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

[CI] fix ci (#2464)

Browse Source 
### What this PR does / why we need it?
1. use action/checkout@v5 instead of v4
2. remove dbo test case because there is issue with it and will be
refactored later
3. make vllm-ascend compatible with vllm v0.10.1.1 and add CI for it
4. fix sampler api changes introduced by
https://github.com/vllm-project/vllm/pull/22387
6. fix qwen3 moe config changes intruoduced by
https://github.com/vllm-project/vllm/pull/20562
7. fix kvcache block changes introduced by
https://github.com/vllm-project/vllm/pull/23262

### 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:
0c6e40bbaa

---------

Signed-off-by: MengqingCao <cmq0113@163.com>
This commit is contained in:
Mengqing Cao
2025-08-22 07:30:48 +08:00
committed by GitHub
parent 0ca3f48c90
commit b0403f8d8a
27 changed files with 389 additions and 199 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
vllm_ascend/core/scheduler.py
View File

@@ -31,6 +31,13 @@ from vllm.v1.outputs import ModelRunnerOutput
from vllm.v1.request import Request, RequestStatus
from vllm.v1.structured_output import StructuredOutputManager
from vllm_ascend.utils import vllm_version_is
if vllm_version_is("0.10.1.1"):
from vllm.v1.core.kv_cache_manager import KVCacheBlocks
else:
KVCacheBlocks = None
class AscendScheduler(Scheduler):
"""This Scheduler extends vllm's original v1 scheduler
@@ -59,7 +66,10 @@ class AscendScheduler(Scheduler):
scheduled_running_reqs: list[Request] = []
preempted_reqs: list[Request] = []
req_to_new_block_ids: dict[str, list[int]] = {}
if vllm_version_is("0.10.1.1"):
req_to_new_block_ids: dict[str, list[int]] = {}
else:
req_to_new_blocks: dict[str, KVCacheBlocks] = {}
num_scheduled_tokens: dict[str, int] = {}
token_budget = self.max_num_scheduled_tokens
# Spec decode-related.
@@ -217,8 +227,11 @@ class AscendScheduler(Scheduler):
if self.lora_config and request.lora_request:
scheduled_loras.add(request.lora_request.lora_int_id)
req_to_new_block_ids[request.request_id] = (
self.kv_cache_manager.get_block_ids(request.request_id))
if vllm_version_is("0.10.1.1"):
req_to_new_block_ids[request.request_id] = (
self.kv_cache_manager.get_block_ids(request.request_id))
else:
req_to_new_blocks[request.request_id] = new_blocks
# Update request info.
num_scheduled_tokens[request.request_id] = num_new_tokens
token_budget -= num_new_tokens
@@ -307,8 +320,11 @@ class AscendScheduler(Scheduler):
# Schedule the request.
scheduled_running_reqs.append(request)
self.scheduled_req_ids.add(request.request_id)
req_to_new_block_ids[request.request_id] = (
new_blocks.get_block_ids())
if vllm_version_is("0.10.1.1"):
req_to_new_block_ids[request.request_id] = (
new_blocks.get_block_ids())
else:
req_to_new_blocks[request.request_id] = new_blocks
num_scheduled_tokens[request.request_id] = num_new_tokens
token_budget -= num_new_tokens
req_index += 1
@@ -346,16 +362,27 @@ class AscendScheduler(Scheduler):
any_request, len(self.running)))
# Construct the scheduler output.
new_reqs_data = [
NewRequestData.from_request(req,
req_to_new_block_ids[req.request_id])
for req in scheduled_new_reqs
]
if vllm_version_is("0.10.1.1"):
new_reqs_data = [
NewRequestData.from_request(
req, req_to_new_block_ids[req.request_id])
for req in scheduled_new_reqs
]
cached_reqs_data = self._make_cached_request_data(
scheduled_running_reqs, scheduled_resumed_reqs,
num_scheduled_tokens, scheduled_spec_decode_tokens,
req_to_new_block_ids)
else:
new_reqs_data = [
NewRequestData.from_request(
req, req_to_new_blocks[req.request_id].get_block_ids())
for req in scheduled_new_reqs
]
cached_reqs_data = self._make_cached_request_data(
scheduled_running_reqs, scheduled_resumed_reqs,
num_scheduled_tokens, scheduled_spec_decode_tokens,
req_to_new_block_ids)
cached_reqs_data = self._make_cached_request_data(
scheduled_running_reqs, scheduled_resumed_reqs,
num_scheduled_tokens, scheduled_spec_decode_tokens,
req_to_new_blocks)
scheduled_cached_reqs = cached_reqs_data
scheduler_output = SchedulerOutput(

7
vllm_ascend/models/qwen3_moe.py
View File

@@ -50,6 +50,7 @@ from vllm.sequence import IntermediateTensors
from vllm_ascend.ops.fused_moe import AscendFusedMoE
from vllm_ascend.ops.sequence_parallel import (MetadataForPadding,
init_metadata_for_sp)
from vllm_ascend.utils import vllm_version_is
class CustomSparseMoeBlock(Qwen3MoeSparseMoeBlock):
@@ -253,7 +254,11 @@ class CustomQwen3MoeModel(Qwen3MoeModel):
quant_config = vllm_config.quant_config
parallel_config = vllm_config.parallel_config
self.num_redundant_experts = parallel_config.num_redundant_experts
if vllm_version_is("0.10.1.1"):
self.num_redundant_experts = parallel_config.num_redundant_experts
else:
eplb_config = parallel_config.eplb_config
self.num_redundant_experts = eplb_config.num_redundant_experts
self.padding_idx = config.pad_token_id
self.vocab_size = config.vocab_size
self.config = config

27
vllm_ascend/sample/sampler.py
View File

@@ -3,12 +3,19 @@ import torch_npu
from vllm.v1.sample.ops.topk_topp_sampler import TopKTopPSampler, random_sample
from vllm.v1.sample.sampler import Sampler
from vllm_ascend.utils import is_310p
from vllm_ascend.utils import is_310p, vllm_version_is
if not vllm_version_is("0.10.1.1"):
from vllm.config import LogprobsMode
DEFAULT_LOGPROBS_MODE = LogprobsMode.RAW_LOGPROBS
else:
LogprobsMode = None
DEFAULT_LOGPROBS_MODE = "raw_logprobs"
class AscendSampler(Sampler):
def __init__(self, logprobs_mode="raw_logprobs"):
def __init__(self, logprobs_mode=DEFAULT_LOGPROBS_MODE):
# TODO: support logprobs_mode in vllm-ascend
super().__init__(logprobs_mode=logprobs_mode)
self.topk_topp_sampler = AscendTopKTopPSampler()
@@ -61,5 +68,19 @@ class AscendTopKTopPSampler(TopKTopPSampler):
def forward_native(self, logits, generators, k, p):
"""Override pytorch native implementation to torch_npu"""
logits = self._apply_top_k_top_p(logits, k, p)
if not vllm_version_is("0.10.1.1"):
logits_to_return = None
if self.logprobs_mode == LogprobsMode.PROCESSED_LOGITS:
logits_to_return = logits
elif self.logprobs_mode == LogprobsMode.PROCESSED_LOGPROBS:
logits_to_return = logits.log_softmax(dim=-1,
dtype=torch.float32)
probs = logits.softmax(dim=-1, dtype=torch.float32)
return random_sample(probs, generators)
output = None
if vllm_version_is("0.10.1.1"):
output = random_sample(probs, generators)
else:
output = (random_sample(probs, generators), logits_to_return)
return output

88
vllm_ascend/worker/model_runner_v1.py
View File

@@ -64,8 +64,8 @@ from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
KVCacheSpec)
from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, DraftTokenIds,
LogprobsTensors, ModelRunnerOutput)
from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, LogprobsTensors,
ModelRunnerOutput)
from vllm.v1.pool.metadata import PoolingMetadata
from vllm.v1.sample.logits_processor import build_logitsprocs
from vllm.v1.sample.metadata import SamplingMetadata
@@ -95,11 +95,17 @@ from vllm_ascend.torchair.torchair_attention import AscendTorchairMetadata
from vllm_ascend.torchair.torchair_mla import AscendMLATorchairMetadata
from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_ND, ACL_FORMAT_FRACTAL_NZ,
ProfileExecuteDuration, is_310p,
maybe_converting_weight_acl_format)
maybe_converting_weight_acl_format,
vllm_version_is)
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 not vllm_version_is("0.10.1.1"):
from vllm.v1.outputs import DraftTokenIds
else:
DraftTokenIds = None
if TYPE_CHECKING:
import xgrammar as xgr # type: ignore[import-untyped]
from vllm.v1.core.sched.output import SchedulerOutput
@@ -514,11 +520,13 @@ class NPUModelRunner(LoRAModelRunnerMixin):
# Update the block IDs.
if not resumed_from_preemption:
# Append the new blocks to the existing block IDs.
for block_ids, new_ids in zip(req_state.block_ids,
new_block_ids):
block_ids.extend(new_ids)
if new_block_ids is not None:
# Append the new blocks to the existing block IDs.
for block_ids, new_ids in zip(req_state.block_ids,
new_block_ids):
block_ids.extend(new_ids)
else:
assert new_block_ids is not None
# The request is resumed from preemption.
# Replace the existing block IDs with the new ones.
req_state.block_ids = new_block_ids
@@ -534,7 +542,9 @@ class NPUModelRunner(LoRAModelRunnerMixin):
# Update the persistent batch.
self.input_batch.num_computed_tokens_cpu[req_index] = (
num_computed_tokens)
self.input_batch.block_table.append_row(new_block_ids, req_index)
if new_block_ids is not None:
self.input_batch.block_table.append_row(
new_block_ids, req_index)
# For the last rank, we don't need to update the token_ids_cpu
# because the sampled tokens are already cached.
@@ -1526,16 +1536,28 @@ class NPUModelRunner(LoRAModelRunnerMixin):
else:
pooler_output.append(None)
extra_args = ({"kv_connector_output": kv_connector_output})
return ModelRunnerOutput(
req_ids=self.input_batch.req_ids,
req_id_to_index=self.input_batch.req_id_to_index,
sampled_token_ids=[],
logprobs=None,
prompt_logprobs_dict={},
pooler_output=pooler_output,
**extra_args,
)
if vllm_version_is("0.10.1.1"):
modelrunner_output = ModelRunnerOutput(
req_ids=self.input_batch.req_ids,
req_id_to_index=self.input_batch.req_id_to_index,
sampled_token_ids=[],
spec_token_ids=None,
logprobs=None,
prompt_logprobs_dict={},
pooler_output=pooler_output,
**extra_args,
)
else:
modelrunner_output = ModelRunnerOutput(
req_ids=self.input_batch.req_ids,
req_id_to_index=self.input_batch.req_id_to_index,
sampled_token_ids=[],
logprobs=None,
prompt_logprobs_dict={},
pooler_output=pooler_output,
**extra_args,
)
return modelrunner_output
@torch.inference_mode()
def execute_model(
@@ -1757,15 +1779,27 @@ class NPUModelRunner(LoRAModelRunnerMixin):
extra_args = ({"kv_connector_output": kv_connector_output})
model_runner_output = ModelRunnerOutput(
req_ids=self.input_batch.req_ids,
req_id_to_index=self.input_batch.req_id_to_index,
sampled_token_ids=valid_sampled_token_ids,
logprobs=logprobs_lists,
prompt_logprobs_dict=prompt_logprobs_dict,
pooler_output=[],
**extra_args,
)
if vllm_version_is("0.10.1.1"):
model_runner_output = ModelRunnerOutput(
req_ids=self.input_batch.req_ids,
req_id_to_index=self.input_batch.req_id_to_index,
sampled_token_ids=valid_sampled_token_ids,
logprobs=logprobs_lists,
spec_token_ids=self._draft_token_ids,
prompt_logprobs_dict=prompt_logprobs_dict,
pooler_output=[],
**extra_args,
)
else:
model_runner_output = ModelRunnerOutput(
req_ids=self.input_batch.req_ids,
req_id_to_index=self.input_batch.req_id_to_index,
sampled_token_ids=valid_sampled_token_ids,
logprobs=logprobs_lists,
prompt_logprobs_dict=prompt_logprobs_dict,
pooler_output=[],
**extra_args,
)
durations = ProfileExecuteDuration().pop_captured_sync()
if durations: