mirror of
https://github.com/hanzoai/engine.git
synced 2026-11-20 14:17:52 +00:00
The ROCm decode-graph path said it mirrored the CUDA gate and had dropped one line of it: CUDA and Vulkan both return early when a draft proposer is attached, ROCm did not. A draft reads what the target's layer loop records as it runs (spec_capture): host-side code that truncates the buffer to the step's start, appends rows and keys them per sequence. A graph replay re-issues only the recorded kernels, so none of that bookkeeping runs and the draft reads rows that were never written. The path is reached in practice — evo logged "ROCm decode graphs disabled after capture/replay error" with DFlash 2 attached. Only a run with a proposer changes path, and it goes to the eager decode that is always correct; nothing without speculation is touched. try_rocm_decode_graph_forward is #[cfg(feature = "rocm")], so this line is not compiled off-ROCm. It is the same statement the CUDA path carries, on the same struct, and draft_proposer is an unconditional field; the default build is checked clean. Co-authored-by: Hanzo Dev <[email protected]>
hanzo-engine
Core crate of hanzo including the models and associated executors.
Documentation: https://docs.rs/hanzo-engine/