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

[CI]Fixed the spell check function in typos.toml (#6753)

Browse Source 
### What this PR does / why we need it?
The incorrect regular expression syntax `.*[UE4M3|ue4m3].*` actually
ignores all words containing any of the following characters: `u, e, 4,
m, 3, |`

```yaml
extend-ignore-identifiers-re = [".*Unc.*", ".*_thw",
    ".*UE8M0.*", ".*[UE4M3|ue4m3].*", ".*eles.*", ".*fo.*", ".*ba.*",
    ".*ot.*", ".*[Tt]h[rR].*"]
```
===fix===>
```yaml
extend-ignore-identifiers-re = [".*Unc.*", ".*_thw",
    ".*UE8M0.*", ".*(UE4M3|ue4m3]).*", ".*eles.*", ".*fo.*", ".*ba.*",
    ".*ot.*", ".*[Tt]h[rR].*"]
```

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.15.0
- vLLM main:
9562912cea

Signed-off-by: MrZ20 <2609716663@qq.com>
This commit is contained in:
SILONG ZENG
2026-02-14 11:57:26 +08:00
committed by GitHub
parent 64aea60f2e
commit e2237819a9
31 changed files with 79 additions and 72 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm_ascend/ascend_config.py
View File

@@ -144,14 +144,14 @@ class AscendConfig:
if os.getenv("VLLM_ASCEND_ENABLE_PREFETCH_MLP", "0") == "1":
MAX_PREFETCH_WEIGHT_SIZE: int = 18 * 1024 * 1024
gate_up_prefetch_size = int(os.getenv("VLLM_ASCEND_MLP_GATE_UP_PREFETCH_SIZE", MAX_PREFETCH_WEIGHT_SIZE))
down_prefetch_szie = int(os.getenv("VLLM_ASCEND_MLP_DOWN_PREFETCH_SIZE", MAX_PREFETCH_WEIGHT_SIZE))
down_prefetch_size = int(os.getenv("VLLM_ASCEND_MLP_DOWN_PREFETCH_SIZE", MAX_PREFETCH_WEIGHT_SIZE))
self.weight_prefetch_config.set_mlp_pre_version_compatibale_config(
gate_up_prefetch_size, down_prefetch_szie
gate_up_prefetch_size, down_prefetch_size
)
logger.info_once(
f"MLP weight prefetch enabled from env variable VLLM_ASCEND_ENABLE_PREFETCH_MLP."
f"gate_up_prefetch_size={gate_up_prefetch_size}, "
f"down_prefetch_szie={down_prefetch_szie}."
f"down_prefetch_size={down_prefetch_size}."
)
warnings.warn(
"VLLM_ASCEND_ENABLE_PREFETCH_MLP is deprecated and will be removed in a v0.16.0 version. "

6
vllm_ascend/compilation/passes/allreduce_rmsnorm_fusion_pass.py
View File

@@ -34,13 +34,13 @@ else:
from vllm.compilation.passes.vllm_inductor_pass import VllmInductorPass
# computation-communication tiling block is 512
ALLREDUCE_NORM_FUSE_THREHOLD = 512
ALLREDUCE_NORM_FUSE_THRESHOLD = 512
def get_compile_range_and_extra_stream_check():
def check_func(match: Match) -> bool:
compile_range = get_pass_context().compile_range
return extra_stream_scope_check(match) and compile_range.start > ALLREDUCE_NORM_FUSE_THREHOLD
return extra_stream_scope_check(match) and compile_range.start > ALLREDUCE_NORM_FUSE_THRESHOLD
return check_func
@@ -176,5 +176,5 @@ class MatmulAllReduceAddRMSNormPass(VllmInductorPass):
"""
Check if the pass is applicable for the current configuration.
"""
applicable = compile_range.start > ALLREDUCE_NORM_FUSE_THREHOLD
applicable = compile_range.start > ALLREDUCE_NORM_FUSE_THRESHOLD
return applicable

10
vllm_ascend/core/scheduler_dynamic_batch.py
View File

@@ -86,9 +86,9 @@ class BudgetRefiner:
return k
return None
def _get_max_budget(self, num_deocde_tokens, num_decode):
def _get_max_budget(self, num_decode_tokens, num_decode):
"""Get the maximum budget according to the number of decoding tokens and the decoding requests."""
aligned_ctx = self._align_key(num_deocde_tokens, self.context_keys)
aligned_ctx = self._align_key(num_decode_tokens, self.context_keys)
aligned_dnum = self._align_key(num_decode, self.dnum_keys)
if aligned_ctx is None or aligned_dnum is None:
return self.default_budget
@@ -99,7 +99,7 @@ class BudgetRefiner:
# For debug.
# logger.info(
# f"budget {budget}, ctx,dnum {aligned_ctx, aligned_dnum}, "
# f"raw ctx,dnum {num_deocde_tokens, num_decode}"
# f"raw ctx,dnum {num_decode_tokens, num_decode}"
# )
return budget
@@ -114,8 +114,8 @@ class BudgetRefiner:
num_decode = len(num_decode_token_lst)
if num_decode <= 0:
return budget
num_deocde_tokens = sum(num_decode_token_lst) / num_decode
return self._get_max_budget(num_deocde_tokens, num_decode)
num_decode_tokens = sum(num_decode_token_lst) / num_decode
return self._get_max_budget(num_decode_tokens, num_decode)
class SchedulerDynamicBatch(Scheduler):

2
vllm_ascend/distributed/device_communicators/pyhccl_wrapper.py
View File

@@ -171,7 +171,7 @@ class HCCLLibrary:
path_to_library_cache: dict[str, Any] = {}
# class attribute to store the mapping from library path
# to the correspongding directory
# to the corresponding directory
path_to_dict_mapping: dict[str, dict[str, Any]] = {}
def __init__(self, so_file: str | None = None):

12
vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_connector.py
View File

@@ -1316,8 +1316,8 @@ class MooncakeConnectorWorker:
"""
prefill_tp_size = meta.remote_ptp_size if getattr(meta, "remote_ptp_size", None) else self._prefill_tp_size
if meta.remote_pcp_size * meta.remote_dcp_size * self.pcp_size * self.dcp_size == 1:
choosen_rank_list = self._get_remote_rank(req_id, prefill_tp_size)
remote_handshake_port_list = [[x + meta.remote_port for x in choosen_rank_list]]
chosen_rank_list = self._get_remote_rank(req_id, prefill_tp_size)
remote_handshake_port_list = [[x + meta.remote_port for x in chosen_rank_list]]
local_block_ids_list, remote_block_ids_list = [meta.local_block_ids], [meta.remote_block_ids]
return remote_handshake_port_list, local_block_ids_list, remote_block_ids_list
@@ -1563,8 +1563,8 @@ class MooncakeConnectorWorker:
),
)
else: # TODO: support prefill context parallel and pipeline parallel open at the same time
choosen_rank_list = self._get_remote_rank(remote_req_id, prefill_tp_size)
remote_handshake_port_list = [[x + meta.remote_port] for x in choosen_rank_list]
chosen_rank_list = self._get_remote_rank(remote_req_id, prefill_tp_size)
remote_handshake_port_list = [[x + meta.remote_port] for x in chosen_rank_list]
for i in range(tp_num_need_pulls * self._prefill_pp_size):
assert self.kv_recv_thread is not None
remote_host, remote_engine_id = self._get_remote_host_info_by_port(
@@ -1651,8 +1651,8 @@ class MooncakeConnectorWorker:
or self.use_sparse
):
tp_ori_data = tp_ori_data.reshape(-1, num_groups)
choosen_group = tp_ori_data[:, [rand_group_index]]
flattened = choosen_group.reshape(-1).tolist()
chosen_group = tp_ori_data[:, [rand_group_index]]
flattened = chosen_group.reshape(-1).tolist()
tp_sampled_nums = [
flattened[i : i + tp_num_need_pulls] for i in range(0, len(flattened), tp_num_need_pulls)
]

4
vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_layerwise_connector.py
View File

@@ -741,7 +741,7 @@ class MooncakeLayerwiseConnectorScheduler:
computed_tokens.get(req_id, 0) + scheduled_tokens - spec_decode_tokens
)
def add_tranfer_task(req_id, send_req_info: SendReqInfo, chunk_finish=False):
def add_transfer_task(req_id, send_req_info: SendReqInfo, chunk_finish=False):
(
local_block_ids,
local_transed_tokens,
@@ -771,7 +771,7 @@ class MooncakeLayerwiseConnectorScheduler:
# whether chunk finish
chunk_finish = send_req_info.local_computed_tokens >= len(send_req_info.request.all_token_ids)
add_tranfer_task(req_id, send_req_info, chunk_finish=chunk_finish)
add_transfer_task(req_id, send_req_info, chunk_finish=chunk_finish)
if chunk_finish:
self._reqs_need_send_layerwise.pop(req_id)
return meta

4
vllm_ascend/eplb/core/eplb_device_transfer_loader.py
View File

@@ -68,7 +68,7 @@ class D2DExpertWeightLoader:
self.updated_log2phy_map = log2phy_map
def asyn_expert_weight_transfer(self, reqs):
# Only when send/recv tasks are parsed into self.comm_op_list, d2d send/recv tasks can be luanched
# Only when send/recv tasks are parsed into self.comm_op_list, d2d send/recv tasks can be launched
if self.state != ExpertWeightUpdateState.READY:
return
@@ -80,7 +80,7 @@ class D2DExpertWeightLoader:
self.state = ExpertWeightUpdateState.TRANSFERRING
def update_expert_map_and_weight(self, reqs):
# Only after send/recv tasks have been luanched, expert_map and weight can be updated
# Only after send/recv tasks have been launched, expert_map and weight can be updated
if self.state != ExpertWeightUpdateState.TRANSFERRING:
return

4
vllm_ascend/eplb/core/policy/policy_flashlb.py
View File

@@ -130,8 +130,8 @@ def jsq_placement(X, pieces, M, stage_weights):
score = 0.0
for s in range(n_stage):
tmp_sj = loads[s, j] + w[s]
numer_sj = tmp_sj if tmp_sj > stage_max[s] else stage_max[s]
score += stage_weights[s] * (numer_sj / denom[s])
number_sj = tmp_sj if tmp_sj > stage_max[s] else stage_max[s]
score += stage_weights[s] * (number_sj / denom[s])
if score < best_val:
best_val = score
best_j = j

10
vllm_ascend/platform.py
View File

@@ -195,10 +195,10 @@ class NPUPlatform(Platform):
)
if vllm_config.additional_config.get("ascend_compilation_config", {}).get("fuse_allreduce_rms", True):
from vllm_ascend.compilation.passes.allreduce_rmsnorm_fusion_pass import ALLREDUCE_NORM_FUSE_THREHOLD
from vllm_ascend.compilation.passes.allreduce_rmsnorm_fusion_pass import ALLREDUCE_NORM_FUSE_THRESHOLD
new_compile_ranges_split_points = vllm_config.compilation_config.compile_ranges_split_points
new_compile_ranges_split_points.append(ALLREDUCE_NORM_FUSE_THREHOLD)
new_compile_ranges_split_points.append(ALLREDUCE_NORM_FUSE_THRESHOLD)
new_compile_ranges_split_points = sorted(new_compile_ranges_split_points)
vllm_config.compilation_config.compile_ranges_split_points = new_compile_ranges_split_points
logger.debug(
@@ -208,10 +208,10 @@ class NPUPlatform(Platform):
npugraph_ex_config = ascend_config.npugraph_ex_config
if npugraph_ex_config and npugraph_ex_config.fuse_allreduce_rms:
from vllm_ascend.compilation.passes.allreduce_rmsnorm_fusion_pass import ALLREDUCE_NORM_FUSE_THREHOLD
from vllm_ascend.compilation.passes.allreduce_rmsnorm_fusion_pass import ALLREDUCE_NORM_FUSE_THRESHOLD
new_compile_ranges_split_points = vllm_config.compilation_config.compile_ranges_split_points
new_compile_ranges_split_points.append(ALLREDUCE_NORM_FUSE_THREHOLD)
new_compile_ranges_split_points.append(ALLREDUCE_NORM_FUSE_THRESHOLD)
new_compile_ranges_split_points = sorted(new_compile_ranges_split_points)
vllm_config.compilation_config.compile_ranges_split_points = new_compile_ranges_split_points
logger.debug(
@@ -558,7 +558,7 @@ class NPUPlatform(Platform):
Args:
attn_metadata (dict[str, Any]): attention metadata for all layers.
vllm_config (VllmConfig): configuration of vllm.
dp_metadata (DpMetada): metadata for data parallelism.
dp_metadata (Dpmetadata): metadata for data parallelism.
lack of typehint because of circular import.
virtual_engine (int, optional): index of virtual engine. Defaults to 0.
num_tokens (int | None, optional): number of tokens. Defaults to None.

4
vllm_ascend/spec_decode/eagle_proposer.py
View File

@@ -941,7 +941,7 @@ class EagleProposer(VllmEagleProposer):
# [0, 1, 2, 3, 4, 5, 6, 7, 8] ->
# [0, 1, 0, 1, 2, 3, 0, 1, 2]
# _r1_ ____r2____ ___r3__
token_offests = self.token_arange_np[:total_num_tokens] - new_query_start_locs_expanded
token_offsets = self.token_arange_np[:total_num_tokens] - new_query_start_locs_expanded
# Expand starting positions to match token pattern
# [0, q1, q1 + q2] ->
@@ -952,7 +952,7 @@ class EagleProposer(VllmEagleProposer):
# [0, 1, // req 1
# q1 + 0, q1 + 1, q1 + 2, q1 + 3, // req 2
# q1 + q2 + 0, q1 + q2 + 1, q1 + q2 + 2] // req 3
token_indices_np = token_offests + old_query_start_locs_expanded
token_indices_np = token_offsets + old_query_start_locs_expanded
token_indices = torch.from_numpy(token_indices_np).to(device, non_blocking=True)
common_attn_metadata.slot_mapping[: token_indices.shape[0]].copy_(

2
vllm_ascend/spec_decode/interface.py
View File

@@ -35,7 +35,7 @@ class Proposer:
aclgraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE,
batch_descriptor=None,
):
"""Called by dummy_run in modle_runner"""
"""Called by dummy_run in model_runner"""
raise NotImplementedError
def generate_token_ids(

2
vllm_ascend/worker/model_runner_v1.py
View File

@@ -2390,7 +2390,7 @@ class NPUModelRunner(GPUModelRunner):
to be reshaped to the desired shape before being used by the models.
NOTE: To support prefill disaggregation, we need to split kvcache tensor into
k_cahce and v cache, and the addr of both are aligned by 2M
k_cache and v cache, and the addr of both are aligned by 2M
Args:
kv_cache_config: The KV cache config

4
vllm_ascend/worker/pcp_utils.py
View File

@@ -459,9 +459,9 @@ class PCPManager:
# draft_len of each request [1, 2, 1]
# then prev_draft_token_indices is [0, 2, 3, 4]
prev_draft_token_indices.extend(range(start, start + draft_len))
num_commmon_tokens = len(sample_flattened_indices)
num_common_tokens = len(sample_flattened_indices)
if num_commmon_tokens == 0:
if num_common_tokens == 0:
# No requests in common with the previous iteration
# So input_ids.cpu will have all the input ids.
return