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

Fix some ci issue and refactor modelrunner (#2445)

Browse Source 
### What this PR does / why we need it?
Fix some ci issue and refactor modelrunner

### Does this PR introduce _any_ user-facing change?
N/A

### How was this patch tested?
CI passed with existing test.

- vLLM version: v0.10.0
- vLLM main:
4d9c61993a

---------

Signed-off-by: wangli <wangli858794774@gmail.com>
Signed-off-by: MengqingCao <cmq0113@163.com>
Signed-off-by: weiguihua2 <weiguihua2@huawei.com>
Co-authored-by: wangli <wangli858794774@gmail.com>
Co-authored-by: weiguihua2 <weiguihua2@huawei.com>
This commit is contained in:
Mengqing Cao
2025-08-20 09:01:04 +08:00
committed by GitHub
parent 955411611c
commit 1327f9be1c
28 changed files with 1612 additions and 1020 deletions
Download Patch File Download Diff File Expand all files Collapse all files

42
vllm_ascend/worker/mtp_proposer_v1.py
View File

@@ -16,7 +16,9 @@ from vllm.v1.sample.metadata import SamplingMetadata
from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.ascend_forward_context import set_ascend_forward_context
from vllm_ascend.attention.utils import AscendCommonAttentionMetadata
from vllm_ascend.models.deepseek_mtp import CustomDeepSeekMTP
from vllm_ascend.torchair.utils import TorchairCommonAttentionMetadata
from vllm_ascend.utils import ProfileExecuteDuration
@@ -88,7 +90,7 @@ class MtpProposer:
# FIXME(woosuk): Avoid synchronization.
num_tokens = cu_num_tokens[-1].item()
token_indices = torch.empty(
token_indices = torch.zeros(
num_tokens,
dtype=torch.int32,
device=cu_num_tokens.device,
@@ -136,9 +138,6 @@ class MtpProposer:
# E.g., [b1, b2, c1, c2, c3, c3] -> [a2, b2, b3, c2, c3, c4]
if token_indices is not None and self.runner.torchair_graph_enabled:
last_token_indices = token_indices
else:
seq_lens = target_positions[last_token_indices] + 1
seq_lens = seq_lens.cpu()
self.input_ids[last_token_indices] = next_token_ids
@@ -155,23 +154,36 @@ class MtpProposer:
# input_batch=self.runner.input_batch,
# scheduler_output=self.runner.scheduler_output,
# )
extra_builder_kwargs = {}
is_running_torchair = self.runner.torchair_graph_enabled and \
not self.runner.with_prefill
if is_running_torchair:
extra_builder_kwargs['graph_pad_size'] = self.runner.graph_pad_size
num_input_tokens = self.runner.graph_pad_size
else:
num_input_tokens = num_tokens
attn_metadata = self.runner.attn_metadata_builder.build(
seq_lens = target_positions[last_token_indices] + 1
seq_lens = seq_lens.int()
common_attn_metadata = AscendCommonAttentionMetadata(
query_start_loc=cu_num_tokens[:batch_size + 1],
query_start_loc_cpu=cu_num_tokens[:batch_size + 1].cpu(),
seq_lens_cpu=seq_lens.cpu(),
num_reqs=batch_size,
num_actual_tokens=num_tokens,
max_query_len=max_query_len,
query_start_loc=cu_num_tokens,
**extra_builder_kwargs)
actual_seq_lengths_q=self.runner.actual_seq_lengths_q,
block_table_tensor=self.runner.input_batch.block_table[0].
get_device_tensor(),
slot_mapping_cpu=target_slot_mapping,
positions=target_positions,
attn_mask=self.runner.attn_mask,
spec_attn_mask=self.runner.spec_attn_mask,
attn_state=self.runner.attn_state,
graph_pad_size=self.runner.graph_pad_size,
decode_token_per_req=self.runner.decode_token_per_req,
)
attn_metadata = self.runner.attn_metadata_builder.build(
common_attn_metadata, self.runner.get_model())
self.positions[:num_tokens] = target_positions
self.hidden_states[:num_tokens] = target_hidden_states
@@ -281,8 +293,16 @@ class MtpProposer:
if skip_attn:
attn_metadata = None
else:
common_attn_metadata = TorchairCommonAttentionMetadata(
num_reqs=num_reqs,
num_actual_tokens=1,
actual_seq_lengths_q=self.runner.actual_seq_lengths_q,
attn_mask=self.runner.attn_mask,
spec_attn_mask=self.runner.spec_attn_mask,
decode_token_per_req=self.runner.decode_token_per_req,
)
attn_metadata = self.runner.attn_metadata_builder.build_torchair_graph_dummy(
num_reqs=num_reqs, num_actual_tokens=1)
common_attn_metadata)
input_ids = self.input_ids[:num_tokens]
positions = self.positions[:num_tokens]