UPSTREAM PR #21082: common: add bounds check in common_init_result::sampler to prevent segfault on failed model load

[loci-dev]

Note

Source pull request: ggml-org/llama.cpp#21082

Overview

This fix supercedes #21049 to fix the segfault when non-existent models are loaded. It reverts the previous fix because it is no longer necessary with this more correct fix in place. In addition a regression test was added to expose the issue and showcase the fix.

Before:

./build/bin/test-init-fail
build: 8557 (43d7135f7) with GNU 13.3.0 for Linux x86_64
common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on
gguf_init_from_file: failed to open GGUF file '/non_existent_model.gguf' (No such file or directory)
llama_model_load: error loading model: llama_model_loader: failed to load model from /non_existent_model.gguf
llama_model_load_from_file_impl: failed to load model
llama_params_fit: encountered an error while trying to fit params to free device memory: failed to load model
llama_params_fit: fitting params to free memory took 0.00 seconds
gguf_init_from_file: failed to open GGUF file '/non_existent_model.gguf' (No such file or directory)
llama_model_load: error loading model: llama_model_loader: failed to load model from /non_existent_model.gguf
llama_model_load_from_file_impl: failed to load model
common_init_from_params: failed to load model '/non_existent_model.gguf'
Segmentation fault (core dumped)

After:

./build/bin/test-init-fail
build: 8557 (43d7135f7) with GNU 13.3.0 for Linux x86_64
common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on
gguf_init_from_file: failed to open GGUF file '/non_existent_model.gguf' (No such file or directory)
llama_model_load: error loading model: llama_model_loader: failed to load model from /non_existent_model.gguf
llama_model_load_from_file_impl: failed to load model
llama_params_fit: encountered an error while trying to fit params to free device memory: failed to load model
llama_params_fit: fitting params to free memory took 0.00 seconds
gguf_init_from_file: failed to open GGUF file '/non_existent_model.gguf' (No such file or directory)
llama_model_load: error loading model: llama_model_loader: failed to load model from /non_existent_model.gguf
llama_model_load_from_file_impl: failed to load model
main: OK
common_init_from_params: failed to load model '/non_existent_model.gguf'

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES AI was used to write the unit test

[loci-review]

Target version:

The target version shows additional execution paths for bounds validation (size() call and conditional branching) before the original pointer accessor chain, explaining the 78% response time increase.

Additional Findings

No performance-critical hot paths affected. Core inference operations (matrix operations, attention mechanisms, KV cache, quantization kernels) show 0% change. All overhead is one-time during initialization. The bounds checking prevents crashes at negligible cost (~171ns per initialization call).

🔎 Full breakdown: Loci Inspector
💬 Questions? Tag @loci-dev