Use unfold request for MoE device route generate.

intelgaoxiong · intelgaoxiong · commit f9e5dff25744 · 2026-06-23T23:03:18.000-07:00
Signed-off-by: intelgaoxiong &lt;xiong.gao@intel.com&gt;
diff --git a/src/plugins/intel_npu/src/plugin/npuw/llm_compiled_model.cpp b/src/plugins/intel_npu/src/plugin/npuw/llm_compiled_model.cpp
@@ -512,7 +512,7 @@ void apply_moe_config(ov::AnyMap& stage_config,
                                  "DEVICE_ROUTED mode uses in-graph gather-based expert selection which is only "
                                  "optimized for GENERATE stage. Please use HOST_ROUTED or DENSE for PREFILL.");
         }
-        stage_config["NPUW_UNFOLD_IREQS"] = "NO";
+        stage_config["NPUW_UNFOLD_IREQS"] = "YES";
     } else if (moe_hint == ::intel_npu::npuw::llm::MoEHint::DENSE) {
         LOG_INFO("MoE config for " << stage_name << " stage: DENSE (all experts active)");
         // DENSE mode requires CPU-only device due to extremely long NPU compilation time and high resource consumption

Original file line number	Diff line number	Diff line change
`@@ -512,7 +512,7 @@ void apply_moe_config(ov::AnyMap& stage_config,`
`512`	`512`	`"DEVICE_ROUTED mode uses in-graph gather-based expert selection which is only "`
`513`	`513`	`"optimized for GENERATE stage. Please use HOST_ROUTED or DENSE for PREFILL.");`
`514`	`514`	`}`
`515`		`- stage_config["NPUW_UNFOLD_IREQS"] = "NO";`
	`515`	`+ stage_config["NPUW_UNFOLD_IREQS"] = "YES";`
`516`	`516`	`} else if (moe_hint == ::intel_npu::npuw::llm::MoEHint::DENSE) {`
`517`	`517`	`LOG_INFO("MoE config for " << stage_name << " stage: DENSE (all experts active)");`
`518`	`518`	`// DENSE mode requires CPU-only device due to extremely long NPU compilation time and high resource consumption`