Skip to content

[bugfix] Fix dummy-run and multi-node issues in MoE routing and MTP #4947

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
wangxiyuan merged 1 commit into from
Dec 15, 2025
Merged

[bugfix] Fix dummy-run and multi-node issues in MoE routing and MTP #4947

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 0 additions & 1 deletion csrc/dispatch_ffn_combine/op_kernel/moe_init_routing_quant_v2/moe_init_routing_quant_v2.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -114,7 +114,6 @@ __aicore__ inline void moe_init_routing_quant_v2(
srcToDstAndGatherOp.Init(x, scale, expandedRowIdx, expandedX, dynamicQuantScale, workspace, tilingData, &srcToDstGatherPipe);
srcToDstAndGatherOp.Process();
srcToDstGatherPipe.Destroy();
return;
}
}

Expand Down
3 changes: 3 additions & 0 deletions vllm_ascend/spec_decode/mtp_proposer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -735,6 +735,9 @@ def _propose(
num_input_tokens, self.runner.with_prefill)

moe_comm_type = self.runner._select_moe_comm_method(num_input_tokens)
# TODO: remove this after moe_comm_type selection logic is finalized
moe_comm_type = (MoECommType.ALLTOALL if moe_comm_type
== MoECommType.FUSED_ALLTOALL else moe_comm_type)

# Enable shared_expert_dp and MTP FULL graph may cause accuracy issues.
if scheduler_output and not self.enable_shared_expert_dp:
Expand Down
7 changes: 4 additions & 3 deletions vllm_ascend/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -912,16 +912,17 @@ def calculate_ep_buffer_size() -> int:
try:
from vllm.config import get_current_vllm_config
vllm_config = get_current_vllm_config()
tp_size = vllm_config.parallel_config.tensor_parallel_size
hf_config = vllm_config.model_config.hf_config

hidden_size = hf_config.hidden_size
topk = getattr(hf_config, "num_experts_per_token", 1)
batch_size = vllm_config.scheduler_config.max_num_batched_tokens
topk = getattr(hf_config, "num_experts_per_tok", 1)
batch_size = vllm_config.scheduler_config.max_num_batched_tokens // tp_size
int8_size = torch.iinfo(torch.int8).bits // 8
bf16_size = torch.finfo(torch.bfloat16).bits // 8
ep_buffer_size = math.ceil(
(batch_size * hidden_size * topk *
(int8_size * 2 + bf16_size)) / (1024 * 1024))
(int8_size + bf16_size) * 3) / (1024 * 1024))
except Exception:
pass
return max(ep_buffer_size, _DEFAULT_BUFFER_SIZE)
Expand Down
Loading