npu-pack: a 1.8 hybrid pack's [0] MatMulNBits placeholders are not weights; refuse by name and point at model.bin (Refs #1503)

npu/pack_import.cpp

@@ -407,10 +407,16 @@ OnnxModel parse_onnx(const std::string& path) {
             continue;
         }
         if (!w) { kept.push_back(mm); continue; }   // names nothing this graph declares: not this genre's business
+        // Two spellings of the placeholder: 1.9 names it `<w>.empty` beside `<w>.packed`; the 1.8 hybrid line names the
+        // node's own `<w>` as the `[0]` placeholder beside `<w>.packed` (Llama-3.2-1B-Instruct_rai_1.8.0_hybrid: all 162
+        // `.MatMulNBits.*` initializers dims [0] and no data, the weights in model.bin). Missing the second spelling kept
+        // the node as a real weight, so `info` reported every projection importable and `import` died on
+        // `...MatMulNBits.bias: no data` (engine#1503).
         std::string packed = w->name;
         if (packed.size() >= 6 && packed.compare(packed.size() - 6, 6, ".empty") == 0)
             packed.replace(packed.size() - 6, 6, ".packed");
-        else { kept.push_back(mm); continue; }
+        else
+            packed += ".packed";
         const OnnxTensor* pw = find_init(packed);
         // No sibling at all: not this genre's pack. Keep the node exactly as it was -- it still names its `*.empty`
         // initializer and the reader below decides what that means. Dropping it here would be a decision made by an
@@ -3596,7 +3602,12 @@ Plan plan(const std::string& pack_dir) {
             pack_dir + ": " + onnx.weightless_example +
             ", and the pack carries no Token .fconst either (no dd_metastate_Llm_Token_*.fconst): these weights are "
             "genuinely absent from the pack's external data, so this pack cannot be imported -- not a tile-layout or "
-            "block-size problem");
+            "block-size problem" +
+            // The 1.8 hybrid line keeps them in a `model.bin` beside the graph (1.3 GB on Llama-3.2-1B-Instruct), a
+            // container this importer does not read; saying where they are is what tells a reader which work is missing.
+            std::string(fs::exists(pack_dir + "/model.bin")
+                            ? ". This pack's weights are in its model.bin, which this importer does not read (engine#1503)"
+                            : ""));
     const bool no_fconst = tok_path.empty();
     // The 1.9 external-data genre (phi-4: 201 `MatMulNBitsBf` ops whose bytes are `*.qweight.packed` siblings in
     // `eager-mm.onnx.data`) must NOT ride the graph-weights path below, which was written for the 1.3/1.6/1.7.1 line's

tests/npu_pack_import_test.cpp

@@ -1350,6 +1350,40 @@ void test_crafted_refusals(const std::string& root) {
     }
     check(!fs::exists(dir / "out3" / "model.qnx") && !fs::exists(dir / "out3" / "model.q4nx"),
           "crafted: nothing written for the empty .packed sibling");
+
+    // 4. engine#1503: the 1.8 hybrid spelling of the same thing. The node names `<w>` ITSELF as the [0] placeholder (no
+    // `.empty` suffix) beside an equally empty `<w>.packed`, and the weights are in a model.bin next to the graph
+    // (Llama-3.2-1B-Instruct_rai_1.8.0_hybrid: all 162 `.MatMulNBits.*` initializers are this shape). Before the fix the
+    // walker only knew the `.empty` spelling, kept the node as a real weight, `info` called every projection importable
+    // and `import` died much later on `...MatMulNBits.bias: no data`.
+    put_file((dir / "model.onnx").string(),
+             pb::model({pb::tensor("model.embed_tokens.weight", {8, 8}, "model.onnx.data", 0, 256),
+                        pb::tensor("rai18_qweight", {0}, "", 0, 0),
+                        pb::tensor("rai18_qweight.packed", {0}, "", 0, 0),
+                        pb::tensor("rai18_scales", {0}, "", 0, 0),
+                        pb::tensor("rai18_zeros", {0}, "", 0, 0)},
+                       {pb::matmulnbits("mm", {"x", "rai18_qweight", "rai18_scales", "rai18_zeros"}, 64, 512, 4, 128)}));
+    put_file((dir / "model.bin").string(), std::string(64, '\0'));
+    try {
+        (void)pi::plan(dir.string());
+        check(false, "crafted 1.8: a [0] qweight with an empty .packed sibling was refused");
+    } catch (const std::exception& e) {
+        const std::string what = e.what();
+        std::printf("  crafted 1.8 placeholder refusal: %s\n", what.c_str());
+        check(what.find("rai18_qweight (its rai18_qweight.packed sibling carries no data either)") != std::string::npos &&
+                  what.find("genuinely absent") != std::string::npos,
+              "crafted 1.8: plan() refuses by name, naming the placeholder and its sibling (" + what + ")");
+        check(what.find("weights are in its model.bin") != std::string::npos,
+              "crafted 1.8: and says where the weights are when a model.bin is there (" + what + ")");
+    }
+    fs::remove(dir / "model.bin");
+    try {
+        (void)pi::plan(dir.string());
+        check(false, "crafted 1.8 without model.bin: refused");
+    } catch (const std::exception& e) {
+        check(std::string(e.what()).find("model.bin") == std::string::npos,
+              "crafted 1.8: no model.bin, so the refusal does not claim one (" + std::string(e.what()) + ")");
+    }
     fs::remove_all(dir);
 }