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

Revert PTA upgrade PR (#3352)

Browse Source 
we notice that torch npu 0919 doesn't work. This PR revert related
change which rely on 0919 version.
Revert PR: #3295  #3205  #3102 

Related: #3353

- vLLM version: v0.11.0
This commit is contained in:
wangxiyuan
2025-10-10 14:09:53 +08:00
committed by GitHub
parent 601a37aeff
commit ba19dd3183
15 changed files with 57 additions and 312 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/ascend_forward_context.py
View File

@@ -156,14 +156,12 @@ def set_ascend_forward_context(
# Once the necessary conditions are met, support for MOE models will also be added.
from vllm_ascend.quantization.quant_config import AscendQuantConfig
addrmsnorm_quant_fusion_enabled = isinstance(vllm_config.quant_config, AscendQuantConfig) and \
vllm_config.model_config.hf_config.model_type in ["llama", "qwen2", "qwen3", "qwen3_moe"] and \
vllm_config.model_config.hf_config.model_type in ["llama", "qwen2", "qwen3"] and \
forward_context.layer_idx is not None
if addrmsnorm_quant_fusion_enabled:
forward_context.model_instance = model_instance
forward_context.num_hidden_layers = vllm_config.model_config.hf_config.num_hidden_layers
forward_context.fusion_linear = "gate_up_dense" if forward_context.layer_idx == 0 else "qkv_dense"
if vllm_config.model_config.hf_config.model_type == "qwen3_moe":
forward_context.fusion_linear = "gate_moe" if forward_context.layer_idx == 0 else "qkv_moe"
forward_context.addrmsnorm_quant_fusion_enabled = addrmsnorm_quant_fusion_enabled
if num_tokens is None and attn_metadata is not None:

24
vllm_ascend/attention/attention_v1.py
View File

@@ -34,8 +34,7 @@ from vllm.v1.kv_cache_interface import AttentionSpec
from vllm_ascend.attention.utils import (AscendCommonAttentionMetadata,
maybe_save_kv_layer_to_connector,
wait_for_kv_layer_from_connector)
from vllm_ascend.compilation.acl_graph import (get_graph_params,
update_graph_params_workspaces)
from vllm_ascend.compilation.acl_graph import get_graph_params
from vllm_ascend.ops.attention import vanilla_chunked_prefill
from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_NZ, aligned_16, is_310p,
nd_to_nz_2d, nd_to_nz_spec)
@@ -394,28 +393,13 @@ class AscendAttentionBackendImpl(AttentionImpl):
forward_context: ForwardContext = get_forward_context()
num_tokens = query.shape[0]
if forward_context.capturing:
# Get workspace from cache or calculate it if not present.
workspace = graph_params.workspaces.get(num_tokens)
if workspace is None:
workspace = torch_npu._npu_paged_attention_get_workspace(
query=query,
key_cache=self.key_cache,
value_cache=self.value_cache,
num_kv_heads=self.num_kv_heads,
num_heads=self.num_heads,
scale_value=self.scale,
block_table=attn_metadata.block_tables,
context_lens=attn_metadata.seq_lens,
out=output)
update_graph_params_workspaces(num_tokens, workspace)
# Handle graph capturing mode
stream = torch_npu.npu.current_stream()
event = torch.npu.ExternalEvent()
event.wait(stream)
event.reset(stream)
graph_params.events[num_tokens].append(event)
graph_params.attn_params[num_tokens].append((
query,
self.key_cache,
@@ -429,7 +413,6 @@ class AscendAttentionBackendImpl(AttentionImpl):
))
torch.npu.graph_task_group_begin(stream)
torch_npu._npu_paged_attention(
query=query,
key_cache=self.key_cache,
@@ -439,8 +422,7 @@ class AscendAttentionBackendImpl(AttentionImpl):
scale_value=self.scale,
block_table=attn_metadata.block_tables,
context_lens=attn_metadata.seq_lens,
out=output,
workspace=workspace)
out=output)
handle = torch.npu.graph_task_group_end(stream)
graph_params.handles[num_tokens].append(handle)
else:

26
vllm_ascend/compilation/acl_graph.py
View File

@@ -215,17 +215,15 @@ def update_attn_params(update_stream, forward_context, runtime_shape):
with torch.npu.stream(update_stream):
torch.npu.graph_task_update_begin(update_stream, handle)
torch_npu._npu_paged_attention(
query=query,
key_cache=key_cache,
value_cache=value_cache,
num_kv_heads=num_kv_heads,
num_heads=num_heads,
scale_value=scale,
block_table=block_table,
context_lens=seq_lens,
out=output,
workspace=graph_params.workspaces.get(runtime_shape))
torch_npu._npu_paged_attention(query=query,
key_cache=key_cache,
value_cache=value_cache,
num_kv_heads=num_kv_heads,
num_heads=num_heads,
scale_value=scale,
block_table=block_table,
context_lens=seq_lens,
out=output)
torch.npu.graph_task_update_end(update_stream)
event.record(update_stream)
@@ -258,11 +256,5 @@ def set_graph_params(aclgraph_capture_sizes: set[int]):
)
def update_graph_params_workspaces(num_tokens: int, workspace: int):
global _graph_params
if _graph_params is not None:
_graph_params.workspaces[num_tokens] = workspace
def get_graph_params():
return _graph_params

61
vllm_ascend/ops/layernorm.py
View File

@@ -15,10 +15,9 @@
# This file is a part of the vllm-ascend project.
#
from typing import Optional, Tuple, Union
from typing import Optional, Tuple, Union, cast
import torch
from vllm.config import get_current_vllm_config
from vllm.forward_context import get_forward_context
from vllm.model_executor.layers.layernorm import GemmaRMSNorm, RMSNorm
@@ -28,7 +27,6 @@ def _addrmsnorm_forward_oot(
x: torch.Tensor,
residual: torch.Tensor,
layer: Optional[torch.nn.Module] = None,
bias: Optional[torch.nn.Parameter] = None,
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
import torch_npu
@@ -41,7 +39,6 @@ def _addrmsnorm_forward_oot(
self.weight,
layer.aclnn_input_scale,
layer.aclnn_input_offset,
beta=bias,
epsilon=self.variance_epsilon)
else:
if is_310p():
@@ -53,31 +50,12 @@ def _addrmsnorm_forward_oot(
else:
x, _, residual = torch_npu.npu_add_rms_norm(
x, residual, self.weight, self.variance_epsilon)
if bias is not None:
x.add_(bias)
torch.ops.vllm.maybe_wait_prefetch_done(x)
return x, residual
class AscendRMSNorm(RMSNorm):
def __init__(
self,
hidden_size: int,
eps: float = 1e-6,
var_hidden_size: Optional[int] = None,
has_weight: bool = True,
dtype: Optional[torch.dtype] = None,
) -> None:
super().__init__(hidden_size, eps, var_hidden_size, has_weight, dtype)
vllm_config = get_current_vllm_config()
self.bias = None
# quantization with anti_method m4 will generate none-zero norm bias
if vllm_config is not None and vllm_config.quant_config is not None and \
any("norm.bias" in name for name in vllm_config.quant_config.quant_description.keys()):
self.bias = torch.nn.Parameter(torch.zeros(hidden_size),
requires_grad=False)
def forward_oot(
self,
x: torch.Tensor,
@@ -89,13 +67,10 @@ class AscendRMSNorm(RMSNorm):
residual = torch.ops.vllm.maybe_chunk_residual(x, residual)
assert x.size(0) == residual.size(0)
x, residual = _addrmsnorm_forward_oot(
self, x, residual, self.next_need_quant_fusion_linear,
self.bias)
self, x, residual, self.next_need_quant_fusion_linear)
return x, residual
x, residual = torch_npu.npu_rms_norm(x, self.weight,
self.variance_epsilon)
if self.bias is not None:
x.add_(self.bias)
return x
@property
@@ -125,13 +100,6 @@ class AscendRMSNorm(RMSNorm):
# does not need to be repeated
if not forward_context.prefetch_mlp_enabled:
forward_context.layer_idx += 1
elif fusion_linear == "qkv_moe":
next_linear = model_instance.model.layers[
layer_idx].self_attn.qkv_proj
forward_context.fusion_linear = "gate_moe"
elif fusion_linear == "gate_moe":
forward_context.fusion_linear = "qkv_moe"
forward_context.layer_idx += 1
from vllm_ascend.quantization.w8a8 import AscendW8A8LinearMethod
if next_linear is not None and \
not isinstance(next_linear.quant_method.quant_method, AscendW8A8LinearMethod):
@@ -139,6 +107,31 @@ class AscendRMSNorm(RMSNorm):
return next_linear
class AscendQuantRMSNorm(AscendRMSNorm):
def __init__(
self,
hidden_size: int,
eps: float = 1e-6,
var_hidden_size: Optional[int] = None,
has_weight: bool = True,
dtype: Optional[torch.dtype] = None,
) -> None:
super().__init__(hidden_size, eps, var_hidden_size, has_weight, dtype)
self.bias = torch.nn.Parameter(torch.zeros(hidden_size),
requires_grad=False)
def forward_oot(
self,
x: torch.Tensor,
residual: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
if residual is not None:
x, residual = super().forward_oot(x, residual)
return x.add_(self.bias), residual
return cast(torch.Tensor, super().forward_oot(x)).add_(self.bias)
class AscendGemmaRMSNorm(GemmaRMSNorm):
def forward_oot(

8
vllm_ascend/utils.py
View File

@@ -501,7 +501,8 @@ def register_ascend_customop(vllm_config: Optional[VllmConfig] = None):
from vllm_ascend.ops.activation import AscendQuickGELU, AscendSiluAndMul
from vllm_ascend.ops.common_fused_moe import (AscendFusedMoE,
AscendSharedFusedMoE)
from vllm_ascend.ops.layernorm import AscendGemmaRMSNorm, AscendRMSNorm
from vllm_ascend.ops.layernorm import (AscendGemmaRMSNorm,
AscendQuantRMSNorm, AscendRMSNorm)
from vllm_ascend.ops.linear import (AscendColumnParallelLinear,
AscendMergedColumnParallelLinear,
AscendQKVParallelLinear,
@@ -532,6 +533,11 @@ def register_ascend_customop(vllm_config: Optional[VllmConfig] = None):
"MultiHeadLatentAttention": AscendMultiHeadLatentAttention,
}
if vllm_config is not None and \
vllm_config.quant_config is not None and \
any("norm.bias" in name for name in vllm_config.quant_config.quant_description.keys()):
REGISTERED_ASCEND_OPS["RMSNorm"] = AscendQuantRMSNorm
for name, op_cls in REGISTERED_ASCEND_OPS.items():
CustomOp.register_oot(_decorated_op_cls=op_cls, name=name)