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

Drop 0.10.2 (#3284)

Browse Source 
Drop v0.10.2 support, we support vLLM 0.11.0rc3 now.
- vLLM version: v0.11.0rc3
- vLLM main:
https://github.com/vllm-project/vllm/commit/releases/v0.11.0

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-10-09 10:28:38 +08:00
committed by GitHub
parent 2dde1268c7
commit f12f76d7ba
17 changed files with 202 additions and 653 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
vllm_ascend/torchair/ops/torchair_fused_moe.py
View File

@@ -50,8 +50,7 @@ from vllm_ascend.torchair.utils import npu_stream_switch, npu_wait_tensor
from vllm_ascend.utils import (AscendSocVersion, dispose_tensor,
get_all_reduce_merge_state,
get_ascend_soc_version,
get_rm_router_logits_state, is_310p,
vllm_version_is)
get_rm_router_logits_state, is_310p)
def torchair_fused_experts_with_mc2(
@@ -1061,26 +1060,14 @@ class TorchairAscendFusedMoE(FusedMoE):
if self.scoring_func != "softmax" and not self.use_grouped_topk:
raise ValueError("Only softmax scoring function is supported for "
"non-grouped topk.")
if vllm_version_is("0.10.2"):
self.moe = FusedMoEConfig.make(
num_experts=self.global_num_experts,
experts_per_token=top_k,
hidden_dim=hidden_size,
num_local_experts=self.local_num_experts,
moe_parallel_config=self.moe_parallel_config,
# TODO (bnell): this needs to be fixed for quantized types.
in_dtype=params_dtype,
quant_config=quant_config)
else:
self.moe = FusedMoEConfig(
num_experts=self.global_num_experts,
experts_per_token=top_k,
hidden_dim=hidden_size,
num_local_experts=self.local_num_experts,
moe_parallel_config=self.moe_parallel_config,
in_dtype=params_dtype,
)
self.moe = FusedMoEConfig(
num_experts=self.global_num_experts,
experts_per_token=top_k,
hidden_dim=hidden_size,
num_local_experts=self.local_num_experts,
moe_parallel_config=self.moe_parallel_config,
in_dtype=params_dtype,
)
if quant_config is None:
self.quant_method = TorchairAscendUnquantizedFusedMoEMethod(
self.moe)
@@ -1242,12 +1229,8 @@ class TorchairAscendFusedMoE(FusedMoE):
router_logits = get_dp_group().all_gather(router_logits, 0)
elif fused_moe_state == FusedMoEState.NaiveMulticast:
if vllm_version_is("0.10.2"):
cu_tokens_across_dp_cpu = get_forward_context(
).dp_metadata.cu_tokens_across_dp_cpu
else:
cu_tokens_across_dp_cpu = get_forward_context(
).dp_metadata.cu_tokens_across_sp(1)
cu_tokens_across_dp_cpu = get_forward_context(
).dp_metadata.cu_tokens_across_sp(1)
hidden_states = self.naive_multicast(hidden_states,
cu_tokens_across_dp_cpu)
if self.rm_router_logits: