ggml-rpc: fix 32-bit ARM (ILP32) serialization bugs

[rovmo]

Fix RPC backend on 32-bit ARM (ILP32) platforms

Problem

The RPC backend is completely non-functional on 32-bit ARM (armv7, ILP32) platforms. Any attempt to use --rpc causes either a silent hang, server-side "Expected HELLO command" errors, set_tensor out-of-bounds crashes, or graph_compute failures with corrupted tensor IDs.

Root cause: several places in the RPC serialization code assume size_t and pointers are 8 bytes. On 32-bit platforms size_t is 4 bytes and pointers are 4 bytes, causing the wire protocol to produce undersized or misaligned data that the receiving side (which reads fixed uint64_t fields) cannot parse.

Fixes

1. send_msg / send_rpc_cmd: message size field is 4 bytes instead of 8

send_msg() sends sizeof(msg_size) bytes for the size header, where msg_size is size_t (4 bytes on ILP32). The receiver recv_msg() reads uint64_t (8 bytes). The receiver blocks forever waiting for the remaining 4 bytes, causing a silent hang on the very first RPC command (HELLO).

2. ggml_backend_rpc_buffer_set_tensor: offset field is 4 bytes instead of 8

The offset parameter is size_t but the wire format specifies 8 bytes. The memcpy copies only sizeof(offset) = 4 bytes, and the subsequent data copy starts at the wrong position. The server reads a garbage 64-bit offset and rejects with "out of buffer bounds".

3. serialize_graph: node pointer memcpy reads 8 bytes from a 4-byte pointer

memcpy(dest, &cgraph->nodes[i], sizeof(uint64_t)) reads 8 bytes starting at the address of a 4-byte pointer. The extra 4 bytes are adjacent memory, producing corrupted tensor IDs that fail with "failed to create graph node".

4. serialize_tensor: zero-initialize rpc_tensor struct

Move memset before the null check so all serializations start zeroed, preventing uninitialized upper bytes in uint64_t fields assigned from narrower types.

Platforms affected

All 32-bit platforms where sizeof(size_t) == 4:

• ARMv7 / armhf (Android Termux, Raspberry Pi 32-bit, etc.)
• Potentially i686 / x86 32-bit builds

Testing

Tested on Android ARMv7 (Termux, clang 21) with Qwen2.5-0.5B-Instruct Q4_K_M:

• Before: llama-cli --rpc hangs indefinitely with zero output
• After: model loads and runs inference successfully over local RPC

[ggml-gh-bot]

Hi @rovmo, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• AI-generated content: This project does not accept PRs, descriptions or commit messages that are fully or predominantly AI-generated. If you have used AI to assist you in writing code, please make sure to disclose that explicitly.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.