Inline the single-use split-prefill setup at its caller (#25722)
This commit is contained in:
@@ -2078,11 +2078,6 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
|
||||
req.mamba_last_track_seqlen = mamba_track_seqlen_aligned
|
||||
mamba_track_seqlens_cpu.append(mamba_track_seqlen)
|
||||
|
||||
def prepare_for_split_prefill(self):
|
||||
self.prepare_for_extend()
|
||||
# For split prefill, we need to set the forward mode to SPLIT_PREFILL
|
||||
self.forward_mode = ForwardMode.SPLIT_PREFILL
|
||||
|
||||
def mix_with_running(self, running_batch: "ScheduleBatch"):
|
||||
self.forward_mode = ForwardMode.MIXED
|
||||
running_bs = running_batch.batch_size()
|
||||
|
||||
@@ -15,7 +15,7 @@ import torch
|
||||
from sglang.bench_one_batch import TreeCacheNamespace
|
||||
from sglang.srt.configs.model_config import ModelConfig
|
||||
from sglang.srt.managers.schedule_batch import Req, ScheduleBatch
|
||||
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
|
||||
from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
|
||||
from sglang.srt.model_executor.model_runner import ModelRunner
|
||||
from sglang.srt.sampling.sampling_params import SamplingParams
|
||||
from sglang.srt.server_args import PortArgs, ServerArgs
|
||||
@@ -115,10 +115,10 @@ class TestForwardSplitPrefill(CustomTestCase):
|
||||
enable_overlap=False,
|
||||
spec_algorithm=SpeculativeAlgorithm.NONE,
|
||||
)
|
||||
batch.prepare_for_extend()
|
||||
if is_split_prefill:
|
||||
batch.prepare_for_split_prefill()
|
||||
else:
|
||||
batch.prepare_for_extend()
|
||||
# For split prefill, we need to set the forward mode to SPLIT_PREFILL
|
||||
batch.forward_mode = ForwardMode.SPLIT_PREFILL
|
||||
|
||||
# Create forward batch
|
||||
forward_batch = ForwardBatch.init_new(batch, self.model_runner)
|
||||
|
||||
Reference in New Issue
Block a user