Repository navigation
wgpu forward: read rope_theta from model config, not hardcoded 1e6 (#4832) #5503
Closed as not planned
Description
Slice of #5338. Checked against origin/main d6ea1ec (2026-10-18): still open.
Files: crates/aprender-compute/src/backends/gpu/device/linalg/wgsl_forward.rs:970 (constructor :312)
Proves it: unit test: the theta used for position>0 RoPE follows the configured value (Llama 1e4 vs Qwen 1e6)
Size: about 90 changed lines, one PR, one ticket (#5498 pull rule).
Label ready: a worker may pull this.
Activity
noahgift commentedon Oct 18, 2026
Branch fix/5503-wgpu-rope-theta @ df0db3c (main merged in). Verified:
cargo test -p aprender-compute --lib --features gpu rope_-> rope_follows_the_configured_theta + rope_matches_the_closed_form pass (the earlier WIP's filter ran 0 tests because it lacked --features gpu); cargo check -p aprender-serve -p aprender-train (gpu) and -p apr-cli --features wgpu clean. Callers set the model's rope_theta. No PR yet (cap).noahgift commentedon Oct 22, 2026
tracked in #3997