github
/
ColossalAI
mirror of https://github.com/hpcaitech/ColossalAI
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

diverse tests

pull/5434/head
Runyu Lu 2024-03-14 10:35:08 +08:00
parent 1821a6dab0
commit ae24b4f025
2 changed files with 5 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
colossalai/inference/core/engine.py
View File

@ -117,7 +117,8 @@ class InferenceEngine:
max_context_len_to_capture = self.inference_config.max_context_len_to_capture max_context_len_to_capture = self.inference_config.max_context_len_to_capture
max_num_blocks = (max_context_len_to_capture + block_size - 1) // block_size max_num_blocks = (max_context_len_to_capture + block_size - 1) // block_size
input_tokens_ids = torch.zeros(max_batch_size, dtype=torch.long).cuda() input_tokens_ids = torch.zeros(max_batch_size, dtype=torch.long).cuda()
self.graph_block_tables = np.zeros((max(_BATCH_SIZES_TO_CAPTURE), max_num_blocks), dtype=np.int32) # self.graph_block_tables = np.zeros((max(_BATCH_SIZES_TO_CAPTURE), max_num_blocks), dtype=np.int32)
self.graph_block_tables = np.full((max(_BATCH_SIZES_TO_CAPTURE), max_num_blocks), -1, dtype=np.int32)
self.graph_block_tables[:, 0] = np.arange(max_num_blocks, max_num_blocks + max(_BATCH_SIZES_TO_CAPTURE)) self.graph_block_tables[:, 0] = np.arange(max_num_blocks, max_num_blocks + max(_BATCH_SIZES_TO_CAPTURE))
self.graph_block_tables[0, :] = np.arange( self.graph_block_tables[0, :] = np.arange(
0, max_num_blocks 0, max_num_blocks

4
tests/test_infer/test_cuda_graph.py
View File

@ -34,7 +34,9 @@ def check_inference_engine(use_cuda_graph=False, batch_size=32):
prompts_token_ids = [] prompts_token_ids = []
for i in range(batch_size): for i in range(batch_size):
prompts_token_ids.append(np.random.randint(low=0, high=100, size=random.randint(1, 1024)).tolist()) prompts_token_ids.append(
np.random.randint(low=0, high=100, size=random.randint(1, max(1024 // batch_size, 32))).tolist()
)
input_len = 1024 input_len = 1024
output_len = 128 output_len = 128