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

[Bugfix] fix dcp_only bug and add e2e accuracy test for dcp only and pcp only (#5565)

Browse Source 
### What this PR does / why we need it?
[Bugfix] fix dcp_only bug and add e2e accuracy test for dcp only and pcp
only
this pr fix the bug of accuracy test when decode_parallel_size>1 and
prefill_context_parallel_size=1.
### Does this PR introduce _any_ user-facing change?
NO

### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
7157596103

---------

Signed-off-by: zhenwenqi2024 <zhenwenqi_2022@qq.com>
This commit is contained in:
zhenwenqi2024
2026-01-06 22:48:21 +08:00
committed by GitHub
parent 77a029979e
commit ad9b711f89
3 changed files with 128 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
vllm_ascend/worker/model_runner_v1.py
View File

@@ -935,22 +935,21 @@ class NPUModelRunner(GPUModelRunner):
blk_table_tensor = blk_table.get_device_tensor()
slot_mapping = blk_table.slot_mapping.gpu[:
maybe_pcp_full_tokens]
if self.pcp_size * self.dcp_size == 1:
if self.pcp_size == 1:
slot_mapping[
total_num_scheduled_tokens:num_input_tokens].fill_(-1)
slot_mapping = blk_table.slot_mapping.gpu
if self.pcp_size * self.dcp_size > 1:
self.long_seq_metadata = self.pcp_manager.generate_pcp_metadata(
total_num_scheduled_tokens, self.query_lens,
self.attn_mask, self.input_batch)
blk_table.slot_mapping.gpu[maybe_pcp_full_tokens:].fill_(-1)
slot_mapping = slot_mapping[:maybe_pcp_full_tokens]
slot_mapping = self.pcp_manager.get_padded_slot_mapping(
total_num_scheduled_tokens,
slot_mapping,
)
blk_table.slot_mapping.gpu[:self.pcp_manager.
num_actual_tokens_pcp_padded] = slot_mapping
if self.pcp_size > 1:
slot_mapping = self.pcp_manager.get_padded_slot_mapping(
total_num_scheduled_tokens,
slot_mapping,
)
blk_table.slot_mapping.gpu[:self.pcp_manager.
num_actual_tokens_pcp_padded] = slot_mapping
# NOTE: This is a temporary hack, now in GPUModelRunner, this prepare_inputs
# has been split to multiple parts, and there are 3 parts that is related to this
@@ -3034,7 +3033,7 @@ def _torch_cuda_wrapper():
torch.cuda.synchronize = torch.npu.synchronize
torch.cuda.mem_get_info = torch.npu.mem_get_info
yield
except Exception:
except Exception as e:
torch.cuda.Event = _EventPlaceholder
torch.cuda.Stream = _StreamPlaceholder
torch.cuda.default_stream = _StreamPlaceholder
@@ -3042,6 +3041,7 @@ def _torch_cuda_wrapper():
torch.cuda.stream = _StreamPlaceholder
torch.cuda.synchronize = _StreamPlaceholder
torch.cuda.mem_get_info = _StreamPlaceholder
raise RuntimeError(f"NPUModelRunner init failed, error is {e}")
finally:
# if anything goes wrong, just patch it with a placeholder
torch.cuda.Event = _EventPlaceholder

4
vllm_ascend/xlite/xlite_worker.py
View File

@@ -13,6 +13,8 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from vllm.v1.worker.workspace import init_workspace_manager
from vllm_ascend.worker.worker import NPUWorker
from vllm_ascend.xlite.xlite_model_runner import XliteModelRunner
@@ -23,4 +25,6 @@ class XliteWorker(NPUWorker):
def init_device(self):
"""Override init_device to init xlite model runner"""
self.device = self._init_device()
num_ubatches = 1
init_workspace_manager(self.device, num_ubatches)
self.model_runner = XliteModelRunner(self.vllm_config, self.device)