[V0 Deprecation] Remove multi-step scheduling (#22138)
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu> Signed-off-by: Woosuk Kwon <woosuk@thinkingmachines.ai>
This commit is contained in:
@@ -508,8 +508,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
|
||||
if inter_data.is_prompt:
|
||||
context_len = seq_data.get_num_computed_tokens()
|
||||
seq_len = min(seq_len, context_len + token_chunk_size)
|
||||
elif self.runner.scheduler_config.is_multi_step or \
|
||||
self.runner.model_config.is_encoder_decoder:
|
||||
elif self.runner.model_config.is_encoder_decoder:
|
||||
context_len = seq_len - 1
|
||||
else:
|
||||
context_len = seq_data.get_num_computed_tokens()
|
||||
@@ -778,9 +777,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
|
||||
int: Returns the determined number of padding sequences. If
|
||||
CUDA graphs is not viable, returns -1.
|
||||
"""
|
||||
is_mscp: bool = self.runner.scheduler_config.is_multi_step and \
|
||||
self.runner.scheduler_config.chunked_prefill_enabled
|
||||
decode_only = self.decode_only or is_mscp
|
||||
decode_only = self.decode_only
|
||||
if not decode_only:
|
||||
# Early exit so we can treat num_seqs as the batch_size below.
|
||||
return -1
|
||||
|
||||
Reference in New Issue
Block a user