hack around slices bug

Author: bnellnm

vllm/v1/engine/async_llm.py

@@ -251,13 +251,14 @@ async def _run_output_handler(self):
  # event loop for too long.
  num_outputs = len(outputs.new_token_id_offsets)
 
- if num_outputs <= VLLM_V1_OUTPUT_PROC_CHUNK_SIZE:
+ if True or num_outputs <= VLLM_V1_OUTPUT_PROC_CHUNK_SIZE:
  slices = ((0, num_outputs), )
  else:
  slices = []
  parts = np.linspace(
  num_outputs,
- cdiv(num_outputs, VLLM_V1_OUTPUT_PROC_CHUNK_SIZE))
+ cdiv(num_outputs, VLLM_V1_OUTPUT_PROC_CHUNK_SIZE),
+ dtype='int')
  last = 0
  for i in parts:
  slices.append((last, i))