Security context

Highlights

Denial of Service: 6 prior fixes. Scrutinize any change in this area.

megatron/core/extensions/transformer_engine.py: most-fixed (1 issue). Treat as high-risk during review.

1 high-severity fix in this history; regressions here are high-impact.

Recurring patterns

Code Injection: Using unsafe Python eval on user-controlled input strings allows arbitrary remote code execution during model configuration and training initialization. Secure implementations must strictly validate input types and use safe parsing alternatives such as ast.literal_eval.

Denial of Service: Memory management routines during KV cache scheduling and Mamba state cache allocations can cause sudden GPU out-of-memory crashes if they fail to budget for scratch spaces or fail to evict overflowed requests under memory pressure.

Denial of Service: Parsing untrusted or malformed inputs (such as truncated tool-calling structures or invalid Unicode sequences) without robust fallbacks causes unhandled exceptions (like ValueError or UnicodeDecodeError) that crash active inference sessions.