From a0af3ea0534eaeb8ab1ebbc5a98c4744442ff23d Mon Sep 17 00:00:00 2001 From: liu lili Date: Thu, 27 Mar 2025 10:10:33 +0800 Subject: [PATCH] lll: optimize pad block tables --- .../model_executor/models/mf_models/mf_model_base.py | 11 ++++------- 1 file changed, 4 insertions(+), 7 deletions(-) diff --git a/vllm_mindspore/model_executor/models/mf_models/mf_model_base.py b/vllm_mindspore/model_executor/models/mf_models/mf_model_base.py index 91a0446e6..7d661d9ba 100644 --- a/vllm_mindspore/model_executor/models/mf_models/mf_model_base.py +++ b/vllm_mindspore/model_executor/models/mf_models/mf_model_base.py @@ -56,13 +56,10 @@ def _pad_block_table(block_tables, seq_length, block_size, pad_size): ) return fake_block_tables - block_tables_list = block_tables.tolist() - padded_block_tables = [ - _pad_to_max(block_table, seq_length // block_size) - for block_table in block_tables_list - ] - - return Tensor(np.array(padded_block_tables).astype(np.int32)) + block_tables_np = block_tables.asnumpy().astype(np.int32) + pad_size = seq_length // block_size - block_tables_np.shape[1] + block_tables_np = np.pad(block_tables_np, ((0, 0),(0, pad_size)), 'constant', constant_values=(-1, -1)) + return block_tables_np def _batch_seq(input_tokens, prefill): -- Gitee