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

[long_seq] remove long_seq env (#4660)

Browse Source 
### What this PR does / why we need it?
remove env VLLM_ASCEND_ENABLE_CONTEXT_PARALLEL 

- vLLM version: v0.12.0

---------

Signed-off-by: LookAround <lixushi@huawei.com>
Signed-off-by: ZhangMingWei716 <2894054457@qq.com>
Co-authored-by: ZhangMingWei716 <2894054457@qq.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
LookAround0301
2025-12-05 10:31:49 +08:00
committed by GitHub
parent ea54388e19
commit b32ef53b3b
16 changed files with 230 additions and 176 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
vllm_ascend/attention/attention_v1.py
View File

@@ -26,10 +26,13 @@ import torch.nn as nn
import torch_npu
from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
AttentionLayer, AttentionType)
from vllm.attention.backends.registry import (AttentionBackendEnum,
register_backend)
from vllm.config import VllmConfig
from vllm.distributed import (get_dcp_group,
get_decode_context_model_parallel_rank,
get_decode_context_model_parallel_world_size)
get_decode_context_model_parallel_world_size,
get_pcp_group)
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.utils.math_utils import cdiv
from vllm.v1.attention.backends.utils import AttentionCGSupport
@@ -41,19 +44,7 @@ from vllm_ascend.attention.utils import (AscendCommonAttentionMetadata,
split_decodes_and_prefills)
from vllm_ascend.compilation.acl_graph import (get_graph_params,
update_graph_params_workspaces)
from vllm_ascend.utils import prefill_context_parallel_enable, weak_ref_tensors
# isort: off
if prefill_context_parallel_enable():
from vllm.distributed import (get_pcp_group,
get_prefill_context_model_parallel_rank,
get_prefill_context_model_parallel_world_size
)
# isort: on
from vllm.attention.backends.registry import (AttentionBackendEnum,
register_backend)
from vllm_ascend.utils import weak_ref_tensors
@register_backend(AttentionBackendEnum.CUSTOM, "ASCEND")
@@ -255,10 +246,9 @@ class AscendAttentionMetadataBuilder:
vllm_config.scheduler_config.max_num_batched_tokens,
dtype=torch.uint8,
device=device)
self.pcp_size = get_prefill_context_model_parallel_world_size(
) if prefill_context_parallel_enable() else 1
self.pcp_rank = get_prefill_context_model_parallel_rank(
) if self.pcp_size > 1 else 0
self.pcp_size = get_pcp_group().world_size
self.pcp_rank = get_pcp_group(
).rank_in_group if self.pcp_size > 1 else 0
self.dcp_size = get_decode_context_model_parallel_world_size()
self.dcp_rank = get_decode_context_model_parallel_rank(
) if self.dcp_size > 1 else 0
@@ -350,8 +340,7 @@ class AscendAttentionMetadataBuilder:
context_lens_cpu = num_computed_tokens_cpu[
num_decodes:num_reqs]
max_context_len_cpu = context_lens_cpu.max().item()
pcp_size = get_prefill_context_model_parallel_world_size(
) if prefill_context_parallel_enable() else 1
pcp_size = get_pcp_group().world_size
if self.chunked_prefill_enabled and max_context_len_cpu > 0:
local_context_lens_allranks = torch.tensor(
num_computed_tokens_of_pcp_dcp
@@ -539,10 +528,9 @@ class AscendAttentionBackendImpl(AttentionImpl):
self.num_queries_per_kv = self.num_heads // self.num_kv_heads
self.key_cache = None
self.value_cache = None
self.pcp_size = get_prefill_context_model_parallel_world_size(
) if prefill_context_parallel_enable() else 1
self.pcp_rank = get_prefill_context_model_parallel_rank(
) if self.pcp_size > 1 else 0
self.pcp_size = get_pcp_group().world_size
self.pcp_rank = get_pcp_group(
).rank_in_group if self.pcp_size > 1 else 0
self.pcp_group = get_pcp_group(
).device_group if self.pcp_size > 1 else None

29
vllm_ascend/attention/mla_v1.py
View File

@@ -13,7 +13,7 @@ from vllm.config import VllmConfig, get_current_vllm_config
from vllm.distributed import (get_dcp_group,
get_decode_context_model_parallel_rank,
get_decode_context_model_parallel_world_size,
get_tensor_model_parallel_rank,
get_pcp_group, get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
get_tp_group)
from vllm.forward_context import ForwardContext, get_forward_context
@@ -37,17 +37,9 @@ from vllm_ascend.compilation.acl_graph import (get_graph_params,
from vllm_ascend.ops.weight_prefetch import maybe_npu_prefetch
from vllm_ascend.quantization.w8a8 import AscendW8A8LinearMethod
from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_ND, ACL_FORMAT_FRACTAL_NZ,
is_enable_nz, prefill_context_parallel_enable,
weak_ref_tensors)
is_enable_nz, weak_ref_tensors)
from vllm_ascend.worker.npu_input_batch import InputBatch
# isort: off
if prefill_context_parallel_enable():
from vllm.distributed import (get_pcp_group,
get_prefill_context_model_parallel_rank,
get_prefill_context_model_parallel_world_size
)
# isort: on
if TYPE_CHECKING:
from vllm.v1.core.sched.output import SchedulerOutput
@@ -265,15 +257,13 @@ class AscendMLAMetadataBuilder:
self.cos_cache = None
self.sin_cache = None
self.pcp_size = get_prefill_context_model_parallel_world_size(
) if prefill_context_parallel_enable() else 1
self.pcp_rank = get_prefill_context_model_parallel_rank(
) if self.pcp_size > 1 else 0
self.pcp_size = get_pcp_group().world_size
self.pcp_rank = get_pcp_group(
).rank_in_group if self.pcp_size > 1 else 0
self.dcp_size = get_decode_context_model_parallel_world_size()
self.dcp_rank = get_decode_context_model_parallel_rank(
) if self.dcp_size > 1 else 0
self.cp_local_block_size = vllm_config.parallel_config.cp_kv_cache_interleave_size if prefill_context_parallel_enable(
) else 1
self.cp_local_block_size = vllm_config.parallel_config.cp_kv_cache_interleave_size
self.cp_virtual_block_size = self.cp_local_block_size * self.dcp_size * self.pcp_size
decode_max_num_seqs = getattr(scheduler_config, 'decode_max_num_seqs',
0)
@@ -868,10 +858,9 @@ class AscendMLAImpl(MLAAttentionImpl):
self.speculative_config = vllm_config.speculative_config
self.enable_mlapo = envs.VLLM_ASCEND_ENABLE_MLAPO
self.pcp_size = get_prefill_context_model_parallel_world_size(
) if prefill_context_parallel_enable() else 1
self.pcp_rank = get_prefill_context_model_parallel_rank(
) if self.pcp_size > 1 else 0
self.pcp_size = get_pcp_group().world_size
self.pcp_rank = get_pcp_group(
).rank_in_group if self.pcp_size > 1 else 0
self.pcp_group = get_pcp_group(
).device_group if self.pcp_size > 1 else None