length=512 \
-    max_target_length=1024 \
+    per_device_batch_size=1 \
+    max_prefill_predict_length=1024 \
+    max_target_length=2048 \
     weight_dtype=bfloat16 \
     ici_fsdp_parallelism=1 \
     ici_tensor_parallelism=-1 \
     scan_layers=false \
-    load_parameters_path=gs://patemotter/checkpoints/quant_llama2-7b-chat/20240906200810/intmp_mp_scale \
+    load_parameters_path=gs://msingh-bkt/checkpoints/quant_llama2-7b-chat/20241120034012/int8_ \
     quantization=int8 \
     checkpoint_is_quantized=true \
-    attention=dot_product \
+    attention=paged \
     num_pages=64 \
     page_size=32 \
     block_size=256