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

Drop vLLM 0.13.0 support (#6069)

Browse Source 
### What this PR does / why we need it?
Drop vLLM 0.13.0 support, upgrade to 0.14.0

- vLLM version: v0.13.0
- vLLM main:
d68209402d

---------

Signed-off-by: hfadzxy <starmoon_zhang@163.com>
This commit is contained in:
zhangxinyuehfad
2026-01-23 09:45:08 +08:00
committed by GitHub
parent 27a513b672
commit 819a4459ce
39 changed files with 86 additions and 272 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
vllm_ascend/worker/model_runner_v1.py
View File

@@ -50,8 +50,10 @@ from vllm.sequence import IntermediateTensors
from vllm.utils.import_utils import LazyLoader
from vllm.utils.math_utils import cdiv
from vllm.utils.mem_utils import DeviceMemoryProfiler
from vllm.v1.attention.backend import AttentionBackend, AttentionType # type: ignore
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
from vllm.v1.attention.backends.utils import CommonAttentionMetadata
from vllm.v1.attention.selector import get_attn_backend # type: ignore
from vllm.v1.core.sched.output import SchedulerOutput
from vllm.v1.kv_cache_interface import (AttentionSpec,
EncoderOnlyAttentionSpec,
@@ -102,7 +104,7 @@ from vllm_ascend.spec_decode.mtp_proposer import MtpProposer
from vllm_ascend.utils import (AscendDeviceType, ProfileExecuteDuration,
enable_sp, get_ascend_device_type, is_moe_model,
lmhead_tp_enable, maybe_trans_nz,
set_weight_prefetch_method, vllm_version_is)
set_weight_prefetch_method)
from vllm_ascend.worker.npu_input_batch import NPUInputBatch
from vllm_ascend.worker.pcp_utils import PCPManager
@@ -115,15 +117,6 @@ if TYPE_CHECKING:
else:
xgr = LazyLoader("xgr", globals(), "xgrammar")
# isort: off
if vllm_version_is('0.13.0'):
from vllm.attention.backends.abstract import ( # type: ignore
AttentionBackend, AttentionType)
from vllm.attention.selector import get_attn_backend # type: ignore
else:
from vllm.v1.attention.selector import get_attn_backend # type: ignore
from vllm.v1.attention.backend import AttentionBackend, AttentionType # type: ignore
# isort: on
import torch_npu
# if true, allow tensor initialization and casting with internal format (e.g., NZ)
@@ -746,10 +739,7 @@ class NPUModelRunner(GPUModelRunner):
# _prepare_inputs may reorder the batch, so we must gather
# multi-modal outputs after that to ensure the correct order
if vllm_version_is('0.13.0'):
model_kwargs = self._init_model_kwargs(num_input_tokens)
else:
model_kwargs = self._init_model_kwargs()
model_kwargs = self._init_model_kwargs()
if self.is_multimodal_model and not self.model_config.is_encoder_decoder:
self.multimodal_cpu_fields = ["grid_thw"]
self._prepare_multimodal_fields()
@@ -1575,16 +1565,10 @@ class NPUModelRunner(GPUModelRunner):
logits = None
else:
if self.input_batch.pooling_params:
if vllm_version_is('0.13.0'):
pool_output = self._pool(
hidden_states,
scheduler_output.total_num_scheduled_tokens,
num_scheduled_tokens_np)
else:
pool_output = self._pool(
hidden_states,
scheduler_output.total_num_scheduled_tokens,
num_scheduled_tokens_np, kv_connector_output)
pool_output = self._pool(
hidden_states,
scheduler_output.total_num_scheduled_tokens,
num_scheduled_tokens_np, kv_connector_output)
if self.debugger is not None:
self.debugger.stop()
self.debugger.step()
@@ -1675,8 +1659,7 @@ class NPUModelRunner(GPUModelRunner):
attn_metadata,
aux_hidden_states,
)
if not vllm_version_is('0.13.0'):
self._copy_draft_token_ids_to_cpu(scheduler_output)
self._copy_draft_token_ids_to_cpu(scheduler_output)
(
logprobs_lists,
@@ -1826,20 +1809,12 @@ class NPUModelRunner(GPUModelRunner):
valid_sampled_token_ids[int(i)].clear()
else:
# Includes spec decode tokens.
if vllm_version_is('0.13.0'):
valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
sampled_token_ids,
self.input_batch.vocab_size,
discard_sampled_tokens_req_indices,
return_cu_num_tokens=logprobs_tensors is not None,
)
else:
valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
sampled_token_ids,
self.input_batch.vocab_size,
discard_sampled_tokens_req_indices,
logprobs_tensors=logprobs_tensors,
)
valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
sampled_token_ids,
self.input_batch.vocab_size,
discard_sampled_tokens_req_indices,
logprobs_tensors=logprobs_tensors,
)
else:
valid_sampled_token_ids = []
invalid_req_indices = discard_sampled_tokens_req_indices.tolist()

14
vllm_ascend/worker/worker.py
View File

@@ -58,16 +58,13 @@ from vllm_ascend.distributed.parallel_state import init_ascend_model_parallel
from vllm_ascend.ops.triton.triton_utils import init_device_properties_triton
from vllm_ascend.utils import (AscendDeviceType, check_ascend_device_type,
enable_sp, get_ascend_device_type,
register_ascend_customop, vllm_version_is)
register_ascend_customop)
from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
torch._dynamo.trace_rules.clear_lru_cache() # noqa: E402
from torch._dynamo.variables import TorchInGraphFunctionVariable # noqa: E402
if vllm_version_is("0.13.0"):
from vllm.model_executor.utils import set_random_seed
else:
from vllm.utils.torch_utils import set_random_seed
from vllm.utils.torch_utils import set_random_seed
torch_non_c_binding_in_graph_functions_npu = dict.fromkeys(
["torch.npu.current_stream"],
@@ -121,13 +118,6 @@ class NPUWorker(WorkerBase):
self.cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[
self.cache_config.cache_dtype]
if vllm_version_is('0.13.0'):
if self.model_config.trust_remote_code:
# note: lazy import to avoid importing torch before initializing
from vllm.utils.import_utils import init_cached_hf_modules
init_cached_hf_modules()
self.profiler = self._init_profiler()
if vllm_config.model_config and vllm_config.model_config.enable_sleep_mode:
# Buffers saved before sleep