EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Misc] Remove useless weight loader patch (#5619)

Browse Source 
The patch for weight loader is useless now. Let's remove it

- vLLM version: v0.13.0
- vLLM main:
8be6432bda

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2026-01-06 20:17:32 +08:00
committed by GitHub
parent 089ca2ddcc
commit cd1162e25a
3 changed files with 5 additions and 59 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
vllm_ascend/patch/__init__.py
View File

@@ -184,19 +184,7 @@
# Future Plan:
# Remove this patch when vLLM support the dispatch function.
#
# ** 7. File: worker/patch_weight_loader.py**
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# 1. `vllm.model_executor.layers.linear.UnquantizedLinearMethod`
# Why:
# vLLM Ascend doesn't work with weight loader v2
# How：
# patch it to fix the bug.
# Related PR (if no, explain why):
# This is a bug by Ascend only. We should fix it soon
# Future Plan:
# Remove this patch when the bug is fixed.
#
# ** 8. File: worker/patch_qwen3_next_mtp.py**
# ** 7. File: worker/patch_qwen3_next_mtp.py**
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# 1. `vllm.v1.worker.utils.bind_kv_cache`
# Why:
@@ -209,7 +197,7 @@
# Future Plan:
# Remove this patch after discussing with vllm community and adapting bind_kv_cache to npu.
#
# ** 9. File: worker/patch_module.py**
# ** 8. File: worker/patch_module.py**
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# 1. `vllm.v1.attention.backends.gdn_attn.torch.argsort`
# Why:
@@ -225,7 +213,7 @@
# Remove this patch when bool is supported in 'torch.argsort' func of npu.
# Make 'torch.argsort' in `vllm.v1.attention.backends.gdn_attn` be stable.
#
# ** 10. File: worker/patch_rejection_sampler.py**
# ** 9. File: worker/patch_rejection_sampler.py**
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# 1. `vllm.v1.sample.rejection_sampler`
# Why:
@@ -241,7 +229,7 @@
# to override them, then delete the patch file `worker/patch_rejection_sampler.py`.
# 2. make these functions as costom op, then remove AscendRejectionSampler
#
# ** 11.File: worker/patch_qwen3_next.py**
# ** 10.File: worker/patch_qwen3_next.py**
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# 1. `vllm.model_executor.models.qwen3_next.Qwen3NextGatedDeltaNet.forward`
# Why:
@@ -253,7 +241,7 @@
# Future Plan:
# Remove this patch when vLLM support these operators.
#
# ** 12. File: worker/patch_qwen3_next.py**
# ** 11. File: worker/patch_qwen3_next.py**
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# 1. `vllm.model_executor.models.qwen3_next.Qwen3NextGatedDeltaNet._forward_core`
# Why:

1
vllm_ascend/patch/worker/__init__.py
View File

@@ -25,7 +25,6 @@ import vllm_ascend.patch.platform.patch_sched_yield # noqa
import vllm_ascend.patch.worker.patch_bert # noqa
import vllm_ascend.patch.worker.patch_distributed # noqa
import vllm_ascend.patch.worker.patch_deepseek # noqa
import vllm_ascend.patch.worker.patch_weight_loader # noqa
import vllm_ascend.patch.worker.patch_multimodal_merge # noqa
import vllm_ascend.patch.worker.patch_minicpm # noqa
import vllm_ascend.patch.worker.patch_rope # noqa

41
vllm_ascend/patch/worker/patch_weight_loader.py
View File

@@ -1,41 +0,0 @@
import torch
from torch.nn.parameter import Parameter
from vllm.logger import init_logger
from vllm.model_executor.layers.linear import UnquantizedLinearMethod
from vllm.model_executor.utils import set_weight_attrs
from vllm.utils.mem_constants import GiB_bytes
logger = init_logger(__name__)
def create_weights(self, layer: torch.nn.Module, input_size_per_partition: int,
output_partition_sizes: list[int], input_size: int,
output_size: int, params_dtype: torch.dtype,
**extra_weight_attrs):
# This method creates unquantized linear weights.
# The weights are not quantized, and they are not sharded.
# The amount of memory allocated for the weights is
# sum(output_partition_sizes) * input_size_per_partition.
try:
weight = Parameter(torch.empty(sum(output_partition_sizes),
input_size_per_partition,
dtype=params_dtype),
requires_grad=False)
except torch.cuda.OutOfMemoryError as e:
logger.error("Failed to create unquantized linear weights: %s", e)
if torch.cuda.is_available():
logger.debug("CUDA device: %s", torch.cuda.current_device())
logger.debug("Allocated: %.2f GiB",
torch.cuda.memory_allocated() / GiB_bytes)
logger.debug("Reserved: %.2f GiB",
torch.cuda.memory_reserved() / GiB_bytes)
raise RuntimeError(
"Failed to create unquantized linear weights. "
"This may be caused by insufficient memory to allocate "
"the weight.") from e
set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0})
layer.register_parameter("weight", weight)
set_weight_attrs(weight, extra_weight_attrs)
UnquantizedLinearMethod.create_weights = create_weights