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

Support v0.10.1 (#2584)

Browse Source 
### What this PR does / why we need it?
This patch also supports v0.10.1

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

### How was this patch tested?
- CI passed
- test 0.10.1: https://github.com/vllm-project/vllm-ascend/pull/2583
- vLLM version: v0.10.1.1
- vLLM main:
321938e9ac

Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
This commit is contained in:
Yikun Jiang
2025-08-28 18:47:53 +08:00
committed by GitHub
parent 6c973361fc
commit 175f6bc445
8 changed files with 40 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm_ascend/core/scheduler.py
View File

@@ -33,7 +33,7 @@ from vllm.v1.structured_output import StructuredOutputManager
from vllm_ascend.utils import vllm_version_is
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
from vllm.v1.core.kv_cache_manager import KVCacheBlocks
else:
KVCacheBlocks = None
@@ -66,7 +66,7 @@ class AscendScheduler(Scheduler):
scheduled_running_reqs: list[Request] = []
preempted_reqs: list[Request] = []
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
req_to_new_block_ids: dict[str, list[int]] = {}
else:
req_to_new_blocks: dict[str, KVCacheBlocks] = {}
@@ -227,7 +227,7 @@ class AscendScheduler(Scheduler):
if self.lora_config and request.lora_request:
scheduled_loras.add(request.lora_request.lora_int_id)
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
req_to_new_block_ids[request.request_id] = (
self.kv_cache_manager.get_block_ids(request.request_id))
else:
@@ -320,7 +320,7 @@ class AscendScheduler(Scheduler):
# Schedule the request.
scheduled_running_reqs.append(request)
self.scheduled_req_ids.add(request.request_id)
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
req_to_new_block_ids[request.request_id] = (
new_blocks.get_block_ids())
else:
@@ -362,7 +362,7 @@ class AscendScheduler(Scheduler):
any_request, len(self.running)))
# Construct the scheduler output.
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
new_reqs_data = [
NewRequestData.from_request(
req, req_to_new_block_ids[req.request_id])
@@ -385,7 +385,7 @@ class AscendScheduler(Scheduler):
req_to_new_blocks)
scheduled_cached_reqs = cached_reqs_data
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
scheduler_output = SchedulerOutput(
scheduled_new_reqs=new_reqs_data,
scheduled_cached_reqs=scheduled_cached_reqs,

2
vllm_ascend/models/qwen3_moe.py
View File

@@ -254,7 +254,7 @@ class CustomQwen3MoeModel(Qwen3MoeModel):
quant_config = vllm_config.quant_config
parallel_config = vllm_config.parallel_config
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
self.num_redundant_experts = parallel_config.num_redundant_experts
else:
eplb_config = parallel_config.eplb_config

6
vllm_ascend/sample/sampler.py
View File

@@ -5,7 +5,7 @@ from vllm.v1.sample.sampler import Sampler
from vllm_ascend.utils import is_310p, vllm_version_is
if not vllm_version_is("0.10.1.1"):
if not (vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1")):
from vllm.config import LogprobsMode
DEFAULT_LOGPROBS_MODE = LogprobsMode.RAW_LOGPROBS
else:
@@ -68,7 +68,7 @@ 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"):
if not (vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1")):
logits_to_return = None
if self.logprobs_mode == LogprobsMode.PROCESSED_LOGITS:
@@ -79,7 +79,7 @@ class AscendTopKTopPSampler(TopKTopPSampler):
probs = logits.softmax(dim=-1, dtype=torch.float32)
output = None
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
output = random_sample(probs, generators)
else:
output = (random_sample(probs, generators), logits_to_return)

30
vllm_ascend/worker/model_runner_v1.py
View File

@@ -96,7 +96,7 @@ 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"):
if not (vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1")):
from vllm.v1.outputs import DraftTokenIds
else:
DraftTokenIds = None
@@ -384,7 +384,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
# Remove finished requests from the cached states.
for req_id in scheduler_output.finished_req_ids:
self.requests.pop(req_id, None)
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
self.encoder_cache.pop(req_id, None)
# Remove the finished requests from the persistent batch.
# NOTE(woosuk): There could be an edge case where finished_req_ids and
@@ -394,7 +394,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
# and handling the second as a new request.
for req_id in scheduler_output.finished_req_ids:
self.input_batch.remove_request(req_id)
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
# Free the cached encoder outputs.
for req_id, input_id in scheduler_output.free_encoder_input_ids:
encoder_outputs = self.encoder_cache.get(req_id)
@@ -455,9 +455,10 @@ class NPUModelRunner(LoRAModelRunnerMixin):
lora_request=new_req_data.lora_request,
**({
"mm_hashes": new_req_data.mm_hashes
} if not vllm_version_is("0.10.1.1") else {
"mm_hashes": None
}),
} if not (vllm_version_is("0.10.1.1")
or vllm_version_is("0.10.1")) else {
"mm_hashes": None
}),
)
# Only relevant for models using M-RoPE (e.g, Qwen2-VL)
@@ -893,13 +894,13 @@ class NPUModelRunner(LoRAModelRunnerMixin):
# Batch the multi-modal inputs.
mm_kwargs = list[MultiModalKwargsItem]()
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
req_ids_pos = list[tuple[str, int, PlaceholderRange]]()
else:
mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
req_state = self.requests[req_id]
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
for mm_input_id in encoder_input_ids:
mm_kwargs.append(req_state.mm_kwargs[mm_input_id])
req_ids_pos.append((req_id, mm_input_id,
@@ -942,7 +943,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
for output in curr_group_outputs:
encoder_outputs.append(output)
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
# Cache the encoder outputs.
for (req_id, input_id, pos_info), output in zip(
req_ids_pos,
@@ -974,7 +975,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
req_state = self.requests[req_id]
num_computed_tokens = req_state.num_computed_tokens
mm_positions = req_state.mm_positions
if not vllm_version_is("0.10.1.1"):
if not (vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1")):
mm_hashes = req_state.mm_hashes
for i, pos_info in enumerate(mm_positions):
start_pos = pos_info.offset
@@ -993,7 +994,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
continue
start_idx = max(num_computed_tokens - start_pos, 0)
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
end_idx = min(
num_computed_tokens - start_pos + num_scheduled_tokens,
num_encoder_tokens)
@@ -1719,7 +1720,8 @@ class NPUModelRunner(LoRAModelRunnerMixin):
logits = None
else:
if self.input_batch.pooling_params:
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is(
"0.10.1"):
return self._pool_v010(
hidden_states,
scheduler_output.total_num_scheduled_tokens,
@@ -1867,7 +1869,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
extra_args = ({"kv_connector_output": kv_connector_output})
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
model_runner_output = ModelRunnerOutput(
req_ids=self.input_batch.req_ids,
req_id_to_index=self.input_batch.req_id_to_index,
@@ -2191,7 +2193,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
dummy_pooling_params = PoolingParams(task=task)
to_update = model.pooler.get_pooling_updates(task)
to_update.apply(dummy_pooling_params)
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
dummy_prompt_lens = torch.tensor(
[h.shape[0] for h in hidden_states_list],
device=self.device,

2
vllm_ascend/worker/npu_input_batch.py
View File

@@ -726,7 +726,7 @@ class InputBatch:
pooling_params = [
self.pooling_params[req_id] for req_id in self.req_ids
]
if vllm_version_is("0.10.1.1"):
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
return PoolingMetadata(
prompt_lens=torch.from_numpy(
self.num_prompt_tokens[:self.num_reqs]).to(self.device),

2
vllm_ascend/worker/worker_v1.py
View File

@@ -50,7 +50,7 @@ from vllm_ascend.utils import (init_ascend_soc_version,
try_register_lib, vllm_version_is)
from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
if not vllm_version_is("0.10.1.1"):
if not (vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1")):
from vllm.v1.outputs import DraftTokenIds
else:
DraftTokenIds = None