anodyine/vllm-omni
1
0 Fork 0
Code Issues 4 Pull Requests 2 Actions Packages Projects Releases Wiki Activity

[NPU] Fix HunyuanImage3 MoE EP+SP parallel mapping (#4344)

Browse Source 
Signed-off-by: Celeste-jq <591998922@qq.com>
This commit is contained in:
Jiaqian Liu authored and GitHub committed 2026-07-02 04:14:12 +00:00
1 parent b0f6b9de96
commit a25636d780
5 files changed
+362 -55
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

vllm_omni/diffusion/worker/diffusion_worker.py
+12 -3
View File
@@ -246,9 +246,18 @@ class DiffusionWorker:
# Create vllm_config for parallel configuration. Pass explicit device_config
# so DeviceConfig does not rely on current_platform in worker subprocesses.
vllm_config = _create_diffusion_worker_vllm_config(self.device, self.od_config)
vllm_config.parallel_config.tensor_parallel_size = self.od_config.parallel_config.tensor_parallel_size
vllm_config.parallel_config.data_parallel_size = self.od_config.parallel_config.data_parallel_size
vllm_config.parallel_config.enable_expert_parallel = self.od_config.parallel_config.enable_expert_parallel
parallel_config = self.od_config.parallel_config
vllm_config.parallel_config.tensor_parallel_size = parallel_config.tensor_parallel_size
vllm_config.parallel_config.data_parallel_size = parallel_config.data_parallel_size
if parallel_config.enable_expert_parallel and self.od_config.is_moe:
# Diffusion uses its own DP/CFG/SP groups normally. vLLM groups are
# only remapped for expert-parallel runtimes that consume vLLM's
# FusedMoE/EP semantics.
vllm_config.parallel_config.data_parallel_size = (
parallel_config.data_parallel_size * parallel_config.cfg_parallel_size
)
vllm_config.parallel_config.prefill_context_parallel_size = parallel_config.sequence_parallel_size
vllm_config.parallel_config.enable_expert_parallel = parallel_config.enable_expert_parallel
vllm_config.profiler_config = self.od_config.profiler_config
vllm_config.model_config = _make_diffusion_vllm_model_config(self.od_config) # type: ignore[assignment]
vllm_config.quant_config = self.od_config.quantization_config