fix: add virtual destructor to npu_cmd to prevent heap corruption (affects v0.9.42+)

[rmanicardi]

Summary

• npu_cmd declares pure virtual methods but has no virtual destructor
• npu_sequence owns a std::vector<std::unique_ptr<npu_cmd>> of heterogeneous derived objects (e.g. npu_write_cmd, 48 bytes)
• When the vector is cleared/destroyed, unique_ptr<npu_cmd> calls operator delete(ptr, sizeof(npu_cmd)) = operator delete(ptr, 8) — wrong-sized deallocation of a 48-byte object, corrupting glibc heap metadata → SIGABRT
• Also adds #include <climits> which is required for UCHAR_MAX used in the op_headers enum (XAIE_IO_CUSTOM_OP_MAX = UCHAR_MAX)

Affected versions

Version	Behaviour
≤ 0.9.40	Latent — npu_write_cmd was ~8 bytes, so wrong-sized delete freed the correct amount by accident
0.9.42, 0.9.43	Fatal — npu_write_cmd is now 48 bytes; wrong-sized delete corrupts allocator chunk metadata → SIGABRT

Root cause

// npu_cmd has pure virtual methods but no virtual destructor:
struct npu_cmd {
    virtual int print_cmd(...) = 0;
    virtual void to_npu(...) = 0;
    virtual void dump_cmd(...) = 0;
    virtual int get_op_lines() = 0;
    // virtual ~npu_cmd() = default;  ← MISSING
};

Deleting through a base pointer without a virtual destructor is undefined behaviour per C++ [expr.delete] §3 when the dynamic type differs from the static type.

Crash — reproduced on v0.9.43 (Fedora 44 / GCC 14 / ASan)

==12869==ERROR: AddressSanitizer: new-delete-type-mismatch on 0x7b3d1bc5bf90 in thread T0:
  object passed to delete has wrong type:
  size of the allocated type:   48 bytes;
  size of the deallocated type: 8 bytes.
    #0  operator delete(void*, unsigned long)          (libasan.so.8)
    #1  npu_sequence::clear_cmds()                     (libqwen2_npu.so)
    #2  qwen3_5vl_npu_sequence::_gen_sequence(...)     (libqwen3_5vl_npu.so)
    #3  qwen3_5vl_npu::Impl::set_context_length(...)   (libqwen3_5vl_npu.so)
    #4  qwen3_5vl_npu::Impl::clear_context()           (libqwen3_5vl_npu.so)
    #5  Qwen3_5VL::load_model(...)                     (flm)
    #6  RestHandler::ensure_model_loaded(...)           (flm)
    #7  RestHandler::RestHandler(...)                   (flm)
    #8  create_lm_server(...)                           (flm)
    #9  main                                            (flm)

0x7b3d1bc5bf90 is located 0 bytes inside of 48-byte region
allocated by thread T0 here:
    #0  operator new(unsigned long)                    (libasan.so.8)
    #1  npu_sequence::rtp_write(...)                   (libllama_npu.so)
    #2  qwen3_5vl_npu_sequence::_gen_sequence(...)     (libqwen3_5vl_npu.so)
    #3  qwen3_5vl_npu::Impl::set_context_length(...)   (libqwen3_5vl_npu.so)

Trigger: flm serve qwen3.5:4b --ctx-len 4096 — process aborts during model loading before serving any request.

Fix

struct npu_cmd {
    virtual ~npu_cmd() = default;   // ← add this line

    virtual int print_cmd(...) = 0;
    virtual void to_npu(...) = 0;
    virtual void dump_cmd(...) = 0;
    virtual int get_op_lines() = 0;
};

Note: all pre-built NPU libraries must be recompiled against this updated header for the fix to take effect in their translation units.

Test plan

• Build flm with ASan (-fsanitize=address)
• flm serve qwen3.5:4b --ctx-len 4096 — confirm no new-delete-type-mismatch at startup
• Send inference request — confirm no crash during prefill
• Rebuild all pre-built NPU .so libraries against the updated header

🤖 Generated with Claude Code

[astrelsky]

This is clearly correct.