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

[Lint]Style: Convert vllm-ascend/ to ruff format(new Batch #8) (#6604)

Browse Source 
### What this PR does / why we need it?
**Scope of Changes**:
| File Path |
| :--- |
| vllm_ascend/ops/\_\_init\_\_.py |
| vllm_ascend/ops/activation.py |
| vllm_ascend/ops/flashcomm2_oshard_manager.py |
| vllm_ascend/ops/layernorm.py |
| vllm_ascend/ops/mla.py |
| vllm_ascend/ops/mm_encoder_attention.py |
| vllm_ascend/ops/register_custom_ops.py |
| vllm_ascend/ops/vocab_parallel_embedding.py |
| vllm_ascend/ops/weight_prefetch.py |
| vllm_ascend/spec_decode/\_\_init\_\_.py |
| vllm_ascend/spec_decode/eagle_proposer.py |
| vllm_ascend/spec_decode/interface.py |
| vllm_ascend/spec_decode/mtp_proposer.py |
| vllm_ascend/spec_decode/ngram_proposer.py |
| vllm_ascend/spec_decode/suffix_proposer.py |

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.15.0
- vLLM main:
d7e17aaacd

Signed-off-by: MrZ20 <2609716663@qq.com>
This commit is contained in:
SILONG ZENG
2026-02-07 09:16:07 +08:00
committed by GitHub
parent c63b7a1188
commit 06aa6036f6
17 changed files with 947 additions and 1148 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
vllm_ascend/ops/mm_encoder_attention.py
View File

@@ -19,18 +19,15 @@ import einops
import torch
import torch.nn.functional as F
import torch_npu
from vllm.config import MultiModalConfig
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention # type: ignore
import vllm_ascend.envs as envs_ascend
MIN_PAD_SIZE = 64 # min_size to pad weight
MAX_PAD_SIZE = 128 # max_size to pad weight
class AscendMMEncoderAttention(MMEncoderAttention):
def __init__(
self,
num_heads: int,
@@ -82,13 +79,12 @@ class AscendMMEncoderAttention(MMEncoderAttention):
return query, key, value
def forward_oot(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
cu_seqlens: torch.Tensor | None = None,
max_seqlen: torch.Tensor
| None = None, # Only used for Flash Attention
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
cu_seqlens: torch.Tensor | None = None,
max_seqlen: torch.Tensor | None = None, # Only used for Flash Attention
):
bsz, q_len = query.size()[:2]
kv_len = key.size(1)
@@ -97,9 +93,7 @@ class AscendMMEncoderAttention(MMEncoderAttention):
# q, k, v: [b, s, head, head_dim] -> [b * s, head, head_dim]
q, k, v = self.reshape_qkv_to_3d(query, key, value, bsz, q_len, kv_len)
enable_pad = (envs_ascend.USE_OPTIMIZED_MODEL
and self.head_size > MIN_PAD_SIZE
and self.head_size < MAX_PAD_SIZE)
enable_pad = envs_ascend.USE_OPTIMIZED_MODEL and self.head_size > MIN_PAD_SIZE and self.head_size < MAX_PAD_SIZE
if enable_pad:
origin_shape = q.shape[-1]
@@ -114,10 +108,7 @@ class AscendMMEncoderAttention(MMEncoderAttention):
context_layer = torch.empty_like(q)
if cu_seqlens is None:
cu_seqlens = torch.arange(0, (bsz + 1) * q_len,
step=q_len,
dtype=torch.int32,
device=query.device)
cu_seqlens = torch.arange(0, (bsz + 1) * q_len, step=q_len, dtype=torch.int32, device=query.device)
cu_seqlens = torch.diff(cu_seqlens).to("cpu")
@@ -137,11 +128,7 @@ class AscendMMEncoderAttention(MMEncoderAttention):
context_layer = context_layer[..., :origin_shape]
if is_reshaped:
context_layer = einops.rearrange(context_layer,
"(b s) h d -> b s h d",
b=bsz).contiguous()
context_layer = einops.rearrange(context_layer, "(b s) h d -> b s h d", b=bsz).contiguous()
else:
context_layer = einops.rearrange(context_layer,
"(b s) h d -> b s (h d)",
b=bsz).contiguous()
context_layer = einops.rearrange(context_layer, "(b s) h d -> b s (h d)", b=bsz).contiguous()
return context_layer