refactor(framework): promote greedy Qwen decoder runtime and Qwen BPE bundle helpers

CMakeLists.txt

@@ -419,6 +419,7 @@ add_library(engine_core OBJECT
     src/framework/runtime/host_ops.cpp
     src/framework/runtime/kv_cache.cpp
     src/framework/runtime/bounded_static_kv_decode.cpp
+    src/framework/runtime/greedy_qwen_decoder.cpp
     src/framework/runtime/options.cpp
     src/framework/runtime/session_base.cpp
     src/framework/runtime/workspace.cpp
@@ -500,6 +501,7 @@ add_library(engine_core OBJECT
     src/framework/modules/optimizations/fast_kv_modules.cpp
     src/framework/tokenizers/hf_tokenizer_json.cpp
     src/framework/tokenizers/llama_bpe.cpp
+    src/framework/tokenizers/qwen_bpe_bundle.cpp
     external/llama_tokenizer/bpe-core.cpp
     external/llama_tokenizer/unicode.cpp
     external/llama_tokenizer/unicode-data.cpp

include/engine/community_models/audio8_asr/thinker.h

@@ -2,16 +2,17 @@
 
 #include "engine/framework/assets/tensor_source.h"
 #include "engine/framework/core/execution_context.h"
+#include "engine/framework/runtime/greedy_qwen_decoder.h"
 #include "engine/community_models/audio8_asr/types.h"
 
 #include <cstddef>
 #include <memory>
 
 namespace engine::community_models::audio8_asr {
 
-// Greedy causal decoder for the Audio8 8-layer Qwen2-style LM. Audio
-// embeddings are injected into the token embedding sequence at the prompt's
-// audio placeholder positions before prefill.
+// The Audio8 8-layer Qwen2-style decoder, expressed through the framework's
+// shared greedy Qwen decoder runtime (prefill with audio-embedding injection
+// plus static-cache step decode). Owns only the family-specific spec.
 class Audio8ThinkerRuntime {
 public:
     Audio8ThinkerRuntime(
@@ -33,8 +34,8 @@ class Audio8ThinkerRuntime {
         const Audio8ASRGenerationOptions & options);
 
 private:
-    struct Impl;
-    std::unique_ptr<Impl> impl_;
+    runtime::GreedyQwenDecoderRuntime runtime_;
+    std::shared_ptr<const Audio8ASRDecoderConfig> config_;
 };
 
 }  // namespace engine::community_models::audio8_asr

include/engine/framework/runtime/greedy_qwen_decoder.h

@@ -0,0 +1,72 @@
+#pragma once
+
+#include "engine/framework/assets/tensor_source.h"
+#include "engine/framework/core/execution_context.h"
+#include "engine/framework/modules/transformers/qwen_causal_decoder.h"
+
+#include <cstddef>
+#include <cstdint>
+#include <memory>
+#include <string>
+#include <vector>
+
+namespace engine::runtime {
+
+// Specification for a greedy Qwen-family causal decoder: how to find its
+// tensors in a weight source and how the shared QwenCausalDecoder stack is
+// configured. Covers Qwen2-style decoders (attention biases, no Q/K norms,
+// as in Audio8-ASR) and Qwen3-style decoders (Q/K norms, no attention
+// biases, as in the qwen3_asr thinker), with separate or packed QKV
+// projections and tied or separate LM heads.
+struct GreedyQwenDecoderSpec {
+    modules::QwenCausalDecoderConfig decoder;
+    int64_t vocab_size = 0;
+    int64_t max_position_embeddings = 0;
+    bool tie_word_embeddings = false;
+    bool attention_bias = false;
+    bool packed_qkv = false;
+    std::string token_embedding_tensor;
+    std::string lm_head_tensor;  // used when !tie_word_embeddings
+    std::string final_norm_tensor;
+    std::string layer_prefix;  // e.g. "language_model.model.layers"
+    std::vector<int64_t> eos_token_ids;
+};
+
+// Greedy autoregressive decoding over a Qwen-style decoder stack: prefill
+// with optional audio-embedding injection (ggml_set_rows at prompt
+// positions) and static-cache step decode, hiding the graph lifetime and
+// K/V state handoff that model families otherwise duplicate.
+class GreedyQwenDecoderRuntime {
+public:
+    struct Injection {
+        std::vector<float> values;        // tokens * hidden, token-major
+        int64_t tokens = 0;
+        std::vector<int32_t> positions;   // prompt positions to replace
+    };
+
+    struct Prompt {
+        std::vector<int32_t> input_ids;
+        Injection injection;  // optional
+    };
+
+    GreedyQwenDecoderRuntime(
+        std::shared_ptr<const assets::TensorSource> weights_source,
+        const GreedyQwenDecoderSpec & spec,
+        core::ExecutionContext & execution,
+        size_t prefill_graph_arena_bytes,
+        size_t decode_graph_arena_bytes,
+        size_t weight_context_bytes,
+        assets::TensorStorageType weight_storage_type);
+    ~GreedyQwenDecoderRuntime();
+
+    GreedyQwenDecoderRuntime(const GreedyQwenDecoderRuntime &) = delete;
+    GreedyQwenDecoderRuntime & operator=(const GreedyQwenDecoderRuntime &) = delete;
+
+    std::vector<int32_t> generate(const Prompt & prompt, int64_t max_new_tokens);
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> impl_;
+};
+
+}  // namespace engine::runtime

include/engine/framework/tokenizers/qwen_bpe_bundle.h

@@ -0,0 +1,22 @@
+#pragma once
+
+#include "engine/framework/assets/resource_bundle.h"
+#include "engine/framework/tokenizers/llama_bpe.h"
+
+#include <memory>
+#include <string_view>
+
+namespace engine::tokenizers {
+
+// Load the Qwen2-pretokenized BPE tokenizer referenced by a model bundle:
+// tokenizer_config.json plus vocab.json/merges.txt or tokenizer.json.
+std::shared_ptr<LlamaBpeTokenizer> load_qwen_bpe_tokenizer(
+    const engine::assets::ResourceBundle & bundle);
+
+// Look up a special token's id from the bundle's tokenizer.json
+// added_tokens list; throws when the token is absent.
+int64_t require_added_token_id(
+    const engine::assets::ResourceBundle & bundle,
+    std::string_view content);
+
+}  // namespace engine::tokenizers

src/community_models/audio8_asr/assets.cpp

@@ -3,6 +3,7 @@
 #include "engine/framework/assets/tensor_source.h"
 #include "engine/framework/io/json.h"
 #include "engine/framework/model_spec/package.h"
+#include "engine/framework/tokenizers/qwen_bpe_bundle.h"
 
 #include <algorithm>
 #include <stdexcept>
@@ -126,19 +127,6 @@ qwen3_asr::Qwen3ASRAudioEncoderConfig parse_audio_encoder_config(const json::Val
     return config;
 }
 
-int64_t require_added_token_id(const assets::ResourceBundle & resources, std::string_view content) {
-    const auto tokenizer = resources.parse_json("tokenizer_json");
-    for (const auto & item : tokenizer.require("added_tokens").as_array()) {
-        const auto * token_content = item.find("content");
-        const auto * token_id = item.find("id");
-        if (token_content != nullptr && token_content->is_string() &&
-            token_id != nullptr && token_id->is_number() && token_content->as_string() == content) {
-            return token_id->as_i64();
-        }
-    }
-    throw std::runtime_error("Audio8 ASR tokenizer.json is missing token: " + std::string(content));
-}
-
 Audio8ASRConfig parse_config(const assets::ResourceBundle & resources) {
     const auto root = resources.parse_json("config");
 
@@ -199,11 +187,11 @@ Audio8ASRConfig parse_config(const assets::ResourceBundle & resources) {
 
     // Prompt special tokens live in tokenizer.json added_tokens; the audio
     // token id from the config must match the tokenizer entry.
-    config.user_token_id = require_added_token_id(resources, "<|user|>");
-    config.begin_audio_token_id = require_added_token_id(resources, "<|begin_of_audio|>");
-    config.end_audio_token_id = require_added_token_id(resources, "<|end_of_audio|>");
-    config.assistant_token_id = require_added_token_id(resources, "<|assistant|>");
-    config.text_decoder.audio_token_id = require_added_token_id(resources, "<|audio|>");
+    config.user_token_id = engine::tokenizers::require_added_token_id(resources, "<|user|>");
+    config.begin_audio_token_id = engine::tokenizers::require_added_token_id(resources, "<|begin_of_audio|>");
+    config.end_audio_token_id = engine::tokenizers::require_added_token_id(resources, "<|end_of_audio|>");
+    config.assistant_token_id = engine::tokenizers::require_added_token_id(resources, "<|assistant|>");
+    config.text_decoder.audio_token_id = engine::tokenizers::require_added_token_id(resources, "<|audio|>");
 
     config.supported_languages = {
         "Chinese", "English", "Cantonese", "French", "German", "Japanese", "Korean"};
@@ -267,19 +255,7 @@ std::shared_ptr<const Audio8ASRAssets> load_audio8_asr_assets(const std::filesys
     encoder_assets->model_weights = std::move(encoder_source);
     assets->encoder_assets = std::move(encoder_assets);
 
-    engine::tokenizers::LlamaBpeTokenizerSpec tokenizer_spec;
-    tokenizer_spec.tokenizer_config_path = assets->resources.require_file("tokenizer_config");
-    if (const auto * path = assets->resources.find_file("vocab")) {
-        tokenizer_spec.vocab_path = *path;
-    }
-    if (const auto * path = assets->resources.find_file("merges")) {
-        tokenizer_spec.merges_path = *path;
-    }
-    if (const auto * path = assets->resources.find_file("tokenizer_json")) {
-        tokenizer_spec.tokenizer_json_path = *path;
-    }
-    tokenizer_spec.pre_type = engine::tokenizers::LlamaBpePreTokenizer::Qwen2;
-    assets->tokenizer = engine::tokenizers::load_llama_bpe_tokenizer(tokenizer_spec);
+    assets->tokenizer = engine::tokenizers::load_qwen_bpe_tokenizer(assets->resources);
     return assets;
 }