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

[Refactor] move the metadata from attention_v1 to util(ready for extract common_cp) & realize Ascendmetadata inherit from the parent class. (#5203)

Browse Source 
RFC: https://github.com/vllm-project/vllm-ascend/issues/4629

1. Remove the pcp-related code from attention_v1.
2. Establish the inheritance relationship of CommonAttentionMetadata.

TODO
1. extract common_cp
2. move cp metadata to common_cp.
3. remove commonAttentionMetadata for aclgraph.

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

---------

Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Co-authored-by: weijinqian_v1 <weijinqian@huawei.com>
This commit is contained in:
weijinqian0
2025-12-23 00:10:52 +08:00
committed by GitHub
parent 3d9954eff0
commit 95e8a52156
9 changed files with 83 additions and 106 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
vllm_ascend/attention/attention_v1.py
View File

@@ -34,7 +34,9 @@ from vllm.v1.core.sched.output import SchedulerOutput
from vllm.v1.kv_cache_interface import AttentionSpec
from vllm_ascend.attention.utils import (AscendCommonAttentionMetadata,
enable_cp, split_decodes_and_prefills,
AscendMetadataForDecode,
AscendMetadataForPrefill, enable_cp,
split_decodes_and_prefills,
using_paged_attention)
from vllm_ascend.compilation.acl_graph import (get_graph_params,
update_graph_params_workspaces)
@@ -118,51 +120,6 @@ class AscendAttentionState(Enum):
SpecDecoding = 4
@dataclass
class AscendMetadataForPrefill:
@dataclass
class AscendPCPMetadata:
q_head_idx: torch.Tensor = None
q_tail_idx: torch.Tensor = None
kv_with_q_head_nomask_idx: torch.Tensor = None
kv_with_q_head_mask_idx: torch.Tensor = None
kv_with_q_tail_nomask_idx: torch.Tensor = None
kv_with_q_tail_mask_idx: torch.Tensor = None
attn_mask_seqlens: torch.Tensor = None
head_attn_nomask_seqlens: torch.Tensor = None
tail_attn_nomask_seqlens: torch.Tensor = None
q_full_idx: torch.Tensor = None
pcp_prefill_mask: torch.Tensor = None
@dataclass
class ChunkedContextMetadata:
actual_chunk_seq_lengths: torch.Tensor
actual_seq_lengths_kv: torch.Tensor
starts: torch.Tensor
chunk_seq_mask_filtered_indices: torch.Tensor
chunked_req_mask: Optional[list[bool]] = None
local_context_lens_allranks: Optional[list[list[int]]] = None
cp_kv_recover_idx_for_chunk: Optional[list[int]] = None
kv_inverse_idx_for_chunk: Optional[list[int]] = None
batch_chunk_seq_mask: Optional[list[bool]] = None
""" Prefill Specific Metadata for Ascend"""
pcp_metadata: Optional[AscendPCPMetadata] = None
pcp_allgather_restore_idx: Optional[List[int]] = None
chunked_context: Optional[ChunkedContextMetadata] = None
block_tables: torch.Tensor = None
actual_seq_lengths_q: torch.Tensor = None
@dataclass
class AscendMetadataForDecode:
""" Decode Specific Metadata for Ascend"""
num_computed_tokens_of_pcp_dcp: Optional[list[list[list[int]]]] = None
batch_seq_mask: torch.Tensor = None
block_tables: torch.Tensor = None
@dataclass
class AscendMetadata:
# **************************** Basic Properties ************************** #
@@ -274,14 +231,7 @@ class AscendAttentionMetadataBuilder:
block_table = common_attn_metadata.block_table_tensor
seq_lens = common_attn_metadata.seq_lens_cpu[:num_reqs]
long_seq_metadata = common_attn_metadata.prefill_context_parallel_metadata
num_actual_tokens_pcp_padded = long_seq_metadata.num_actual_tokens_pcp_padded if long_seq_metadata else None
if num_actual_tokens_pcp_padded is None:
num_actual_tokens_pcp_padded = num_actual_tokens
slot_mapping = common_attn_metadata.slot_mapping[:
num_actual_tokens_pcp_padded]
slot_mapping = common_attn_metadata.slot_mapping[:num_actual_tokens]
attn_mask = common_attn_metadata.attn_mask
attn_state = common_attn_metadata.attn_state
@@ -292,7 +242,6 @@ class AscendAttentionMetadataBuilder:
attn_metadata = AscendMetadata(
num_actual_tokens=num_actual_tokens,
num_decode_tokens=num_decode_tokens,
num_actual_tokens_pcp_padded=num_actual_tokens_pcp_padded,
block_tables=block_table,
query_start_loc=query_start_loc,
seq_lens=seq_lens,