chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Revert "[Feature] PD-Multiplexing Context and Scheduler." (#12267)

Browse Source
This commit is contained in:
Yineng Zhang
2025-10-28 02:00:37 -07:00
committed by GitHub
parent 41a113356a
commit ea39952797
9 changed files with 26 additions and 457 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/tp_worker.py
+1 -24
View File
@@ -35,7 +35,7 @@ from sglang.srt.managers.io_struct import (
UpdateWeightsFromIPCReqInput,
UpdateWeightsFromTensorReqInput,
)
from sglang.srt.managers.schedule_batch import ModelWorkerBatch, ScheduleBatch
from sglang.srt.managers.schedule_batch import ModelWorkerBatch
from sglang.srt.managers.scheduler import GenerationBatchResult
from sglang.srt.mem_cache.allocator import BaseTokenToKVPoolAllocator
from sglang.srt.mem_cache.memory_pool import ReqToTokenPool
@@ -425,26 +425,3 @@ class TpModelWorker(BaseTpWorker):
pp_hidden_states_proxy_tensors=pp_proxy_tensors,
can_run_cuda_graph=can_run_cuda_graph,
)
def forward_batch_split_prefill(self, batch: ScheduleBatch):
if batch.split_index == 0:
model_worker_batch = batch.get_model_worker_batch()
forward_batch = ForwardBatch.init_new(model_worker_batch, self.model_runner)
batch.split_forward_batch = forward_batch
batch.seq_lens_cpu_cache = model_worker_batch.seq_lens_cpu
else:
model_worker_batch = batch.get_model_worker_batch(batch.seq_lens_cpu_cache)
logits_output, can_run_cuda_graph = self.model_runner.forward(
batch.split_forward_batch, split_forward_count=batch.split_forward_count
)
if logits_output:
next_token_ids = self.model_runner.sample(logits_output, model_worker_batch)
else:
next_token_ids = None
batch_result = GenerationBatchResult(
logits_output=logits_output,
can_run_cuda_graph=can_run_cuda_graph,
)
batch_result.next_token_ids = next_token_ids
return batch_result