refactor: config plumbing and coding-style cleanup (no functional changes)

.claude/settings.json

@@ -0,0 +1,36 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(g++ -std=c++20 -fsyntax-only -DUSEAVX2=1 -D__WINDOWS__ -DDISABLE_ABI_CHECK=1 -DDEV_BUILD -I include test/qwen3_5_omni_npu/test.cpp)",
+      "Read(//scratch/alfxu/FastFlowLM_IRON/FLM_DLL/**)",
+      "Read(//scratch/alfxu/FastFlowLM_IRON/**)",
+      "Bash(nm -DC /scratch/alfxu/FastFlowLM/src/lib/libqwen3_5_omni_npu.so)",
+      "Bash(cd /scratch/alfxu/FastFlowLM_IRON/FLM_DLL; echo \"=== qwen3_5_omni_npu.cpp \\(public wrapper\\) ===\"; wc -l detail/qwen3_5_omni/qwen3_5_omni_npu.cpp; echo; grep -n \"prefill\\\\|payload\\\\|image_soft\\\\|soft_token\\\\|forward\\\\|say\\\\|image_payload\\\\|audio_payload\\\\|num_soft\\\\|image_token\\\\|248\\\\|preprocess\" detail/qwen3_5_omni/qwen3_5_omni_npu.cpp | head -60)",
+      "Read(//scratch/alfxu/FastFlowLM_IRON/FLM_DLL/detail/qwen3_5_omni/**)",
+      "Bash(cd /scratch/alfxu/FastFlowLM_IRON/FLM_DLL; echo \"=== thinker hpp ===\"; sed -n '1,80p' detail/qwen3_5_omni/qwen3_5_omni_thinker/qwen3_5_omni_thinker.hpp 2>/dev/null; echo \"=== grep patch constants in omni thinker/image embed ===\"; grep -rn \"PATCH_SIZE\\\\|MERGE_SIZE\\\\|SHORTEST_EDGE\\\\|LONGEST_EDGE\\\\|TEMPORAL\\\\|RESCALE\\\\|IMAGE_MEAN\\\\|IMAGE_STD\\\\|soft_token\\\\|image_soft\\\\|248\" detail/qwen3_5_omni/ include/models/qwen3_5_omni/ 2>/dev/null | grep -iv \"//\" | head -40)",
+      "Bash(cd /scratch/alfxu/FastFlowLM_IRON/FLM_DLL; echo \"=== image embed hpp ===\"; cat detail/qwen3_5_omni/qwen3_5_omni_thinker/qwen3_5_omni_image_embed/qwen3_5_omni_image_embed.hpp 2>/dev/null | head -80; echo \"=== vision_config in LM_Config / omni ===\"; grep -rn \"vision_config\\\\|_vision_config\\\\|PATCH_SIZE\\\\|patch_size\\\\|merge_size\\\\|temporal\\\\|shortest\\\\|longest\\\\|image_mean\\\\|image_std\\\\|rescale\" detail/qwen3_5_omni/ 2>/dev/null | head -40)",
+      "Bash(cd /scratch/alfxu/FastFlowLM; echo \"=== LM_Config public members ===\"; grep -n \"_json_config\\\\|_vision_config\\\\|from_pretrained\\\\|vocab_size\\\\|struct\\\\|class LM_Config\\\\|public\\\\|nlohmann\" src/include/lm_config.hpp | head -40; echo \"=== omni model config.json vision/image processor ===\"; find /scratch/alfxu -maxdepth 4 -name \"config.json\" -path \"*mni*\" 2>/dev/null | head; find /scratch/alfxu -maxdepth 4 -iname \"*omni*\" -type d 2>/dev/null | head)",
+      "Read(//scratch/alfxu/**)",
+      "Bash(python3 -c ' *)",
+      "Bash(python3 -c \"import json,sys; d=json.load\\(sys.stdin\\); print\\(json.dumps\\({k:d[k] for k in d if k in ['image_mean','image_std','rescale_factor','patch_size','temporal_patch_size','merge_size','min_pixels','max_pixels','size','shortest_edge','longest_edge']},indent=1\\)\\)\")",
+      "Read(//scratch/alfxu/FastFlowLM_IRON/FLM_DLL/detail/qwen3_5_omni/qwen3_5_omni_talker/**)",
+      "Bash(cat /scratch/alfxu/models/Qwen3.5-omni-NPU2/config.json 2>/dev/null | grep -iE \"audio|vision|image|token_id\" | head -50)",
+      "Read(//scratch/alfxu/models/Qwen3.5-omni-NPU2/**)",
+      "Bash(grep -rniE \"write_wav|write.wav|\\\\.wav|save.*wav|pcm\" /scratch/alfxu/FastFlowLM/src /scratch/alfxu/FastFlowLM_IRON 2>/dev/null | grep -viE \"\\\\.o:|Binary\" | head -40)",
+      "Bash(grep -n \"audio_end_token_id\\\\|audio_start_token_id\\\\|audio_token_id\\\\|codec_bos\\\\|codec_eos\\\\|tts\" /scratch/alfxu/models/Qwen3.5-omni-NPU2/config.json)",
+      "Bash(grep -nE \"\\\\\"thinker_config\\\\\"|\\\\\"talker_config\\\\\"|\\\\\"text_config\\\\\"|\\\\\"vision_config\\\\\"|\\\\\"audio_config\\\\\"|audio_end_token_id|audio_start_token_id|audio_token_id\" /scratch/alfxu/models/Qwen3.5-omni-NPU2/config.json | head -40)",
+      "Bash(cd /scratch/alfxu/FastFlowLM/src; echo \"=== audio_reader.hpp API ===\"; grep -n \"load_audio\\\\|audio_data_t\\\\|MonoDownmix\\\\|sample_rate\\\\|num_frames\\\\|num_samples\\\\|struct\\\\|class AudioReader\" include/audio/audio_reader.hpp | head -30; echo \"=== omni engine resample rate / audio config ===\"; grep -rn \"resample\\\\|sample_rate\\\\|16000\\\\|Audio_resample\" /scratch/alfxu/FastFlowLM_IRON/FLM_DLL/detail/qwen3_5_omni/qwen3_5_omni_thinker/qwen3_5_omni_audio_embed/*.cpp 2>/dev/null | head)",
+      "Read(//scratch/alfxu/FastFlowLM_IRON/FLM_DLL/detail/qwen3_5_omni/qwen3_5_omni_thinker/qwen3_5_omni_audio_embed/**)",
+      "Bash(cd /scratch/alfxu/FastFlowLM_IRON/FLM_DLL/detail/qwen3_5_omni/qwen3_5_omni_thinker/qwen3_5_omni_audio_embed; ls; echo \"=== audio embed hpp ===\"; cat *.hpp 2>/dev/null | head -60)",
+      "Bash(grep -rn \"say\\\\|talker\\\\|codec\\\\|thinker_res\" /scratch/alfxu/FastFlowLM/src/include/models/qwen3_5_omni/qwen3_5_omni.hpp; echo \"===say impl===\"; grep -rn \"::say\\\\|buffer<bf16> say\\\\|say\\(\" /scratch/alfxu/FastFlowLM/src /scratch/alfxu/FastFlowLM_IRON 2>/dev/null | grep -v \"\\\\.o:\" | head)",
+      "Bash(make clean *)",
+      "Bash(cd /scratch/alfxu/FastFlowLM/src; grep -rn \"model.q4nx\\\\|\\\\.q4nx\\\\|safetensors\" common/tensor_utils/*.cpp 2>/dev/null | head; echo \"=== how other models' model dirs look \\(do they have model.q4nx?\\) ===\"; ls /scratch/alfxu/models/Qwen3.5-4B-NPU2/ 2>/dev/null | head)",
+      "Read(//scratch/alfxu/models/Qwen3.5-4B-NPU2/**)",
+      "Bash(g++ -std=c++20 -fsyntax-only -I include -I . common/qwen3_5_omni/modeling_qwen3_5_omni.cpp)",
+      "Bash(g++ -std=c++20 -fsyntax-only -I include -I . test/qwen3_5_omni_npu/test.cpp)",
+      "Bash(g++ -std=c++20 -fsyntax-only -I include -I . common/AutoModel/automodel.cpp)",
+      "Bash(g++ -std=c++20 -fsyntax-only -D__FLM_VERSION__='\"test\"' -I include -I . test/qwen3_5_omni_npu/test.cpp)",
+      "Bash(g++ -std=c++20 -fsyntax-only -I include -I . common/AutoModel/modeling_qwen3_5_omni.cpp)"
+    ]
+  }
+}

src/common/AutoModel/automodel.cpp

@@ -483,7 +483,7 @@ void AutoModel::set_sampler(sampler_config& sampler_config) {
     if (this->sampler != nullptr) {
         this->sampler.reset();
     }
-    this->sampler = std::make_unique<Sampler>(this->lm_config->vocab_size, sampler_config);
+    this->sampler = std::make_unique<Sampler>(this->lm_config->get("vocab_size"), sampler_config);
 }
 
 /// \brief Set the max length
@@ -575,9 +575,9 @@ void AutoModel::set_topk(int topk) {
         header_print("WARNING", "Top-k must be greater than 0");
         return;
     }
-    if (topk > this->lm_config->vocab_size) {
-        header_print("WARNING", "Top-k is greater than vocab size, set to vocab size: " << this->lm_config->vocab_size);
-        topk = this->lm_config->vocab_size;
+    if (topk > this->lm_config->get("vocab_size")) {
+        header_print("WARNING", "Top-k is greater than vocab size, set to vocab size: " << this->lm_config->get("vocab_size"));
+        topk = this->lm_config->get("vocab_size");
     }
 
     this->sampler->top_k = topk;

src/common/AutoModel/modeling_qwen2.cpp

@@ -16,7 +16,7 @@ void Qwen2::load_model(std::string model_path, json model_info, int default_cont
 
     this->q4nx = std::make_unique<Q4NX>(this->model_path);
 
-    // lm_config->model_type == qwen2
+    // lm_config->get<std::string>("model_type", "") == qwen2
     this->lm_engine = std::make_unique<qwen2_npu>(*this->lm_config, this->npu.get(), this->MAX_L);
 
     this->lm_engine->load_weights(*this->q4nx);

src/common/AutoModel/modeling_qwen2vl.cpp

@@ -18,7 +18,7 @@ void Qwen2VL::load_model(std::string model_path, json model_info, int default_co
     this->_shared_load_model(model_path, model_info, default_context_length, enable_preemption);
 
     this->q4nx = std::make_unique<Q4NX>(this->model_path);
-    // lm_config->model_type == qwen2
+    // lm_config->get<std::string>("model_type", "") == qwen2
     this->lm_engine = std::make_unique<qwen2vl_npu>(*this->lm_config, this->npu.get(), this->MAX_L);
 
     this->lm_engine->load_weights(*this->q4nx);

src/common/AutoModel/modeling_qwen3.cpp

@@ -16,7 +16,7 @@ void Qwen3::load_model(std::string model_path, json model_info, int default_cont
     this->_shared_load_model(model_path, model_info, default_context_length, enable_preemption);
 
     this->q4nx = std::make_unique<Q4NX>(this->model_path);
-    // lm_config->model_type == qwen3
+    // lm_config->get<std::string>("model_type", "") == qwen3
     this->lm_engine = std::make_unique<qwen3_npu>(*this->lm_config, this->npu.get(), this->MAX_L);
 
     this->lm_engine->load_weights(*this->q4nx);
@@ -314,7 +314,7 @@ void Qwen3_IT::load_model(std::string model_path, json model_info, int default_c
 
     this->q4nx = std::make_unique<Q4NX>(this->model_path);
 
-    // lm_config->model_type == qwen3
+    // lm_config->get<std::string>("model_type", "") == qwen3
     this->lm_engine = std::make_unique<qwen3_npu>(*this->lm_config, this->npu.get(), this->MAX_L);
 
     this->lm_engine->load_weights(*this->q4nx);
@@ -465,7 +465,7 @@ void Qwen3_TK::load_model(std::string model_path, json model_info, int default_c
 
     this->q4nx = std::make_unique<Q4NX>(this->model_path);
 
-    // lm_config->model_type == qwen3
+    // lm_config->get<std::string>("model_type", "") == qwen3
     this->lm_engine = std::make_unique<qwen3_npu>(*this->lm_config, this->npu.get(), this->MAX_L);
 
     this->lm_engine->load_weights(*this->q4nx);

src/common/AutoModel/modeling_qwen3_5_omni.cpp

@@ -75,13 +75,13 @@ void Qwen3_5_Omni::load_model(std::string model_path, json model_info, int defau
     this->lm_config->from_pretrained(this->model_path);
 
     // The omni thinker vocab lives in thinker_config.text_config, so the
-    // top-level "vocab_size" is 0. Resolve it here so every downstream reader
-    // (set_sampler, set_topk, ...) sees the real vocab instead of 0.
+    // top-level "vocab_size" is 0. Republish it at the top level so every
+    // downstream reader (set_sampler, set_topk, ...) sees the real vocab.
     {
-        const auto& jc = this->lm_config->_json_config;
+        auto& jc = this->lm_config->_json_config;
         if (jc.contains("thinker_config") && jc["thinker_config"].contains("text_config")) {
-            this->lm_config->vocab_size =
-                json_u32(jc["thinker_config"]["text_config"], "vocab_size", this->lm_config->vocab_size);
+            jc["vocab_size"] =
+                json_u32(jc["thinker_config"]["text_config"], "vocab_size", this->lm_config->get("vocab_size"));
         }
     }
 
@@ -694,7 +694,7 @@ buffer<bf16> Qwen3_5_Omni::say(std::string wav_out_path) {
 
 void Qwen3_5_Omni::set_sampler(sampler_config& sampler_config) {
     if (this->sampler != nullptr) this->sampler.reset();
-    this->sampler = std::make_unique<Sampler>((int)this->lm_config->vocab_size, sampler_config);
+    this->sampler = std::make_unique<Sampler>((int)this->lm_config->get("vocab_size"), sampler_config);
 }
 
 void Qwen3_5_Omni::clear_context() {

src/common/AutoModel/modeling_qwen3_5vl.cpp

@@ -17,7 +17,7 @@ void Qwen3_5VL::load_model(std::string model_path, json model_info, int default_
     this->_shared_load_model(model_path, model_info, default_context_length, enable_preemption);
 
     this->q4nx = std::make_unique<Q4NX>(this->model_path);
-    // lm_config->model_type == qwen3
+    // lm_config->get<std::string>("model_type", "") == qwen3
     this->lm_engine = std::make_unique<qwen3_5vl_npu>(*this->lm_config, this->npu.get(), this->MAX_L);
 
     this->lm_engine->load_weights(*this->q4nx);

src/common/AutoModel/modeling_qwen3_6_moe.cpp

@@ -17,7 +17,7 @@ void Qwen3_6_MOE::load_model(std::string model_path, json model_info, int defaul
     this->_shared_load_model(model_path, model_info, default_context_length, enable_preemption);
 
     this->q4nx = std::make_unique<Q4NX>(this->model_path);
-    // lm_config->model_type == qwen3
+    // lm_config->get<std::string>("model_type", "") == qwen3
     this->lm_engine = std::make_unique<qwen3_6_moe_npu>(*this->lm_config, this->npu.get(), this->MAX_L);
 
     this->lm_engine->load_weights(*this->q4nx);

src/common/AutoModel/modeling_qwen3vl.cpp

@@ -17,7 +17,7 @@ void Qwen3VL::load_model(std::string model_path, json model_info, int default_co
     this->_shared_load_model(model_path, model_info, default_context_length, enable_preemption);
 
     this->q4nx = std::make_unique<Q4NX>(this->model_path);
-    // lm_config->model_type == qwen3
+    // lm_config->get<std::string>("model_type", "") == qwen3
     this->lm_engine = std::make_unique<qwen3vl_npu>(*this->lm_config, this->npu.get(), this->MAX_L);
 
     this->lm_engine->load_weights(*this->q4nx);

src/common/whisper/modeling_whisper.cpp

@@ -50,7 +50,7 @@ void Whisper::load_model(std::string model_path, nlohmann::ordered_json model_in
     s_config.min_p = 0.1;
     s_config.temperature = 0.4;
 
-    this->sampler = std::make_unique<Sampler>(this->lm_config->vocab_size, s_config);
+    this->sampler = std::make_unique<Sampler>(this->lm_config->get("vocab_size"), s_config);
 }
 
 void Whisper::setup_tokenizer(std::string model_path) {