Code clean up

Signed-off-by: intelgaoxiong <xiong.gao@intel.com>

Author: intelgaoxiong

src/plugins/intel_npu/src/plugin/npuw/llm_compiled_model.cpp

@@ -1205,21 +1205,6 @@ ov::npuw::LLMCompiledModel::LLMCompiledModel(const std::shared_ptr<ov::Model>& m
     }
 
     // Compile multiple generate model variants with different sizes
-    // DBG: dump final effective config keys relevant to online partitioning / attention isolation
-    {
-        auto dump_key = [](const ov::AnyMap& cfg, const std::string& key, const std::string& label) {
-            auto it = cfg.find(key);
-            std::cout << "[LLM_CFG] " << label << " " << key << " = "
-                      << (it != cfg.end() ? it->second.as<std::string>() : "(not set)") << std::endl;
-        };
-        dump_key(prefill_config, "NPUW_ONLINE_PIPELINE", "PREFILL");
-        dump_key(prefill_config, "NPUW_ONLINE_ISOLATE", "PREFILL");
-        dump_key(prefill_config, "NPUW_ATTN", "PREFILL");
-        dump_key(generate_config, "NPUW_ONLINE_PIPELINE", "GENERATE");
-        dump_key(generate_config, "NPUW_ONLINE_ISOLATE", "GENERATE");
-        dump_key(generate_config, "NPUW_ATTN", "GENERATE");
-        dump_key(generate_config, "NPUW_UNFOLD_IREQS", "GENERATE");
-    }
     compile_generate_model_variants(generate_model_variants, plugin, generate_config);
 
     m_prefill_compiled = m_compiled_model_factory(prefill_model, plugin, prefill_config);

src/plugins/intel_npu/src/plugin/npuw/moe/moe_infer_utils.cpp

@@ -66,24 +66,20 @@ ov::Tensor slice_expert_weight(const ov::Tensor& batched_weight, size_t expert_i
 }
 
 namespace {
-// Returns the token-count dimension from a 4-D router output shape.
-// Two layout variants exist depending on the opset version used during model export:
-//   Layout A: [num_experts, 1, token_num, 1]  → returns shape[2]
-//   Layout B: [num_experts, token_num, 1, 1]  → returns shape[1]
-// ASSERTs if neither layout is recognised.
+// Returns the token-count dimension from a router output shape.
+// All supported layouts have a trailing singleton dim:
+//   [N, token, 1]     (3-D)          → token at dim 1
+//   [N, 1, token, 1]  (4-D, dim1=1)  → token at dim 2
+//   [N, token, 1, 1]  (4-D, dim1≠1)  → token at dim 1
+// ASSERTs if rank ∉ {3, 4} or trailing dim ≠ 1.
 static size_t get_router_token_count(const ov::Shape& router_shape) {
-    if (router_shape.size() == 3) {
-        // 3-D [num_experts, token_count, 1] (Qwen3 prefill)
-        NPUW_ASSERT(router_shape[2] == 1 && "Unexpected 3-D router shape: trailing dim must be 1");
-        return router_shape[1];
-    }
-    NPUW_ASSERT(router_shape.size() == 4);
-    if (router_shape[1] == 1 && router_shape[3] == 1)
-        return router_shape[2];  // Layout A: [num_experts, 1, token_count, 1]
-    if (router_shape[2] == 1 && router_shape[3] == 1)
-        return router_shape[1];  // Layout B: [num_experts, token_count, 1, 1]
-    NPUW_ASSERT(false && "Unexpected router output shape - cannot determine token dimension!");
-    return 0;  // unreachable, suppress warning
+    const auto rank = router_shape.size();
+    NPUW_ASSERT((rank == 3 || rank == 4) && router_shape.back() == 1 &&
+                "Router shape must be 3-D or 4-D with trailing singleton");
+    // [N, 1, token, 1] is the only layout where dim 1 is a singleton rather than the token dim.
+    if (rank == 4 && router_shape[1] == 1)
+        return router_shape[2];  // [N, 1, token, 1]
+    return router_shape[1];      // [N, token, 1, 1] or [N, token, 1]
 }
 }  // namespace
 

src/plugins/intel_npu/src/plugin/npuw/moe_transformations/moe_transformation.cpp

@@ -385,38 +385,32 @@ std::optional<MoEDownstream> detect_and_transform_moe_downstream(const std::shar
     LOG_DEBUG("Detecting MoE downstream pattern...");
     LOG_BLOCK();
 
-    // Pattern to match: Parameter -> [Convert] -> ReduceSum
-    // Looking for a parameter with shape [N, ..., W] where:
-    //   dim 0  = N (total experts, > 1)
-    //   dim n-1 = W (hidden_dim)
-    //   Accepted layouts:
-    //     4-D: [N, 1, H, W] (GPT-OSS) or [N, H, 1, W] (Qwen, decoding)
-    //     3-D: [N, token, W] (Qwen, prefill — token may be > 1)
-    // The primary discriminator is check_downstream_pattern() (Parameter -> [Convert] -> ReduceSum).
+    // Pattern: Parameter -> [Convert] -> ReduceSum, shape [N, ..., W] where:
+    //   rank 3: [N, token, W]
+    //   rank 4: [N, 1, H, W] or [N, H, 1, W]
+    //   shape[0] = N > 1 (num_experts).
     const auto& params = model->get_parameters();
 
     for (size_t param_idx = 0; param_idx < params.size(); ++param_idx) {
         const auto& param = params[param_idx];
 
-        // Validate shape: must be 3-D or 4-D with dim 0 > 1
-        auto param_shape = param->get_partial_shape();
-        if (!param_shape.rank().is_static()) {
-            continue;
-        }
-        const auto rank = param_shape.rank().get_length();
-        if (rank < 3 || rank > 4) {
-            continue;
-        }
-
-        auto shape = param_shape.to_shape();
-        if (shape[0] <= 1) {
+        auto shape_opt = [&]() -> std::optional<ov::Shape> {
+            const auto ps = param->get_partial_shape();
+            if (!ps.rank().is_static())
+                return std::nullopt;
+            const auto rank = ps.rank().get_length();
+            if (rank < 3 || rank > 4)
+                return std::nullopt;
+            const auto s = ps.to_shape();
+            if (s[0] <= 1)
+                return std::nullopt;
+            if (rank == 4 && s[1] != 1 && s[2] != 1)
+                return std::nullopt;
+            return s;
+        }();
+        if (!shape_opt || !check_downstream_pattern(param))
             continue;
-        }
-
-        // Check pattern: Parameter -> Convert -> ReduceSum
-        if (!check_downstream_pattern(param)) {
-            continue;
-        }
+        const auto& shape = *shape_opt;
 
         LOG_DEBUG("  Found downstream pattern for Parameter: " << param->get_friendly_name());
         LOG_DEBUG("  Pattern matched: Parameter -> Convert -> ReduceSum");
@@ -468,7 +462,6 @@ std::optional<MoEDownstream> create_moe_downstream(const std::shared_ptr<ov::Mod
     LOG_BLOCK();
 
     LOG_INFO("Using K=" << active_experts_num << " active experts for downstream");
-    std::cout << "Using K=" << active_experts_num << " active experts for downstream" << std::endl;
 
     auto downstream_info = detect_and_transform_moe_downstream(model, active_experts_num);
     if (downstream_info && downstream_info->is_valid()) {
@@ -482,7 +475,6 @@ std::optional<MoEDownstream> create_moe_downstream(const std::shared_ptr<ov::Mod
     }
 
     LOG_WARN("Failed to create MoEDownstream - downstream pattern not found");
-    std::cout << "Failed to create MoEDownstream - downstream pattern not found" << std::endl;
     return std::nullopt;
 }