fix(qairt): name recognized dsp_arch values when device creation fails

[MAN$I VERMA (mansiverma897993)]

Fixes #1254.

A bundle whose htp_backend_ext_config.json specifies a dsp_arch this QAIRT release doesn't recognize (e.g. "v85") fails with a bare QNN error 1008, which reads as unsupported silicon.

This adds dsp_arch_diagnostic() to the qairt runtime utils: it scans the config (schema-agnostic walk) and produces a message naming the offending value and the recognized set (v68-v81). Both LLM and VLM create paths log it as a warning, and on pipeline-creation failure return PARAM_NOT_SUPPORTED with the diagnostic instead of the bare error.

Not a hard gate: a missing/unparseable config or a QAIRT that accepts the value changes nothing. Test report in the comment below.

[MAN$I VERMA (mansiverma897993)]

Test report (logic verification; no Snapdragon 8 Elite Gen 5 hardware on my side)

• Walk/validation logic checked against 9 cases: the exact QAIRT: unrecognized dsp_arch (e.g. v85) fails with bare error 1008 instead of naming supported values #1254 config (v85 under devices -> flagged), v81/v73 pass untouched, nested graphs schema, duplicate dedup, numeric dsp_arch skipped, missing key / invalid JSON / empty object -> no diagnostic.
• clang-format 18.1.8 (CI's version): all three files clean.
• Compile is covered by CI. The failure-path message on a real v85-exported bundle would be good to confirm on-device.

[zdc (Hahahaooovvv)]

Hi MAN$I VERMA (@mansiverma897993) , Compiled this branch for arm64-android-snapdragon-release (same toolchain image CI uses) and plugins/qairt/src/llm.cpp.o / vlm.cpp.o both fail — 20 errors each, e.g. error: no member named 'nlohmann' in the global namespace.

Root cause: qnn_runtime_utils.h's new #include "external/json.hpp" (standard nlohmann) collides with QAIRT's own vendored JSON fork (qnn-api/src/utils/detail/json.hpp, namespace renamed nlohmann → qualla). Both headers guard identical macro names (NLOHMANN_JSON_NAMESPACE etc.) with the same include-guard idiom — whichever gets included first through llm.cpp's include chain (llm.h → llm_spec_loader.h → input_provider.h → graph.h → QnnApi.hpp → config/Config.hpp) locks the macros and breaks the other's type definitions.

Fix: use the qualla::json fork already used elsewhere in this codebase (e.g. core/src/llm/llm_spec_loader.cpp) instead of pulling in a second, conflicting JSON library:

--- a/sdk/plugins/qairt/include/qnn_runtime_utils.h
+++ b/sdk/plugins/qairt/include/qnn_runtime_utils.h
@@ -16,8 +16,8 @@
 #include <string>
 #include <vector>
 
-#include "external/json.hpp"
 #include "types.h"
+#include "utils/detail/json.hpp"
 
 namespace geniex::qairt::runtime {
 
@@ -48,16 +48,16 @@ inline std::optional<std::string> dsp_arch_diagnostic(const std::string& htp_con
     if (!file.is_open()) {
         return std::nullopt;
     }
-    const nlohmann::json root = nlohmann::json::parse(file, nullptr, /*allow_exceptions=*/false);
+    const qualla::json root = qualla::json::parse(file, nullptr, /*allow_exceptions=*/false);
     if (root.is_discarded()) {
         return std::nullopt;
     }
 
     // Exporters nest dsp_arch differently, so walk the whole document.
-    std::vector<std::string>           unrecognized;
-    std::vector<const nlohmann::json*> stack{&root};
+    std::vector<std::string>         unrecognized;
+    std::vector<const qualla::json*> stack{&root};
     while (!stack.empty()) {
-        const nlohmann::json* node = stack.back();
+        const qualla::json* node = stack.back();
         stack.pop_back();
         if (node->is_object()) {
             for (auto it = node->begin(); it != node->end(); ++it) {

With this change, compiles clean and verified end-to-end on a real Snapdragon 8 Elite Gen 5 (SM8850) device via QDC: patched htp_backend_ext_config.json to dsp_arch: "v85" and confirmed the WARN fires with the expected message, and pipeline creation fails with PARAM_NOT_SUPPORTED (code -100016) instead of the bare 1008. Control run with the original v81 value shows no false-positive diagnostic.

[MAN$I VERMA (mansiverma897993)]

Hi MAN$I VERMA (MAN$I VERMA (@mansiverma897993)) , Compiled this branch for arm64-android-snapdragon-release (same toolchain image CI uses) and plugins/qairt/src/llm.cpp.o / vlm.cpp.o both fail — 20 errors each, e.g. error: no member named 'nlohmann' in the global namespace.

Root cause: qnn_runtime_utils.h's new #include "external/json.hpp" (standard nlohmann) collides with QAIRT's own vendored JSON fork (qnn-api/src/utils/detail/json.hpp, namespace renamed nlohmann → qualla). Both headers guard identical macro names (NLOHMANN_JSON_NAMESPACE etc.) with the same include-guard idiom — whichever gets included first through llm.cpp's include chain (llm.h → llm_spec_loader.h → input_provider.h → graph.h → QnnApi.hpp → config/Config.hpp) locks the macros and breaks the other's type definitions.

Fix: use the qualla::json fork already used elsewhere in this codebase (e.g. core/src/llm/llm_spec_loader.cpp) instead of pulling in a second, conflicting JSON library:

--- a/sdk/plugins/qairt/include/qnn_runtime_utils.h
+++ b/sdk/plugins/qairt/include/qnn_runtime_utils.h
@@ -16,8 +16,8 @@
 #include <string>
 #include <vector>
 
-#include "external/json.hpp"
 #include "types.h"
+#include "utils/detail/json.hpp"
 
 namespace geniex::qairt::runtime {
 
@@ -48,16 +48,16 @@ inline std::optional<std::string> dsp_arch_diagnostic(const std::string& htp_con
     if (!file.is_open()) {
         return std::nullopt;
     }
-    const nlohmann::json root = nlohmann::json::parse(file, nullptr, /*allow_exceptions=*/false);
+    const qualla::json root = qualla::json::parse(file, nullptr, /*allow_exceptions=*/false);
     if (root.is_discarded()) {
         return std::nullopt;
     }
 
     // Exporters nest dsp_arch differently, so walk the whole document.
-    std::vector<std::string>           unrecognized;
-    std::vector<const nlohmann::json*> stack{&root};
+    std::vector<std::string>         unrecognized;
+    std::vector<const qualla::json*> stack{&root};
     while (!stack.empty()) {
-        const nlohmann::json* node = stack.back();
+        const qualla::json* node = stack.back();
         stack.pop_back();
         if (node->is_object()) {
             for (auto it = node->begin(); it != node->end(); ++it) {

With this change, compiles clean and verified end-to-end on a real Snapdragon 8 Elite Gen 5 (SM8850) device via QDC: patched htp_backend_ext_config.json to dsp_arch: "v85" and confirmed the WARN fires with the expected message, and pipeline creation fails with PARAM_NOT_SUPPORTED (code -100016) instead of the bare 1008. Control run with the original v81 value shows no false-positive diagnostic.

Thanks zdc (@Hahahaooovvv) for the detailed root-cause and the diff! Applied it as suggested switched to the vendored qualla::json (utils/detail/json.hpp) and dropped the conflicting nlohmann include. Verified llm.cpp / vlm.cpp now compile clean with the same include chain that broke before. Force-pushed to the branch, ready for another look.

[MAN$I VERMA (mansiverma897993)]

zdc (@Hahahaooovvv) RemiliaForever (@RemiliaForever) CI fails resolved: rebased onto the latest main, and all CI runs (QC Preflight, PR Check, zizmor) are fresh on commit dbefb8b. Code is verified against current main (vendored qualla::json, PARAM_NOT_SUPPORTED). Please Approve and run the checks so they can pass.