starvla: add Qwen2.5-VL FAST policy

Author: JJJYmmm

src/models/starvla/fast_codec.cpp

@@ -0,0 +1,90 @@
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <filesystem>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+namespace robotcpp::starvla {
+
+struct FastCodecConfig {
+    double scale = 0.0;
+    int32_t min_token = 0;
+    size_t vocab_size = 0;
+    size_t time_horizon = 0;
+    size_t action_dim = 0;
+};
+
+struct FastDecodeResult {
+    size_t batch_size = 0;
+    size_t time_horizon = 0;
+    size_t action_dim = 0;
+    std::vector<double> actions;
+};
+
+class FastCodec {
+public:
+    static std::unique_ptr<FastCodec> create(
+        FastCodecConfig config, std::vector<std::string> vocab_by_id,
+        std::vector<int32_t> fast_to_vlm_id, std::string & error);
+
+    // Constructs directly from the converter-compiled ByteLevel pieces stored
+    // in policy GGUF. offsets has vocab_size + 1 entries and indexes the flat
+    // byte buffer; no external tokenizer JSON is consulted.
+    static std::unique_ptr<FastCodec> create_compiled(
+        FastCodecConfig config, std::vector<int32_t> token_offsets,
+        std::vector<uint8_t> token_bytes,
+        std::vector<int32_t> fast_to_vlm_id, std::string & error);
+
+    static std::unique_ptr<FastCodec> load_hf_assets(
+        const std::filesystem::path & tokenizer_json,
+        const std::filesystem::path & processor_config_json,
+        const std::filesystem::path & action_token_map_json,
+        size_t time_horizon, size_t action_dim, std::string & error);
+
+    const FastCodecConfig & config() const;
+    const std::vector<int32_t> & fast_to_vlm_ids() const;
+
+    bool map_fast_to_vlm(const std::vector<int32_t> & fast_ids,
+                         std::vector<int32_t> & vlm_ids, std::string & error) const;
+    bool map_vlm_to_fast(const std::vector<int32_t> & vlm_ids,
+                         std::vector<int32_t> & fast_ids, std::string & error) const;
+
+    // Extracts every mapped action token from a generated Qwen sequence in order.
+    // EOS stopping remains the generator's responsibility; ordinary EOS/pad/text
+    // IDs in the returned sequence are ignored and do not terminate this scan.
+    bool extract_fast_tokens(const std::vector<int32_t> & generated_ids,
+                             std::vector<int32_t> & fast_ids, std::string & error) const;
+
+    // Exposed for focused parity diagnostics. This is the Hugging Face ByteLevel
+    // decoder output before min_token adjustment and inverse DCT.
+    bool byte_level_decode(const std::vector<int32_t> & fast_ids,
+                           std::vector<uint32_t> & codepoints, std::string & error) const;
+
+    bool decode_fast_tokens(const std::vector<std::vector<int32_t>> & batch_fast_ids,
+                            FastDecodeResult & result, std::string & error) const;
+
+    // Strict low-level API: every input ID must be an action token. Use
+    // decode_generated_tokens for complete Qwen sequences containing text.
+    bool decode_vlm_action_tokens(const std::vector<std::vector<int32_t>> & batch_vlm_ids,
+                                  FastDecodeResult & result, std::string & error) const;
+
+    // Production entry point for complete Qwen generated_ids. Ordinary text and
+    // control tokens are filtered through the explicit inverse action-token map.
+    bool decode_generated_tokens(const std::vector<std::vector<int32_t>> & batch_generated_ids,
+                                 FastDecodeResult & result, std::string & error) const;
+
+private:
+    FastCodec(FastCodecConfig config, std::vector<std::vector<uint8_t>> token_bytes,
+              std::vector<int32_t> fast_to_vlm_id);
+
+    FastCodecConfig config_;
+    std::vector<std::vector<uint8_t>> token_bytes_;
+    std::vector<int32_t> fast_to_vlm_id_;
+    std::vector<std::pair<int32_t, int32_t>> vlm_to_fast_id_;
+};
+
+} // namespace robotcpp::starvla