[Compile] Remove torch.compile-based sequence parallelism and AsyncTP (#60517)
This commit is contained in:
@@ -12,7 +12,6 @@ from fnmatch import filter as fnmatch_filter
|
||||
from types import NoneType
|
||||
from typing import TYPE_CHECKING, Any, cast
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
@@ -98,7 +97,6 @@ from vllm.v1.worker.startup_plan import (
|
||||
maybe_apply_startup_plan,
|
||||
maybe_save_startup_plan,
|
||||
)
|
||||
from vllm.v1.worker.utils import is_residual_scattered_for_sp
|
||||
from vllm.v1.worker.worker_base import CompilationTimes, WorkerBase
|
||||
from vllm.v1.worker.workspace import init_workspace_manager
|
||||
|
||||
@@ -1273,45 +1271,10 @@ class Worker(WorkerBase):
|
||||
|
||||
intermediate_tensors = None
|
||||
forward_pass = scheduler_output.total_num_scheduled_tokens > 0
|
||||
num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
|
||||
all_gather_tensors = {}
|
||||
compilation_config = self.vllm_config.compilation_config
|
||||
parallel_config = self.vllm_config.parallel_config
|
||||
|
||||
if (
|
||||
parallel_config.pipeline_parallel_size > 1
|
||||
and compilation_config.pass_config.enable_sp
|
||||
and forward_pass
|
||||
):
|
||||
# currently only supported by V1 GPUModelRunner
|
||||
assert not self.use_v2_model_runner
|
||||
num_scheduled_tokens_np = np.array(
|
||||
list(scheduler_output.num_scheduled_tokens.values()),
|
||||
dtype=np.int32,
|
||||
)
|
||||
# TODO(lucas): This is pretty gross; ideally we should only ever call
|
||||
# `_determine_batch_execution_and_padding` once (will get called again
|
||||
# in `execute_model`) but this requires a larger refactor of PP.
|
||||
_, batch_desc, _, _, _ = (
|
||||
self.model_runner._determine_batch_execution_and_padding(
|
||||
num_tokens=num_scheduled_tokens,
|
||||
num_reqs=len(num_scheduled_tokens_np),
|
||||
num_scheduled_tokens_np=num_scheduled_tokens_np,
|
||||
max_num_scheduled_tokens=num_scheduled_tokens_np.max(),
|
||||
use_cascade_attn=False, # TODO(lucas): Handle cascade attention
|
||||
)
|
||||
)
|
||||
all_gather_tensors = {
|
||||
"residual": not is_residual_scattered_for_sp(
|
||||
self.vllm_config, batch_desc.num_tokens
|
||||
)
|
||||
}
|
||||
|
||||
if forward_pass and not get_pp_group().is_first_rank:
|
||||
tensor_dict, comm_handles, comm_postprocess = (
|
||||
get_pp_group().irecv_tensor_dict(
|
||||
all_gather_group=get_tp_group(),
|
||||
all_gather_tensors=all_gather_tensors,
|
||||
)
|
||||
)
|
||||
assert tensor_dict is not None
|
||||
@@ -1349,7 +1312,6 @@ class Worker(WorkerBase):
|
||||
handles = get_pp_group().isend_tensor_dict(
|
||||
output.tensors,
|
||||
all_gather_group=get_tp_group(),
|
||||
all_gather_tensors=all_gather_tensors,
|
||||
)
|
||||
self._pp_send_work = handles[1:]
|
||||
|
||||
|
||||
Reference in New Issue
Block a user