Files
enginex-ascend-910-vllm/vllm_ascend/model_loader/rfork/transfer_backend.py
Sun Ruoxi 7f8a1b1f7a init v0.23.0
Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>
2026-08-27 15:11:51 +08:00

586 lines
21 KiB
Python

#
# Copyright (c) 2026 Huawei Technologies Co., Ltd. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import time
from bisect import bisect_left
from typing import Any
import requests
import torch
from torch import nn
from vllm.logger import logger
from vllm.utils.network_utils import get_ip, get_open_port, join_host_port
MAX_TRANSFER_CHUNK_BYTES = 1024**3
MAX_TRANSFER_CHUNK_WEIGHTS = 512
def _normalize_weight_shape(shape: Any) -> tuple[int, ...] | None:
if shape is None:
return None
if not isinstance(shape, (list, tuple)):
return None
if not all(isinstance(dim, int) and dim >= 0 for dim in shape):
return None
return tuple(shape)
def _parse_weight_info(weight_info: Any):
if not isinstance(weight_info, (list, tuple)) or len(weight_info) not in (3, 4):
return None
seed_ptr, seed_len, seed_size = weight_info[:3]
if not all(isinstance(value, int) for value in (seed_ptr, seed_len, seed_size)):
return None
seed_shape = None
if len(weight_info) == 4:
seed_shape = _normalize_weight_shape(weight_info[3])
if seed_shape is None:
return None
return seed_ptr, seed_len, seed_size, seed_shape
def _reshape_tensor_to_seed_shape(
name: str,
tensor: torch.Tensor,
seed_shape: tuple[int, ...] | None,
reshape_events: list[tuple[str, tuple[int, ...], tuple[int, ...]]] | None = None,
) -> bool:
if seed_shape is None or tuple(tensor.shape) == seed_shape:
return True
if tensor.numel() != _numel_from_shape(seed_shape):
logger.error(
"Weight shape mismatch for %s, local shape %s cannot view as seed shape %s",
name,
tuple(tensor.shape),
seed_shape,
)
return False
local_shape = tuple(tensor.shape)
try:
tensor.data = tensor.data.view(seed_shape)
except Exception as e:
logger.error(
"Failed to reshape RFork tensor %s from %s to seed shape %s: %s",
name,
local_shape,
seed_shape,
e,
)
return False
if reshape_events is not None:
reshape_events.append((name, local_shape, seed_shape))
return True
def _update_registered_weight_shape(
weight_shape_dict: dict[str, tuple[int, ...]] | None,
name: str,
tensor: torch.Tensor,
) -> None:
if isinstance(weight_shape_dict, dict):
weight_shape_dict[name] = tuple(tensor.shape)
def _numel_from_shape(shape: tuple[int, ...]) -> int:
numel = 1
for dim in shape:
numel *= dim
return numel
def _is_transferable_tensor(tensor: torch.Tensor) -> bool:
return not tensor.is_meta and tensor.numel() > 0 and _is_tensor_on_transfer_device(tensor)
def _is_tensor_on_transfer_device(tensor: torch.Tensor) -> bool:
return tensor.device.type == "npu"
def _iter_tensors_in_value(prefix: str, value: Any, visited_object_ids: set[int], scan_objects: bool = False):
if isinstance(value, torch.Tensor):
yield prefix, value
return
if isinstance(value, (nn.Module, str, bytes)) or callable(value):
return
if isinstance(value, (list, tuple)):
for index, item in enumerate(value):
yield from _iter_tensors_in_value(f"{prefix}.{index}", item, visited_object_ids, scan_objects)
return
if isinstance(value, dict):
for key, item in value.items():
yield from _iter_tensors_in_value(f"{prefix}.{key}", item, visited_object_ids, scan_objects)
return
if not scan_objects or not hasattr(value, "__dict__"):
return
value_id = id(value)
if value_id in visited_object_ids:
return
visited_object_ids.add(value_id)
for attr_name, attr_value in vars(value).items():
if attr_name.startswith("_"):
continue
yield from _iter_tensors_in_value(f"{prefix}.{attr_name}", attr_value, visited_object_ids, scan_objects)
def _try_collect_transferable_tensor(
name: str,
tensor: torch.Tensor,
seen_data_ptrs: set[int],
collected_tensors: list[tuple[str, torch.Tensor]],
) -> tuple[bool, bool]:
if not _is_transferable_tensor(tensor):
return False, False
data_ptr = tensor.data_ptr()
if data_ptr in seen_data_ptrs:
return False, True
seen_data_ptrs.add(data_ptr)
collected_tensors.append((name, tensor))
return True, False
def _collect_transferable_tensors(model: nn.Module) -> list[tuple[str, torch.Tensor]]:
seen_data_ptrs: set[int] = set()
collected_tensors: list[tuple[str, torch.Tensor]] = []
for name, tensor in model.named_parameters():
_try_collect_transferable_tensor(
name,
tensor,
seen_data_ptrs,
collected_tensors,
)
for name, tensor in model.named_buffers():
_try_collect_transferable_tensor(
name,
tensor,
seen_data_ptrs,
collected_tensors,
)
# Some Ascend post-load paths replace checkpoint parameters with runtime
# tensors stored as plain module attributes, e.g. MLA/SFA W_UV and W_UK_T.
for module_prefix, module in model.named_modules():
for attr_name, attr_value in vars(module).items():
if attr_name.startswith("_") or isinstance(attr_value, nn.Module):
continue
scan_objects = attr_name == "impl"
for tensor_name, tensor in _iter_tensors_in_value(attr_name, attr_value, set(), scan_objects):
full_name = f"{module_prefix}.{tensor_name}" if module_prefix else tensor_name
_try_collect_transferable_tensor(
full_name,
tensor,
seen_data_ptrs,
collected_tensors,
)
return collected_tensors
def _iter_transferable_tensors(model: nn.Module):
yield from _collect_transferable_tensors(model)
def _block_contains_weight_ptr(address: int, size: int, sorted_weight_ptrs: list[int]) -> bool:
index = bisect_left(sorted_weight_ptrs, address)
return index < len(sorted_weight_ptrs) and sorted_weight_ptrs[index] < address + size
def _iter_transfer_chunks(
weight_names: list[str],
seed_ptr_list: list[int],
client_ptr_list: list[int],
client_len_list: list[int],
):
chunk_start = 0
chunk_bytes = 0
chunk_weights = 0
for index, length in enumerate(client_len_list):
should_flush = chunk_weights > 0 and (
chunk_bytes + length > MAX_TRANSFER_CHUNK_BYTES or chunk_weights >= MAX_TRANSFER_CHUNK_WEIGHTS
)
if should_flush:
yield (
weight_names[chunk_start:index],
seed_ptr_list[chunk_start:index],
client_ptr_list[chunk_start:index],
client_len_list[chunk_start:index],
)
chunk_start = index
chunk_bytes = 0
chunk_weights = 0
chunk_bytes += length
chunk_weights += 1
if chunk_weights > 0:
yield (
weight_names[chunk_start:],
seed_ptr_list[chunk_start:],
client_ptr_list[chunk_start:],
client_len_list[chunk_start:],
)
class RForkTransferBackend:
def __init__(self):
self.rfork_transfer_engine: Any | None = None
self.rfork_transfer_engine_session_id = None
self.rfork_transfer_engine_weights_info_dict = None
self.rfork_transfer_engine_weights_shape_dict = None
self.registered_weight_blocks = []
self._registered_transferable_tensors: list[tuple[str, torch.Tensor]] | None = None
self._is_initialized = False
self.init_transfer_engine()
def init_transfer_engine(self):
try:
from yr.datasystem import TransferEngine # type: ignore[import-not-found]
except ImportError as e:
err_msg = (
"Failed to import TransferEngine from yr.datasystem. "
"Please install @yuanrong-datasystem/transfer_engine."
)
logger.error(err_msg)
raise ImportError(err_msg) from e
transfer_engine = TransferEngine()
local_hostname = join_host_port(get_ip(), get_open_port())
ret = transfer_engine.initialize(local_hostname, "ascend", f"npu:{torch.npu.current_device()}")
if ret.is_error():
err_msg = (
f"TransferEngine initialization failed: "
f"initialize({local_hostname}, 'ascend', "
f"'npu:{int(torch.npu.current_device())}') -> {ret.to_string()}"
)
logger.error(err_msg)
raise RuntimeError(err_msg)
self.rfork_transfer_engine = transfer_engine
self.rfork_transfer_engine_session_id = local_hostname
self._is_initialized = True
def is_initialized(self) -> bool:
return self._is_initialized
def _get_transfer_engine(self) -> Any:
if self.rfork_transfer_engine is None:
raise RuntimeError("TransferEngine is not initialized.")
return self.rfork_transfer_engine
def register_memory_region(self, model):
transfer_engine = self._get_transfer_engine()
start_reg_mr_time = time.perf_counter()
self._registered_transferable_tensors = None
weight_mr_dict = {}
weight_shape_dict = {}
weight_addr_set = set()
transferable_tensors = list(_iter_transferable_tensors(model))
for name, weight in transferable_tensors:
weight_mr_dict[name] = (
weight.data_ptr(),
weight.numel(),
weight.element_size(),
)
weight_shape_dict[name] = tuple(weight.shape)
weight_addr_set.add(weight.data_ptr())
sorted_weight_ptrs = sorted(weight_addr_set)
memory_snapshot = torch.npu.memory.memory_snapshot()
weight_blocks_for_reg_mr = []
for segment in memory_snapshot:
current_weight_block = None
for block in segment.get("blocks", []):
address = block.get("address", -1)
size = block.get("size", -1)
state = block.get("state", "")
if address < 0 or size < 0 or state == "":
continue
if state == "active_allocated" and _block_contains_weight_ptr(address, size, sorted_weight_ptrs):
if current_weight_block is None:
current_weight_block = (address, size)
elif current_weight_block[0] + current_weight_block[1] == address:
current_weight_block = (
current_weight_block[0],
current_weight_block[1] + size,
)
else:
weight_blocks_for_reg_mr.append(current_weight_block)
current_weight_block = (address, size)
if current_weight_block is not None:
weight_blocks_for_reg_mr.append(current_weight_block)
addresses, sizes = zip(*weight_blocks_for_reg_mr) if weight_blocks_for_reg_mr else ((), ())
ret = transfer_engine.batch_register_memory(addresses, sizes)
if ret.is_error():
self._registered_transferable_tensors = None
logger.error(
"batch_register_memory failed for %d blocks, ret: %s",
len(weight_blocks_for_reg_mr),
ret.to_string(),
)
return False
self.rfork_transfer_engine_weights_info_dict = weight_mr_dict
self.rfork_transfer_engine_weights_shape_dict = weight_shape_dict
self.registered_weight_blocks = weight_blocks_for_reg_mr
self._registered_transferable_tensors = transferable_tensors
logger.info(
"register_memory_region time: %.4fs, weights: %d",
time.perf_counter() - start_reg_mr_time,
len(weight_mr_dict),
)
return True
def unregister_memory_region(self) -> bool:
transfer_engine = self._get_transfer_engine()
start_unreg_mr_time = time.perf_counter()
if not self.registered_weight_blocks:
self.rfork_transfer_engine_weights_info_dict = None
self.rfork_transfer_engine_weights_shape_dict = None
self._registered_transferable_tensors = None
logger.debug("unregister_memory_region skipped because no blocks are registered.")
return True
ret = transfer_engine.batch_unregister_memory([address for address, _ in self.registered_weight_blocks])
if ret.is_error():
logger.error(
"batch_unregister_memory failed for %d blocks, ret: %s",
len(self.registered_weight_blocks),
ret.to_string(),
)
return False
self.rfork_transfer_engine_weights_info_dict = None
self.rfork_transfer_engine_weights_shape_dict = None
self.registered_weight_blocks = []
self._registered_transferable_tensors = None
logger.info(
"unregister_memory_region time: %.4fs",
time.perf_counter() - start_unreg_mr_time,
)
return True
def recv_from_source(
self,
model,
seed_instance_ip,
seed_instance_service_port,
local_seed_key,
):
transfer_engine = self._get_transfer_engine()
seed_url = f"http://{seed_instance_ip}:{seed_instance_service_port}"
seed_session_id, seed_weight_info, seed_weight_shapes = get_remote_instance_transfer_engine_info(
seed_url,
local_seed_key,
)
if seed_session_id is None or seed_weight_info is None:
self._registered_transferable_tensors = None
logger.error("Cannot get transfer engine session or weight info.")
return False
transferable_tensors = getattr(self, "_registered_transferable_tensors", None)
if transferable_tensors is None:
transferable_tensors = list(_iter_transferable_tensors(model))
seed_ptr_list = []
client_ptr_list = []
client_len_list = []
weight_names = []
reshape_events: list[tuple[str, tuple[int, ...], tuple[int, ...]]] = []
try:
for name, tensor in transferable_tensors:
weight_info = seed_weight_info.get(name, None)
if weight_info is None:
logger.error("Cannot find weight info for %s.", name)
return False
parsed_weight_info = _parse_weight_info(weight_info)
if parsed_weight_info is None:
logger.error("Invalid weight info for %s: %s", name, weight_info)
return False
seed_ptr, seed_len, seed_size, seed_shape = parsed_weight_info
if seed_shape is None and isinstance(seed_weight_shapes, dict):
seed_shape = _normalize_weight_shape(seed_weight_shapes.get(name))
if seed_len != tensor.numel() or seed_size != tensor.element_size():
logger.error(
"Weight info mismatch for %s, expected (%s, %s), got (%s, %s)",
name,
seed_len,
seed_size,
tensor.numel(),
tensor.element_size(),
)
return False
if not _reshape_tensor_to_seed_shape(name, tensor, seed_shape, reshape_events):
return False
_update_registered_weight_shape(
self.rfork_transfer_engine_weights_shape_dict,
name,
tensor,
)
seed_ptr_list.append(seed_ptr)
client_ptr_list.append(tensor.data_ptr())
client_len_list.append(tensor.numel() * tensor.element_size())
weight_names.append(name)
finally:
self._registered_transferable_tensors = None
transferable_tensors = None
if reshape_events:
sample_events = ", ".join(
f"{name}: {local_shape}->{seed_shape}" for name, local_shape, seed_shape in reshape_events[:3]
)
if len(reshape_events) > 3:
sample_events += ", ..."
logger.debug(
"RFork reshaped %d tensors to match seed shapes: %s",
len(reshape_events),
sample_events,
)
transfer_chunks = list(
_iter_transfer_chunks(
weight_names,
seed_ptr_list,
client_ptr_list,
client_len_list,
)
)
total_transfer_bytes = sum(client_len_list)
transfer_start_time = time.perf_counter()
logger.info(
"transfer weights starts, weights: %d, chunks: %d, total bytes: %.2f GiB",
len(client_len_list),
len(transfer_chunks),
total_transfer_bytes / (1024**3),
)
for index, (chunk_names, chunk_seed_ptrs, chunk_client_ptrs, chunk_lengths) in enumerate(transfer_chunks, 1):
chunk_start_time = time.perf_counter()
logger.debug(
"transfer weights chunk %d/%d starts, weights: %d, bytes: %.2f GiB, first: %s, last: %s",
index,
len(transfer_chunks),
len(chunk_lengths),
sum(chunk_lengths) / (1024**3),
chunk_names[0],
chunk_names[-1],
)
ret = transfer_engine.batch_transfer_sync_read(
seed_session_id,
chunk_client_ptrs,
chunk_seed_ptrs,
chunk_lengths,
)
if ret.is_error():
logger.error(
"Failed to transfer weights chunk %d/%d, first: %s, last: %s, ret=%s",
index,
len(transfer_chunks),
chunk_names[0],
chunk_names[-1],
ret.to_string(),
)
return False
logger.debug(
"transfer weights chunk %d/%d done, time: %.4fs",
index,
len(transfer_chunks),
time.perf_counter() - chunk_start_time,
)
transfer_time = time.perf_counter() - transfer_start_time
logger.info("transfer weights time: %.4fs", transfer_time)
return True
def get_remote_instance_transfer_engine_info(seed_url: str, local_seed_key: str):
try:
response = requests.get(
f"{seed_url}/get_rfork_transfer_engine_info",
params={"seed_key": local_seed_key},
)
if response.status_code != 200:
logger.error(
"GET %s/get_rfork_transfer_engine_info failed: %s",
seed_url,
response.status_code,
)
return None, None, None
data = response.json()
info = data.get("rfork_transfer_engine_info", None)
if info is not None and isinstance(info, list) and len(info) == 2:
shape_info = get_remote_instance_weight_shape_info(seed_url, local_seed_key)
return info[0], info[1], shape_info
logger.error(
"Failed to get rfork_transfer_engine_info in response from %s.",
seed_url,
)
return None, None, None
except Exception as e:
logger.error("Exception getting transfer engine info from %s: %s", seed_url, e)
return None, None, None
def get_remote_instance_weight_shape_info(seed_url: str, local_seed_key: str):
try:
response = requests.get(
f"{seed_url}/get_rfork_transfer_engine_shape_info",
params={"seed_key": local_seed_key},
)
if response.status_code != 200:
logger.debug(
"GET %s/get_rfork_transfer_engine_shape_info failed: %s",
seed_url,
response.status_code,
)
return None
data = response.json()
info = data.get("rfork_transfer_engine_shape_info", None)
if info is None or isinstance(info, dict):
return info
logger.error(
"Failed to get rfork_transfer_engine_shape_info in response from %s.",
seed_url,
)
return None
except Exception as e:
logger.debug("Exception getting transfer engine shape info from %s: %s", seed_url, e)
return None