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

[main][refactor] Refactoring forward_context and model_runner_v1 (#1979)

Browse Source 
### What this PR does / why we need it?

A refactoring of forward_context and model_runner_v1, add some context
which is necessary in model inference into forward_context, and refactor
dummy_run logic, make it more reasonable.
Some details for this PR:

Add `ascend_forward_context`;
Update mc2_v2 op, and support `active_mask` param;
Update scripts in examples dir;
refactor `dummy_run` logic;
Add soc_version for A2 and A3;

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

No change at user-facing.

### How was this patch tested?


- vLLM version: v0.10.0
- vLLM main:
57c22e57f9

Signed-off-by: zzzzwwjj <1183291235@qq.com>
This commit is contained in:
zzzzwwjj
2025-07-28 14:06:20 +08:00
committed by GitHub
parent e3a2443c3a
commit ba3dfbd59e
22 changed files with 629 additions and 347 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
vllm_ascend/attention/attention_v1.py
View File

@@ -119,6 +119,7 @@ class AscendAttentionState(Enum):
@dataclass
class AscendMetadata:
# **************************** Basic Properties ****************************
attn_mask: Optional[torch.Tensor] = None
# Current state of this attention run.
@@ -149,11 +150,6 @@ class AscendMetadata:
# (num_tokens,)
slot_mapping: torch.Tensor = None
# ************************* DP Related Properties **************************
with_prefill_across_dp: bool = False
# Maximum number of tokens across dp group
max_num_tokens_across_dp: int = 0
class AscendAttentionMetadataBuilder:
@@ -164,12 +160,7 @@ class AscendAttentionMetadataBuilder:
scheduler_output: "SchedulerOutput") -> bool:
return False
def build(self,
num_reqs,
num_actual_tokens,
max_query_len,
max_num_tokens_across_dp: int = 0,
with_prefill_across_dp: bool = False):
def build(self, num_reqs, num_actual_tokens, max_query_len):
block_table = self.runner.input_batch.block_table[0].get_device_tensor(
)
@@ -196,18 +187,15 @@ class AscendAttentionMetadataBuilder:
attn_mask = torch_npu.npu_format_cast(mask_nz.contiguous(),
ACL_FORMAT_FRACTAL_NZ)
attn_metadata = AscendMetadata(
num_actual_tokens=num_actual_tokens,
block_tables=block_table,
query_start_loc=query_start_loc,
query_lens=query_lens,
seq_lens=seq_lens,
max_query_len=max_query_len,
slot_mapping=slot_mapping,
attn_mask=attn_mask,
attn_state=attn_state,
max_num_tokens_across_dp=max_num_tokens_across_dp,
with_prefill_across_dp=with_prefill_across_dp)
attn_metadata = AscendMetadata(num_actual_tokens=num_actual_tokens,
block_tables=block_table,
query_start_loc=query_start_loc,
query_lens=query_lens,
seq_lens=seq_lens,
max_query_len=max_query_len,
slot_mapping=slot_mapping,
attn_mask=attn_mask,
attn_state=attn_state)
return attn_metadata