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

upgrade vLLM to main (#4608)

Browse Source 
1. fix https://github.com/vllm-project/vllm/pull/28542
The model structure modifications we involved in are:
     - Qwen2.5-VL(still exist some patch)
     - Qwen2-VL
     - Qwen2
     - DeepSeek series
     - Qwen-moe series
2. fix https://github.com/vllm-project/vllm/pull/29121
   the output token now  type changed from np to `list[list[int]]`

3. fix https://github.com/vllm-project/vllm/pull/29262
    `xformers` backend for multimodal now has been deprecated
4. fix https://github.com/vllm-project/vllm/pull/29342

5. fix https://github.com/vllm-project/vllm/pull/28579
6. fix https://github.com/vllm-project/vllm/pull/28718
7. fix https://github.com/vllm-project/vllm/issues/28665
8. fix https://github.com/vllm-project/vllm/pull/26847
vllm introduced the `optimization-level`, some default config has been
changed, and the param `--enforce-eager` has been deprecated
9. fix http://github.com/vllm-project/vllm/pull/29223 it retuns tuple
for sampler.
10. fix https://github.com/vllm-project/vllm/pull/29471 we'll remove the
related patch to avoid this kind of error.

Co-authored-by: hfadzxy <starmoon_zhang@163.com>
Co-authored-by: wangli <wangli858794774@gmail.com>


- vLLM version: v0.11.2

---------

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Signed-off-by: wangli <wangli858794774@gmail.com>
Signed-off-by: hfadzxy <starmoon_zhang@163.com>
Co-authored-by: wangli <wangli858794774@gmail.com>
Co-authored-by: hfadzxy <starmoon_zhang@163.com>
This commit is contained in:
wangxiyuan
2025-12-02 22:10:52 +08:00
committed by GitHub
parent 4588cdac02
commit 7f2673ea2d
60 changed files with 383 additions and 374 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
vllm_ascend/torchair/models/qwen2.py
View File

@@ -23,7 +23,7 @@ import torch.nn.functional as F
import vllm
from torch import nn
from transformers import Qwen2Config
from vllm.attention import AttentionMetadata, AttentionType
from vllm.attention.backends.abstract import AttentionMetadata, AttentionType
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (get_pp_group, tensor_model_parallel_all_gather,
@@ -40,6 +40,7 @@ from vllm.model_executor.models.qwen2 import Qwen2MLP, Qwen2Model
from vllm.model_executor.models.utils import (AutoWeightsLoader,
PPMissingLayer, maybe_prefix)
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.config import set_default_rope_theta
from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.attention.attention_v1 import AscendAttentionState
@@ -72,11 +73,10 @@ class CustomQwen2Attention(Qwen2Attention):
hidden_size: int,
num_heads: int,
num_kv_heads: int,
rope_parameters: Optional[dict[str, Any]] = None,
max_position: int = 4096 * 32,
rope_theta: float = 10000,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
rope_scaling: Optional[tuple] = None,
prefix: str = "",
attn_type: str = AttentionType.DECODER,
dual_chunk_attention_config: Optional[dict[str, Any]] = None,
@@ -86,13 +86,13 @@ class CustomQwen2Attention(Qwen2Attention):
num_heads=num_heads,
num_kv_heads=num_kv_heads,
max_position=max_position,
rope_theta=rope_theta,
cache_config=cache_config,
quant_config=quant_config,
rope_scaling=rope_scaling,
prefix=prefix,
attn_type=attn_type,
dual_chunk_attention_config=dual_chunk_attention_config)
dual_chunk_attention_config=dual_chunk_attention_config,
rope_parameters=rope_parameters)
ascend_config = get_ascend_config()
self.torchair_graph_enabled = ascend_config.torchair_graph_config.enabled
@@ -145,9 +145,9 @@ class CustomQwen2DecoderLayer(nn.Module):
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
# Requires transformers > 4.32.0
rope_theta = getattr(config, "rope_theta", 1000000)
rope_scaling = getattr(config, "rope_scaling", None)
set_default_rope_theta(config, default_theta=1000000)
dual_chunk_attention_config = getattr(config,
"dual_chunk_attention_config",
None)
@@ -166,10 +166,9 @@ class CustomQwen2DecoderLayer(nn.Module):
num_heads=config.num_attention_heads,
max_position=config.max_position_embeddings,
num_kv_heads=config.num_key_value_heads,
rope_theta=rope_theta,
rope_parameters=config.rope_parameters,
cache_config=cache_config,
quant_config=quant_config,
rope_scaling=rope_scaling,
prefix=f"{prefix}.self_attn",
attn_type=attn_type,
dual_chunk_attention_config=dual_chunk_attention_config,

15
vllm_ascend/torchair/models/qwen3_moe.py
View File

@@ -21,7 +21,8 @@ from typing import Any, List, Optional, Union
import torch
from torch import nn
from transformers import PretrainedConfig
from vllm.attention import Attention, AttentionMetadata
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, CompilationMode, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
@@ -137,8 +138,7 @@ class CustomQwen3MoeAttention(Qwen3MoeAttention):
hidden_size: int,
num_heads: int,
num_kv_heads: int,
rope_theta: float = 10000,
rope_scaling: Optional[dict[str, Any]] = None,
rope_parameters: dict[str, Any],
max_position_embeddings: int = 8192,
head_dim: Optional[int] = None,
rms_norm_eps: float = 1e-06,
@@ -167,7 +167,6 @@ class CustomQwen3MoeAttention(Qwen3MoeAttention):
self.q_size = self.num_heads * self.head_dim
self.kv_size = self.num_kv_heads * self.head_dim
self.scaling = self.head_dim**-0.5
self.rope_theta = rope_theta
self.max_position_embeddings = max_position_embeddings
self.qkv_proj = QKVParallelLinear(hidden_size,
@@ -188,8 +187,7 @@ class CustomQwen3MoeAttention(Qwen3MoeAttention):
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
base=rope_theta,
rope_scaling=rope_scaling,
rope_parameters=rope_parameters,
)
self.attn = Attention(self.num_heads,
self.head_dim,
@@ -270,16 +268,13 @@ class CustomQwen3MoeDecoderLayer(Qwen3MoeDecoderLayer):
nn.Module.__init__(self)
self.hidden_size = config.hidden_size
rope_theta = getattr(config, "rope_theta", 10000)
rope_scaling = getattr(config, "rope_scaling", None)
max_position_embeddings = getattr(config, "max_position_embeddings",
8192)
self.self_attn = CustomQwen3MoeAttention(
hidden_size=self.hidden_size,
num_heads=config.num_attention_heads,
num_kv_heads=config.num_key_value_heads,
rope_theta=rope_theta,
rope_scaling=rope_scaling,
rope_parameters=config.rope_parameters,
max_position_embeddings=max_position_embeddings,
rms_norm_eps=config.rms_norm_eps,
qkv_bias=getattr(config, 'attention_bias', False),

52
vllm_ascend/torchair/models/torchair_deepseek_v2.py
View File

@@ -25,13 +25,13 @@
# # vllm-project/vllm/vllm/model_executor/models/deepseek_v2.py
# """Inference-only DeepseekV2/DeepseekV3 model."""
from typing import Any, Callable, Dict, Iterable, List, Optional, Tuple, Union
from typing import Callable, Iterable, List, Optional, Tuple, Union
import torch
import torch_npu
from torch import nn
from transformers import PretrainedConfig
from vllm.attention import AttentionMetadata
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.attention.layer import MLAAttention
from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
@@ -492,8 +492,6 @@ class TorchairDeepseekV2MLAAttention(DeepseekV2MLAAttention):
v_head_dim: int,
q_lora_rank: Optional[int],
kv_lora_rank: int,
rope_theta: float = 10000,
rope_scaling: Optional[Dict[str, Any]] = None,
max_position_embeddings: int = 8192,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
@@ -518,7 +516,6 @@ class TorchairDeepseekV2MLAAttention(DeepseekV2MLAAttention):
self.first_k_dense_replace = config.first_k_dense_replace
self.scaling = self.qk_head_dim**-0.5
self.rope_theta = rope_theta
self.max_position_embeddings = max_position_embeddings
self.prefix = prefix
@@ -592,17 +589,17 @@ class TorchairDeepseekV2MLAAttention(DeepseekV2MLAAttention):
quant_config=quant_config,
prefix=f"{prefix}.o_proj")
if rope_scaling:
rope_scaling["rope_type"] = 'deepseek_yarn'
if config.rope_parameters["rope_type"] != "default":
config.rope_parameters["rope_type"] = "deepseek_yarn"
self.rotary_emb = get_rope(qk_rope_head_dim,
rotary_dim=qk_rope_head_dim,
max_position=max_position_embeddings,
base=rope_theta,
rope_scaling=rope_scaling,
rope_parameters=config.rope_parameters,
is_neox_style=False)
if rope_scaling:
mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
scaling_factor = rope_scaling["factor"]
if config.rope_parameters["rope_type"] != "default":
mscale_all_dim = config.rope_parameters.get(
"mscale_all_dim", False)
scaling_factor = config.rope_parameters["factor"]
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
self.scaling = self.scaling * mscale * mscale
@@ -708,8 +705,6 @@ class TorchairDeepseekV2SFAAttention(DeepseekV2MLAAttention):
v_head_dim: int,
q_lora_rank: Optional[int],
kv_lora_rank: int,
rope_theta: float = 10000,
rope_scaling: Optional[Dict[str, Any]] = None,
max_position_embeddings: int = 8192,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
@@ -734,7 +729,6 @@ class TorchairDeepseekV2SFAAttention(DeepseekV2MLAAttention):
self.first_k_dense_replace = config.first_k_dense_replace
self.scaling = self.qk_head_dim**-0.5
self.rope_theta = rope_theta
self.max_position_embeddings = max_position_embeddings
self.prefix = prefix
@@ -814,17 +808,19 @@ class TorchairDeepseekV2SFAAttention(DeepseekV2MLAAttention):
return_bias=False,
)
if rope_scaling:
rope_scaling["rope_type"] = 'deepseek_yarn'
self.rotary_emb = get_rope(qk_rope_head_dim,
rotary_dim=qk_rope_head_dim,
max_position=max_position_embeddings,
base=rope_theta,
rope_scaling=rope_scaling,
is_neox_style=False)
if rope_scaling:
mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
scaling_factor = rope_scaling["factor"]
if config.rope_parameters["rope_type"] != "default":
config.rope_parameters["rope_type"] = "deepseek_yarn"
self.rotary_emb = get_rope(
qk_rope_head_dim,
rotary_dim=qk_rope_head_dim,
max_position=max_position_embeddings,
rope_parameters=config.rope_parameters,
is_neox_style=False,
)
if config.rope_parameters["rope_type"] != "default":
mscale_all_dim = config.rope_parameters.get(
"mscale_all_dim", False)
scaling_factor = config.rope_parameters["factor"]
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
self.scaling = self.scaling * mscale * mscale
@@ -921,8 +917,6 @@ class TorchairDeepseekV2DecoderLayer(DeepseekV2DecoderLayer):
) -> None:
nn.Module.__init__(self)
self.hidden_size = config.hidden_size
rope_theta = getattr(config, "rope_theta", 10000)
rope_scaling = getattr(config, "rope_scaling", None)
max_position_embeddings = getattr(config, "max_position_embeddings",
8192)
# DecoderLayers are created with `make_layers` which passes the prefix
@@ -955,8 +949,6 @@ class TorchairDeepseekV2DecoderLayer(DeepseekV2DecoderLayer):
q_lora_rank=config.q_lora_rank
if hasattr(config, "q_lora_rank") else None,
kv_lora_rank=config.kv_lora_rank,
rope_theta=rope_theta,
rope_scaling=rope_scaling,
max_position_embeddings=max_position_embeddings,
cache_config=cache_config,
quant_config=quant_config,

15
vllm_ascend/torchair/models/torchair_pangu_moe.py
View File

@@ -24,7 +24,8 @@ import torch_npu
from torch import nn
from torch.nn import Parameter
from transformers import PretrainedConfig
from vllm.attention import Attention, AttentionMetadata
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (divide, get_pp_group,
@@ -539,8 +540,7 @@ class PanguProMoEAttention(nn.Module):
hidden_size: int,
num_heads: int,
num_kv_heads: int,
rope_theta: float = 10000,
rope_scaling: Optional[Dict[str, Any]] = None,
rope_parameters: Dict[str, Any],
max_position_embeddings: int = 8192,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
@@ -566,7 +566,6 @@ class PanguProMoEAttention(nn.Module):
self.q_size = self.num_heads * self.head_dim
self.kv_size = self.num_kv_heads * self.head_dim
self.scaling = self.head_dim**-0.5
self.rope_theta = rope_theta
self.max_position_embeddings = max_position_embeddings
self.qkv_proj = QKVParallelLinear(
@@ -600,8 +599,7 @@ class PanguProMoEAttention(nn.Module):
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
base=rope_theta,
rope_scaling=rope_scaling,
rope_parameters=rope_parameters,
)
self.attn = Attention(
self.num_heads,
@@ -654,8 +652,6 @@ class PanguProMoEDecoderLayer(nn.Module):
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
rope_theta = getattr(config, "rope_theta", 10000)
rope_scaling = getattr(config, "rope_scaling", None)
max_position_embeddings = getattr(config, "max_position_embeddings",
8192)
@@ -663,8 +659,7 @@ class PanguProMoEDecoderLayer(nn.Module):
hidden_size=self.hidden_size,
num_heads=config.num_attention_heads,
num_kv_heads=config.num_key_value_heads,
rope_theta=rope_theta,
rope_scaling=rope_scaling,
rope_parameters=config.rope_parameters,
max_position_embeddings=max_position_embeddings,
cache_config=cache_config,
quant_config=quant_config,

3
vllm_ascend/torchair/ops/torchair_fused_moe.py
View File

@@ -993,6 +993,7 @@ class TorchairAscendFusedMoE(FusedMoE):
tp_size=tp_size,
ep_size=ep_size,
dp_size=dp_size,
pcp_size=1,
prefix=prefix,
custom_routing_function=custom_routing_function,
scoring_func=scoring_func,
@@ -1011,6 +1012,8 @@ class TorchairAscendFusedMoE(FusedMoE):
self.moe_parallel_config = FusedMoEParallelConfig.make(
tp_size_=(tp_size if tp_size is not None else
get_tensor_model_parallel_world_size()),
# TODO: support pcp
pcp_size_=1,
dp_size_=(dp_size
if dp_size is not None else get_dp_group().world_size),
vllm_parallel_config=vllm_config.parallel_config)

2
vllm_ascend/torchair/torchair_mla.py
View File

@@ -170,7 +170,7 @@ class AscendMLATorchairMetadataBuilder:
self.block_size = vllm_config.cache_config.block_size
self.max_blocks = (vllm_config.model_config.max_model_len +
self.block_size - 1) // self.block_size
self.chunked_prefill_enabled = scheduler_config.chunked_prefill_enabled
self.chunked_prefill_enabled = scheduler_config.enable_chunked_prefill
if self.chunked_prefill_enabled:
self.chunked_prefill_workspace_size = min(
# Max sure there is enough for 8 full length request or at least

17
vllm_ascend/torchair/torchair_mtp_proposer.py
View File

@@ -1,13 +1,12 @@
import types
import numpy as np
import torch
import torch.nn as nn
import torchair
from torchair import patch_for_hcom
from vllm.config import (CUDAGraphMode, VllmConfig,
get_layers_from_vllm_config, set_current_vllm_config)
from vllm.forward_context import BatchDescriptor, get_forward_context
from vllm.forward_context import get_forward_context
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.model_loader import get_model_loader
from vllm.model_executor.model_loader.utils import \
@@ -149,7 +148,7 @@ class TorchairMtpProposer(MtpProposer):
break
def generate_token_ids(self,
valid_sampled_token_ids: list[np.ndarray],
valid_sampled_token_ids: list[list[int]],
sampling_metadata: SamplingMetadata = None,
scheduler_output: SchedulerOutput = None,
spec_decode_metadata: SpecDecodeMetadata = None,
@@ -162,7 +161,7 @@ class TorchairMtpProposer(MtpProposer):
attn_metadata = attn_metadata['model.layers.0.self_attn.attn']
next_token_ids: list[int] = []
for i, token_ids in enumerate(valid_sampled_token_ids):
if token_ids.shape[0] > 0:
if token_ids:
# Common case.
next_token_id = token_ids[-1]
else:
@@ -173,7 +172,7 @@ class TorchairMtpProposer(MtpProposer):
seq_len = (req_state.num_computed_tokens +
scheduler_output.num_scheduled_tokens[req_id])
next_token_id = req_state.get_token_id(seq_len)
next_token_ids.append(next_token_id.item())
next_token_ids.append(next_token_id)
next_token_ids = torch.tensor(next_token_ids,
dtype=torch.int32,
device=self.device)
@@ -189,7 +188,7 @@ class TorchairMtpProposer(MtpProposer):
# TODO(woosuk): Refactor this.
num_draft_tokens = spec_decode_metadata.num_draft_tokens
num_rejected_tokens = [
n + 1 - valid_sampled_token_ids[i].shape[0] if n > 0 else 0
n + 1 - len(valid_sampled_token_ids[i]) if n > 0 else 0
for i, n in enumerate(num_draft_tokens)
]
num_rejected_tokens = torch.tensor(
@@ -343,12 +342,7 @@ class TorchairMtpProposer(MtpProposer):
# torchair mode can reuse self.runner.num_tokens_across_dp
num_tokens_across_dp = self.runner.num_tokens_across_dp
with_prefill = self.runner.with_prefill
moe_comm_type = self.runner._select_moe_comm_method(num_input_tokens)
batch_descriptor = BatchDescriptor(num_tokens=num_input_tokens,
uniform_decode=False)
aclgraph_runtime_mode, batch_descriptor = \
self.runner.aclgraph_dispatcher.dispatch(batch_descriptor)
for step in range(self.num_speculative_tokens):
with set_ascend_forward_context(
@@ -359,7 +353,6 @@ class TorchairMtpProposer(MtpProposer):
num_tokens_across_dp=num_tokens_across_dp,
reserved_mc2_mask=self.runner.reserved_mc2_mask,
moe_comm_type=moe_comm_type,
aclgraph_runtime_mode=aclgraph_runtime_mode,
in_profile_run=self.runner.in_profile_run,
num_actual_tokens=num_tokens):
with ProfileExecuteDuration().capture_async('mtp_forward'):

2
vllm_ascend/torchair/torchair_sfa.py
View File

@@ -171,7 +171,7 @@ class AscendSFATorchairMetadataBuilder:
self.block_size = vllm_config.cache_config.block_size
self.max_blocks = (vllm_config.model_config.max_model_len +
self.block_size - 1) // self.block_size
self.chunked_prefill_enabled = scheduler_config.chunked_prefill_enabled
self.chunked_prefill_enabled = scheduler_config.enable_chunked_prefill
if self.chunked_prefill_enabled:
self.chunked_prefill_workspace_size = min(
# Max sure there is enough for 8 full length request or at least