refactor(runtime): remove Python interpreter bridges

CMakeLists.txt

@@ -232,7 +232,6 @@ add_library(trtmc_core SHARED
   src/tokenizer/wordpiece_tokenizer.cpp
   src/tokenizer/unigram_tokenizer.cpp
   # Utilities
-  src/utils/data_dir.cpp
   src/utils/text_parsers.cpp
   src/utils/json_helpers.cpp
   src/cli/jsonl_io.cpp
@@ -780,6 +779,34 @@ set_target_properties(trtmc_test_optimized_embedding PROPERTIES
   LIBRARY_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/tests/optimized-runtimes/embedding"
 )
 
+add_library(trtmc_test_model_plugin_missing_abi SHARED EXCLUDE_FROM_ALL
+  tests/cpp/fakes/fake_model_plugin_abi.cpp
+)
+target_include_directories(trtmc_test_model_plugin_missing_abi PRIVATE
+  ${PROJECT_SOURCE_DIR}/include
+)
+target_compile_definitions(trtmc_test_model_plugin_missing_abi PRIVATE
+  TRTMC_FAKE_MODEL_PLUGIN_OMIT_ABI=1
+)
+set_target_properties(trtmc_test_model_plugin_missing_abi PROPERTIES
+  OUTPUT_NAME trtmc_test_model_plugin_abi
+  LIBRARY_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/tests/model-plugin-abi/missing"
+)
+
+add_library(trtmc_test_model_plugin_wrong_abi SHARED EXCLUDE_FROM_ALL
+  tests/cpp/fakes/fake_model_plugin_abi.cpp
+)
+target_include_directories(trtmc_test_model_plugin_wrong_abi PRIVATE
+  ${PROJECT_SOURCE_DIR}/include
+)
+target_compile_definitions(trtmc_test_model_plugin_wrong_abi PRIVATE
+  TRTMC_FAKE_MODEL_PLUGIN_ABI_VERSION=1
+)
+set_target_properties(trtmc_test_model_plugin_wrong_abi PROPERTIES
+  OUTPUT_NAME trtmc_test_model_plugin_abi
+  LIBRARY_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/tests/model-plugin-abi/wrong"
+)
+
 # --- Test helper function ---
 # Usage:
 #   trtmc_add_test(test_name)               -- links trtmc_core, adds src/ include
@@ -902,6 +929,15 @@ trtmc_add_test(test_bundle_view)
 trtmc_add_test(test_pipeline_registry)
 trtmc_add_test(test_model_plugin_loader MODEL_OWNED)
 add_dependencies(test_model_plugin_loader trtmc_model_plugins)
+trtmc_add_test(test_model_plugin_abi)
+add_dependencies(test_model_plugin_abi
+  trtmc_test_model_plugin_missing_abi
+  trtmc_test_model_plugin_wrong_abi
+)
+target_compile_definitions(test_model_plugin_abi PRIVATE
+  TRTMC_TEST_MODEL_PLUGIN_MISSING_ABI_DSO="$<TARGET_FILE:trtmc_test_model_plugin_missing_abi>"
+  TRTMC_TEST_MODEL_PLUGIN_WRONG_ABI_DSO="$<TARGET_FILE:trtmc_test_model_plugin_wrong_abi>"
+)
 trtmc_add_test(test_backend_loader)
 trtmc_add_test(test_trt_version)
 trtmc_add_test(test_runtime_measurement)
@@ -940,7 +976,6 @@ trtmc_add_test(test_cabi_batch                 NO_SRC_INCLUDE)
 # Config / parsing
 trtmc_add_test(test_text_parsers)
 trtmc_add_test(test_json_helpers)
-trtmc_add_test(test_data_dir)
 trtmc_add_test(test_cli_args)
 target_sources(test_cli_args PRIVATE src/cli/args.cpp)
 trtmc_add_test(test_jsonl_dataset_and_output

benchmarks/performance/release.yaml

@@ -724,24 +724,6 @@ entries:
       runner: hf-transformers
       mode: torch-compile
       compile_scope: model.forward
-  - id: qwen3_omni.generate_audio
-    family: qwen3_omni
-    operation: generate_audio
-    model: qwen3-omni-30b-a3b-instruct
-    workload:
-      testcase: qwen3-omni-30b-a3b-instruct
-    baseline:
-      runner: task-reference
-      adapter: hf-qwen3-omni
-      mode: hf-eager
-      reference_backend: hf_transformers
-      timing_scope: task-model-call-wall
-      input_preparation_included: false
-      output_contract: audio-shape
-      adapter_options:
-        device_map: cuda:0
-        speaker: Ethan
-        talker_max_new_tokens: 32
   - id: qwen_image.generate_image
     family: qwen_image
     operation: generate_image

cmake/register_model_plugin.cpp.in

@@ -6,6 +6,7 @@
 // Generated by cmake/trtmc_pipeline_plugins.cmake. Do not edit.
 
 #include "trtmc/config/schema_registry.h"
+#include "trtmc/runtime/pipeline_plugin_loader.h"
 #include "trtmc/runtime/pipeline_registry.h"
 
 namespace trtmc {
@@ -18,6 +19,10 @@ namespace config::schemas {
 
 } // namespace trtmc
 
+extern "C" std::uint32_t trtmc_model_plugin_abi_version() {
+    return ::trtmc::kModelPluginAbiVersion;
+}
+
 extern "C" const char* trtmc_model_plugin_id() {
     return "@TRTMC_MODEL_PLUGIN_ID@";
 }

examples/trtmc_benchmark_worker.cpp

@@ -152,7 +152,6 @@ std::vector<float> optional_float32_values(const Json& request, const std::strin
 
 trtmc::LoadOptions load_options(const Json& runtime) {
     trtmc::LoadOptions options;
-    options.hf_python = optional_value<std::string>(runtime, "hf_python", "");
     options.runtime_cache_path = optional_value<std::string>(runtime, "runtime_cache_path", "");
     options.cuda_graphs = optional_value<bool>(runtime, "cuda_graphs", false);
     options.kv_cache_size_bytes = optional_value<std::uint64_t>(runtime, "kv_cache_size_bytes", 0);

examples/trtmc_dataset_benchmark.cpp

@@ -101,7 +101,7 @@ std::optional<std::string> extract_answer_from_text(const std::string& text) {
 
 void usage() {
     std::cerr << "Usage: trtmc_dataset_benchmark <bundle.bundle> <dataset.jsonl> <output.jsonl> "
-                 "[--max-new-tokens N] [--hf-python PATH] [--kv-cache-size SIZE] "
+                 "[--max-new-tokens N] [--kv-cache-size SIZE] "
                  "[--backend-dir PATH] [--model-plugin-dir PATH] "
                  "[--temperature F] [--top-k N] [--top-p F] [--min-p F] [--seed N] "
                  "[--chat-template] [--no-thinking] [--stop-on-answer] "
@@ -153,8 +153,6 @@ int main(int argc, char** argv) {
             set_tokens.emplace_back(need_value(arg));
         } else if (arg == "--max-new-tokens") {
             max_new_tokens = std::stoi(need_value(arg));
-        } else if (arg == "--hf-python") {
-            load_options.hf_python = need_value(arg);
         } else if (arg == "--kv-cache-size") {
             load_options.kv_cache_size_bytes = parse_size_bytes(need_value(arg));
         } else if (arg == "--backend-dir") {

include/trtmc/pipeline.h

@@ -691,9 +691,8 @@ class IPipeline {
 
 // --- Factory ---
 // LoadOptions bundles every knob the factory understands. Users who only want
-// the defaults can still call the positional overload below.
+// the defaults can call the single-argument overload below.
 struct LoadOptions {
-    std::string hf_python;
     std::string runtime_cache_path;
     bool cuda_graphs{false};
     std::uint64_t kv_cache_size_bytes{0};               // 0 = use bundle's max_cache_length
@@ -703,9 +702,7 @@ struct LoadOptions {
     std::vector<std::string> model_plugin_search_paths; // Extra dirs for libtrtmc_model_*.so
 };
 
-std::unique_ptr<IPipeline> load(const std::string& bundle_path, const std::string& hf_python = "",
-                                const std::string& runtime_cache_path = "",
-                                bool cuda_graphs = false);
+std::unique_ptr<IPipeline> load(const std::string& bundle_path);
 std::unique_ptr<IPipeline> load(const std::string& bundle_path, const LoadOptions& options);
 std::unique_ptr<IPipeline> load(const std::string& bundle_path, const LoadOptions& options,
                                 const std::string& kernel_bindings_path);
@@ -718,7 +715,6 @@ extern "C" {
 
 struct TrtmcPipelineOptions {
     int max_new_tokens;        // 0 = use model default
-    const char* hf_python;     // nullptr = auto-detect
     const char* image_path;    // nullptr = text-only
     const char* runtime_cache; // nullptr = no RTX cache
     int cuda_graphs;           // 0 = disabled

include/trtmc/runtime/pipeline_factory.h

@@ -23,10 +23,7 @@ class PipelinePool;
 
 class PipelineFactory {
   public:
-    static std::unique_ptr<IPipeline> from_bundle(const std::string& bundle_path,
-                                                  const std::string& hf_python = "",
-                                                  const std::string& runtime_cache_path = "",
-                                                  bool cuda_graphs = false);
+    static std::unique_ptr<IPipeline> from_bundle(const std::string& bundle_path);
     static std::unique_ptr<IPipeline> from_bundle(const std::string& bundle_path,
                                                   const LoadOptions& options);
     static std::unique_ptr<IPipeline> from_bundle(const std::string& bundle_path,

include/trtmc/runtime/pipeline_plugin.h

@@ -73,7 +73,6 @@ struct PipelineContext {
     const BundleFile& bundle;
     const BaseConfig& config;
     const std::string& config_json;        // raw JSON text from bundle
-    const std::string& hf_python;          // path to HF Python interpreter
     const std::string& bundle_path;        // filesystem path to .bundle file
     IBackend* backend;                     // Backend for creating ITrtModule instances
     const std::string& runtime_cache_path; // RTX: JIT kernel cache file path

include/trtmc/runtime/pipeline_plugin_loader.h

@@ -5,12 +5,17 @@
 
 #pragma once
 
+#include <cstdint>
 #include <optional>
 #include <string>
 #include <vector>
 
 namespace trtmc {
 
+// Exact-match epoch: pre-1.0 runtimes accept only model DSOs built for this version.
+inline constexpr std::uint32_t kModelPluginAbiVersion = 2;
+inline constexpr const char* kModelPluginAbiEntrypoint = "trtmc_model_plugin_abi_version";
+
 struct ModelPluginInfo {
     const char* model_id;
     const char* runtime_strategy;

plugins/trtmc-agent-skills/skills/debug-trt-mismatch/SKILL.md

@@ -124,7 +124,6 @@ PYTHONPATH=python:. python3 tools/diff_vl.py \
   --image /path/to/test.jpg \
   --model <model> \
   --binary ./build/trtmc \
-  --hf-python <python> \
   --debug-layers
 ```
 
@@ -135,7 +134,6 @@ PYTHONPATH=python:. python3 tools/diff_audio.py \
   --bundle /tmp/model.bundle \
   --binary ./build/trtmc \
   --model <model> \
-  --hf-python <python> \
   --stage 1
 ```
 
@@ -162,7 +160,6 @@ When Python/reference parity passes but the C++ result differs:
 PYTHONPATH=python:. python3 tools/test_runner_parity.py \
   --bundle /tmp/model.bundle \
   --binary ./build/trtmc \
-  --hf-python <python> \
   --prompt "The capital of France is" \
   --max-new-tokens 20
 ```

plugins/trtmc-agent-skills/skills/profile-model/SKILL.md

@@ -74,7 +74,6 @@ PYTHONPATH=python:. python3 tools/trtmc_profile.py \
   --iterations 10 \
   --dtype float16 \
   --trtmc-binary ./build/trtmc \
-  --hf-python <python> \
   --json \
   --output-dir <profile-dir>
 ```