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

Upgrade vLLM to v0.10.0 (#1927)

Browse Source 
### What this PR does / why we need it?
- Upgrade to v0.10.0
- Drop v0.9.2 version compatibility
- Add patch for
`vllm_ascend/patch/worker/patch_common/patch_sampler_gather_logprobs.py`
as workaround of
f3a683b7c9
for v0.10.0 and also add e2e test `test_models_prompt_logprobs`
- Pin transformers<4.54.0 as workaround of
https://github.com/vllm-project/vllm-ascend/issues/2034

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
- Test locally:
`VLLM_USE_MODELSCOPE=true pytest -sv
tests/e2e/singlecard/test_offline_inference.py::test_models_prompt_logprobs`
- CI passed

- vLLM version: v0.9.2
- vLLM main:
7728dd77bb

---------

Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
This commit is contained in:
Yikun Jiang
2025-07-26 15:43:29 +08:00
committed by GitHub
parent 2f50304c19
commit 17a430f7b8
29 changed files with 198 additions and 251 deletions
Download Patch File Download Diff File Expand all files Collapse all files

103
vllm_ascend/worker/model_runner_v1.py
View File

@@ -45,8 +45,9 @@ from vllm.logger import logger
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
from vllm.model_executor.model_loader import get_model
from vllm.model_executor.models.interfaces_base import (VllmModelForPooling,
is_pooling_model)
from vllm.model_executor.models.interfaces import supports_transcription
from vllm.model_executor.models.interfaces_base import (
VllmModelForPooling, is_pooling_model, is_text_generation_model)
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import MultiModalKwargs, PlaceholderRange
from vllm.multimodal.utils import group_mm_inputs_by_modality
@@ -66,7 +67,7 @@ from vllm.v1.sample.sampler import Sampler
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
from vllm.v1.worker.utils import (gather_mm_placeholders,
from vllm.v1.worker.utils import (bind_kv_cache, gather_mm_placeholders,
sanity_check_mm_encoder_outputs,
scatter_mm_placeholders)
@@ -88,15 +89,8 @@ from vllm_ascend.worker.eagle_proposer_v1 import EagleProposer
from vllm_ascend.worker.mtp_proposer_v1 import MtpProposer
from vllm_ascend.worker.npu_input_batch import CachedRequestState, InputBatch
if vllm_version_is("0.9.2"):
from vllm.model_executor.models.interfaces import has_step_pooler
from vllm.v1.utils import bind_kv_cache
else:
from vllm.model_executor.models.interfaces import supports_transcription
from vllm.model_executor.models.interfaces_base import \
is_text_generation_model
if not vllm_version_is("0.10.0"):
from vllm.tasks import GenerationTask, SupportedTask
from vllm.v1.worker.utils import bind_kv_cache
if TYPE_CHECKING:
import xgrammar as xgr # type: ignore[import-untyped]
@@ -409,7 +403,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
else:
generator = None
if not vllm_version_is("0.9.2") and pooling_params:
if pooling_params:
assert (task := pooling_params.task) is not None, (
"You did not set `task` in the API")
model = cast(VllmModelForPooling, self.model)
@@ -585,10 +579,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
# OPTIMIZATION: Start copying the block table first.
# This way, we can overlap the copy with the following CPU operations.
if vllm_version_is("0.9.2"):
self.input_batch.block_table.commit(num_reqs)
else:
self.input_batch.block_table.commit_block_table(num_reqs)
self.input_batch.block_table.commit_block_table(num_reqs)
# Get the number of scheduled tokens for each request.
req_ids = self.input_batch.req_ids
@@ -939,10 +930,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
# OPTIMIZATION: Start copying the block table first.
# This way, we can overlap the copy with the following CPU operations.
if vllm_version_is("0.9.2"):
self.input_batch.block_table.commit(num_reqs)
else:
self.input_batch.block_table.commit_block_table(num_reqs)
self.input_batch.block_table.commit_block_table(num_reqs)
# Get the number of scheduled tokens for each request.
# TODO: The Python loop can be slow. Optimize.
@@ -1771,57 +1759,33 @@ class NPUModelRunner(LoRAModelRunnerMixin):
req_num_tokens = num_tokens // num_reqs
if vllm_version_is("0.9.2"):
dummy_metadata = PoolingMetadata(
prompt_lens=torch.tensor(
[h.shape[0] for h in hidden_states_list],
device=self.device),
prompt_token_ids=torch.zeros((num_reqs, req_num_tokens),
dtype=torch.int32,
device=self.device),
pooling_params=[PoolingParams()] * num_reqs)
try:
pooler_output = self.model.pooler(
hidden_states=hidden_states_list,
pooling_metadata=dummy_metadata)
except RuntimeError as e:
if 'out of memory' in str(e):
raise RuntimeError(
"NPU out of memory occurred when warming up pooler with "
f"{num_reqs} dummy requests. Please try lowering "
"`max_num_seqs` or `gpu_memory_utilization` when "
"initializing the engine.") from e
else:
raise e
else:
model = cast(VllmModelForPooling, self.model)
dummy_task = self.get_supported_pooling_tasks()[0]
dummy_pooling_params = PoolingParams(task=dummy_task)
model = cast(VllmModelForPooling, self.model)
dummy_task = self.get_supported_pooling_tasks()[0]
dummy_pooling_params = PoolingParams(task=dummy_task)
to_update = model.pooler.get_pooling_updates(dummy_task)
to_update.apply(dummy_pooling_params)
to_update = model.pooler.get_pooling_updates(dummy_task)
to_update.apply(dummy_pooling_params)
dummy_metadata = PoolingMetadata(
prompt_lens=torch.tensor(
[h.shape[0] for h in hidden_states_list],
device=self.device),
prompt_token_ids=torch.zeros((num_reqs, req_num_tokens),
dtype=torch.int32,
device=self.device),
pooling_params=[dummy_pooling_params] * num_reqs)
dummy_metadata = PoolingMetadata(
prompt_lens=torch.tensor([h.shape[0] for h in hidden_states_list],
device=self.device),
prompt_token_ids=torch.zeros((num_reqs, req_num_tokens),
dtype=torch.int32,
device=self.device),
pooling_params=[dummy_pooling_params] * num_reqs)
try:
pooler_output = model.pooler(hidden_states=hidden_states_list,
pooling_metadata=dummy_metadata)
except RuntimeError as e:
if 'out of memory' in str(e):
raise RuntimeError(
"NPU out of memory occurred when warming up pooler with "
f"{num_reqs} dummy requests. Please try lowering "
"`max_num_seqs` or `gpu_memory_utilization` when "
"initializing the engine.") from e
else:
raise e
try:
pooler_output = model.pooler(hidden_states=hidden_states_list,
pooling_metadata=dummy_metadata)
except RuntimeError as e:
if 'out of memory' in str(e):
raise RuntimeError(
"NPU out of memory occurred when warming up pooler with "
f"{num_reqs} dummy requests. Please try lowering "
"`max_num_seqs` or `gpu_memory_utilization` when "
"initializing the engine.") from e
else:
raise e
return pooler_output
@@ -1841,9 +1805,6 @@ class NPUModelRunner(LoRAModelRunnerMixin):
QKVParallelLinear, RowParallelLinear)):
module.weight.data = torch_npu.npu_format_cast(
module.weight.data, ACL_FORMAT_FRACTAL_NZ)
if vllm_version_is("0.9.2") and has_step_pooler(self.model):
self.input_batch.logits_processing_needs_token_ids_bool = True
if self.drafter:
logger.info("Loading drafter model...")
if isinstance(self.drafter, EagleProposer):

25
vllm_ascend/worker/npu_input_batch.py
View File

@@ -35,8 +35,6 @@ from vllm.v1.spec_decode.utils import is_spec_decode_unsupported
from vllm.v1.utils import copy_slice
from vllm.v1.worker.block_table import MultiGroupBlockTable
from vllm_ascend.utils import vllm_version_is
_SAMPLING_EPS = 1e-5
@@ -246,11 +244,8 @@ class InputBatch:
# req_index -> bad_words_token_ids
self.bad_words_token_ids: dict[int, list[list[int]]] = {}
if vllm_version_is("0.9.2"):
self.logits_processing_needs_token_ids_bool = False
else:
self.logits_processing_needs_token_ids = np.zeros(max_num_reqs,
dtype=bool)
self.logits_processing_needs_token_ids = np.zeros(max_num_reqs,
dtype=bool)
self.req_output_token_ids: list[Optional[list[int]]] = []
@@ -387,9 +382,6 @@ class InputBatch:
if sampling_params.bad_words_token_ids:
self.bad_words_token_ids[
req_index] = sampling_params.bad_words_token_ids
elif vllm_version_is("0.9.2"):
assert request.pooling_params is not None
self.pooling_params[req_id] = request.pooling_params
elif pooling_params := request.pooling_params:
self.pooling_params[req_id] = pooling_params
self.logits_processing_needs_token_ids[req_index] = (
@@ -624,15 +616,10 @@ class InputBatch:
self.presence_penalties, num_reqs)
copy_slice(self.repetition_penalties_cpu_tensor,
self.repetition_penalties, num_reqs)
if vllm_version_is("0.9.2"):
needs_prompt_token_ids = (
not self.no_penalties
or (self.num_reqs > 0
and self.logits_processing_needs_token_ids_bool))
else:
needs_prompt_token_ids = (
not self.no_penalties
or self.logits_processing_needs_token_ids[:num_reqs].any())
needs_prompt_token_ids = (
not self.no_penalties
or self.logits_processing_needs_token_ids[:num_reqs].any())
if needs_prompt_token_ids:
# The prompt tokens are used only for applying penalties or
# step pooling during the sampling/pooling process.

2
vllm_ascend/worker/worker_v1.py
View File

@@ -45,7 +45,7 @@ from vllm_ascend.utils import (sleep_mode_enabled, try_register_lib,
vllm_version_is)
from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
if not vllm_version_is("0.9.2"):
if not vllm_version_is("0.10.0"):
from vllm.tasks import SupportedTask