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

Drop 0.11.0 support (#4377)

Browse Source 
There is a lot hack code for v0.11.0, which makes the code hard to
upgrade to newer vLLM version. Since v0.11.0 will release soon. Let's
drop v0.11.0 support first. Then we'll upgrade to v0.11.2 soon.


- vLLM version: v0.11.0
- vLLM main:
2918c1b49c

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-11-24 17:08:20 +08:00
committed by GitHub
parent 41ddb06554
commit a1f142b7ad
80 changed files with 467 additions and 1755 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
vllm_ascend/patch/platform/patch_mamba_config.py
View File

@@ -3,23 +3,10 @@ import vllm.model_executor.models.config
from vllm.logger import init_logger
from vllm.model_executor.models import ModelRegistry
from vllm.model_executor.models.config import MambaModelConfig
from vllm_ascend.utils import vllm_version_is
if vllm_version_is("0.11.0"):
from vllm.utils import cdiv
else:
from vllm.utils.math_utils import cdiv
from vllm.utils.math_utils import cdiv
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
from vllm.v1.kv_cache_interface import FullAttentionSpec, MambaSpec
from vllm_ascend.utils import vllm_version_is
if vllm_version_is("0.11.0"):
from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
else:
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
@classmethod
def verify_and_update_config(cls, vllm_config) -> None:

13
vllm_ascend/patch/platform/patch_multiproc_executor.py
View File

@@ -8,21 +8,14 @@ import vllm.v1.executor.multiproc_executor
from vllm import envs
from vllm.config import VllmConfig
from vllm.distributed.device_communicators.shm_broadcast import MessageQueue
from vllm.utils.network_utils import (get_distributed_init_method,
get_loopback_ip, get_open_port)
from vllm.utils.system_utils import get_mp_context
from vllm.v1.executor.abstract import FailureCallback
from vllm.v1.executor.multiproc_executor import (
MultiprocExecutor, UnreadyWorkerProcHandle, WorkerProc,
set_multiprocessing_worker_envs)
from vllm_ascend.utils import vllm_version_is
if vllm_version_is("0.11.0"):
from vllm.utils import (get_distributed_init_method, get_loopback_ip,
get_mp_context, get_open_port)
else:
from vllm.utils.network_utils import (get_distributed_init_method,
get_loopback_ip, get_open_port)
from vllm.utils.system_utils import get_mp_context
class AscendMultiprocExecutor(MultiprocExecutor):
supports_pp: bool = True

6
vllm_ascend/patch/worker/__init__.py
View File

@@ -28,9 +28,3 @@ import vllm_ascend.patch.worker.patch_roberta # 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
from vllm_ascend.utils import vllm_version_is
if vllm_version_is("0.11.0"):
import vllm_ascend.patch.worker.patch_deepseek_mtp # noqa
import vllm_ascend.patch.worker.patch_deepseek_v3_2 # noqa

94
vllm_ascend/patch/worker/patch_deepseek_mtp.py
View File

@@ -1,94 +0,0 @@
import torch
import torch.nn as nn
from transformers import PretrainedConfig
from vllm.config import VllmConfig
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.model_executor.models.deepseek_mtp import \
DeepSeekMultiTokenPredictorLayer
from vllm.model_executor.models.deepseek_v2 import DeepseekV2DecoderLayer
from vllm.model_executor.models.utils import maybe_prefix
from vllm_ascend.utils import vllm_version_is
if vllm_version_is("0.11.0"):
from vllm.compilation.decorators import support_torch_compile
from vllm.model_executor.models.deepseek_mtp import DeepSeekMTP
class SharedHead(nn.Module):
def __init__(
self,
config: PretrainedConfig,
prefix: str,
quant_config: QuantizationConfig = None,
) -> None:
super().__init__()
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.head = ParallelLMHead(
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "head"),
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
return self.norm(hidden_states)
def predictor_init(self, vllm_config: VllmConfig, prefix: str) -> None:
nn.Module.__init__(self)
config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
self.enorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.hnorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.eh_proj = nn.Linear(config.hidden_size * 2,
config.hidden_size,
bias=False)
# We don't need topk_indices_buffer in Ascend
topk_indices_buffer = None
self.shared_head = SharedHead(config=config,
prefix=prefix,
quant_config=quant_config)
self.mtp_block = DeepseekV2DecoderLayer(vllm_config, prefix,
topk_indices_buffer)
def predictor_forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
previous_hidden_states: torch.Tensor,
inputs_embeds: torch.Tensor | None = None,
spec_step_index: int = 0,
) -> torch.Tensor:
assert inputs_embeds is not None
# masking inputs at position 0, as not needed by MTP
inputs_embeds = torch.where(positions.unsqueeze(-1) == 0, 0, inputs_embeds)
inputs_embeds = self.enorm(inputs_embeds)
previous_hidden_states = self.hnorm(previous_hidden_states)
hidden_states = self.eh_proj(
torch.cat([inputs_embeds, previous_hidden_states], dim=-1))
hidden_states, residual = self.mtp_block(positions=positions,
hidden_states=hidden_states,
residual=None)
hidden_states = residual + hidden_states
return hidden_states
# Patch this only for aclgraph support, as this is not support in vLLM 0.11.0
@support_torch_compile
class AscendDeepSeekMTP(DeepSeekMTP):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__(vllm_config=vllm_config, prefix=prefix)
DeepSeekMultiTokenPredictorLayer.__init__ = predictor_init
if vllm_version_is("0.11.0"):
DeepSeekMultiTokenPredictorLayer.forward = predictor_forward

108
vllm_ascend/patch/worker/patch_deepseek_v3_2.py
View File

@@ -1,108 +0,0 @@
#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# This file is a part of the vllm-ascend project.
#
# 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.
#
from itertools import islice
from typing import Optional, Union
import torch
import vllm.model_executor.models.deepseek_v2
from torch import nn
from vllm.compilation.decorators import support_torch_compile
from vllm.config import VllmConfig
from vllm.distributed import get_pp_group
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.vocab_parallel_embedding import \
VocabParallelEmbedding
from vllm.model_executor.models.deepseek_v2 import DeepseekV2DecoderLayer
from vllm.model_executor.models.utils import (
PPMissingLayer, make_empty_intermediate_tensors_factory, make_layers)
from vllm.sequence import IntermediateTensors
@support_torch_compile
class DeepseekV2Model(nn.Module):
fall_back_to_pt_during_load = False
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
self.config = config
self.vocab_size = config.vocab_size
self.is_v32 = hasattr(config, "index_topk")
topk_indices_buffer = None
if get_pp_group().is_first_rank:
self.embed_tokens = VocabParallelEmbedding(
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=f"{prefix}.embed_tokens")
else:
self.embed_tokens = PPMissingLayer()
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: DeepseekV2DecoderLayer(vllm_config, prefix,
topk_indices_buffer),
prefix=f"{prefix}.layers")
if get_pp_group().is_last_rank:
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
else:
self.norm = PPMissingLayer()
self.make_empty_intermediate_tensors = (
make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], config.hidden_size))
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.embed_tokens(input_ids)
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
intermediate_tensors: Optional[IntermediateTensors],
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
if get_pp_group().is_first_rank:
if inputs_embeds is not None:
hidden_states = inputs_embeds
else:
hidden_states = self.get_input_embeddings(input_ids)
residual = None
else:
assert intermediate_tensors is not None
hidden_states = intermediate_tensors["hidden_states"]
residual = intermediate_tensors["residual"]
for layer in islice(self.layers, self.start_layer, self.end_layer):
hidden_states, residual = layer(positions, hidden_states, residual)
if not get_pp_group().is_last_rank:
return IntermediateTensors({
"hidden_states": hidden_states,
"residual": residual
})
hidden_states, _ = self.norm(hidden_states, residual)
return hidden_states
vllm.model_executor.models.deepseek_v2.DeepseekV2Model = DeepseekV2Model

11
vllm_ascend/patch/worker/patch_triton.py
View File

@@ -6,16 +6,11 @@ import vllm.model_executor.layers.mamba.ops.causal_conv1d
from vllm_ascend.ops.casual_conv1d import (causal_conv1d_fn,
causal_conv1d_update_npu)
from vllm_ascend.ops.fla import LayerNormFn, torch_chunk_gated_delta_rule
from vllm_ascend.ops.sigmoid_gating import (
fused_recurrent_gated_delta_rule_fwd_kernel,
fused_recurrent_gated_delta_rule_fwd_kernel_0_11_0)
from vllm_ascend.utils import vllm_version_is
from vllm_ascend.ops.sigmoid_gating import \
fused_recurrent_gated_delta_rule_fwd_kernel
vllm.model_executor.layers.mamba.ops.causal_conv1d.causal_conv1d_update = causal_conv1d_update_npu
vllm.model_executor.layers.mamba.ops.causal_conv1d.causal_conv1d_fn = causal_conv1d_fn
if vllm_version_is('0.11.0'):
vllm.model_executor.layers.fla.ops.fused_recurrent.fused_recurrent_gated_delta_rule_fwd_kernel = fused_recurrent_gated_delta_rule_fwd_kernel_0_11_0
else:
vllm.model_executor.layers.fla.ops.fused_recurrent.fused_recurrent_gated_delta_rule_fwd_kernel = fused_recurrent_gated_delta_rule_fwd_kernel
vllm.model_executor.layers.fla.ops.fused_recurrent.fused_recurrent_gated_delta_rule_fwd_kernel = fused_recurrent_gated_delta_rule_fwd_kernel
vllm.model_executor.layers.fla.ops.layernorm_guard.LayerNormFn = LayerNormFn
vllm.model_executor.layers.fla.ops.chunk.chunk_gated_delta_rule = torch_chunk_gated_delta_rule

8
vllm_ascend/patch/worker/patch_weight_loader.py
View File

@@ -3,13 +3,7 @@ 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_ascend.utils import vllm_version_is
if vllm_version_is("0.11.0"):
from vllm.utils import GiB_bytes
else:
from vllm.utils.mem_constants import GiB_bytes
from vllm.utils.mem_constants import GiB_bytes
logger = init_logger(__name__)