rob thijssen
7ebcfba5ca
All checks were successful
CI / CUDA type-check (push) Successful in 33s
build-prerelease / Resolve version stamps (push) Successful in 32s
CI / Format (push) Successful in 33s
CI / Clippy (push) Successful in 2m34s
build-prerelease / Build cortex binary (push) Successful in 4m33s
build-prerelease / Build neuron-blackwell (push) Successful in 6m14s
CI / Test (push) Successful in 6m50s
CI / Build cortex SRPM (push) Has been skipped
CI / Publish cortex to COPR (push) Has been skipped
CI / Build neuron SRPM (push) Has been skipped
CI / Publish neuron to COPR (push) Has been skipped
CI / Bump version in source (push) Has been skipped
build-prerelease / Build neuron-ampere (push) Successful in 8m12s
build-prerelease / Package cortex RPM (push) Successful in 1m23s
build-prerelease / Build neuron-ada (push) Successful in 5m9s
build-prerelease / Package helexa-neuron-ada RPM (push) Successful in 2m59s
build-prerelease / Package helexa-neuron-ampere RPM (push) Successful in 3m3s
build-prerelease / Package helexa-neuron-blackwell RPM (push) Successful in 3m52s
build-prerelease / Publish to rpm.lair.cafe (unstable) (push) Successful in 1m2s
On beast the interleaved M-RoPE degraded image understanding rather than fixing it: the model misread spatial layout (a horizontal row of people described as a "diagonal receding line"), got attributes wrong, and rambled — a "how many people" follow-up generated 4459 tokens over 3.5 minutes, past agent-0's HTTP timeout (the "fails to respond without an error"). The interleave is evidently not numerically correct, and it can't be validated remotely without a transformers reference. Gate it: `get_rope_index` now returns plain sequential identity positions unless NEURON_MROPE is truthy, so mrope_cos_sin reduces to plain RoPE and image tokens behave exactly as pre-M-RoPE (content recognition works; spatial layout approximate; no rambling). The real computation moves to `compute_mrope_index` (still unit-tested). Default off restores the working vision and unblocks agent-0; the M-RoPE code stays in place to debug + validate before flipping the default on. Pure non-cuda change (rope.rs); both single-GPU and TP forwards call the gated get_rope_index unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>