优化tokenizer和 kv cache内存

main_ax650

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bdbd72d706b584081c103af2fe6a59426140f0ed0b79d9baf4282a601e2e014c
-size 6593472
+oid sha256:9f6cfcf0021a15a3baea0513e0eb6b17bdfbe08928de0e74619e6684a13a1493
+size 6808392

run_qwen2_5_vl_image.sh

@@ -7,7 +7,7 @@ AXMODEL_DIR=./Qwen2.5-VL-3B-Instruct-AX650-chunk_prefill_512
 --bos 0 --eos 0 \
 --dynamic_load_axmodel_layer 0 \
 --use_mmap_load_embed 1 \
---filename_tokenizer_model "http://127.0.0.1:12345" \
+--filename_tokenizer_model "qwen2.5_tokenizer.txt" \
 --filename_post_axmodel "${AXMODEL_DIR}/qwen2_5_vl_post.axmodel" \
 --use_topk 0 \
 --filename_tokens_embed "${AXMODEL_DIR}/model.embed_tokens.weight.bfloat16.bin" \

run_qwen2_5_vl_video.sh

@@ -7,7 +7,7 @@ AXMODEL_DIR=./Qwen2.5-VL-3B-Instruct-AX650-chunk_prefill_512
 --bos 0 --eos 0 \
 --dynamic_load_axmodel_layer 0 \
 --use_mmap_load_embed 1 \
---filename_tokenizer_model "http://127.0.0.1:12345" \
+--filename_tokenizer_model "qwen2.5_tokenizer.txt" \
 --filename_post_axmodel "${AXMODEL_DIR}/qwen2_5_vl_post.axmodel" \
 --use_topk 0 \
 --filename_tokens_embed "${AXMODEL_DIR}/model.embed_tokens.weight.bfloat16.bin" \