Expand description
Highway taps for the qwen4_exp bisect.
Every PIECE of this port is pinned to the reference — the n-gram ids are bit-exact, the mHC kernels and the PLE gate/conv match to cosine 0.99999, the NVMe gather is bit-exact — and the model still does not produce coherent text. That combination says the fault is in the COMPOSITION, and composition is exactly what per-kernel probes cannot see.
So: dump the hc_mult-wide residual highway at named points and diff it
against the reference layer by layer.
The sub-layer boundary is what makes this affordable. Tapping after a
block’s hc_post but BEFORE the next hc_pre means the reference only has
to reproduce that block — for layer 0 that is the GDN projections alone,
with none of the 512-expert MoE. Only the taps that come after an MoE need
experts, and even then top-10 routing over a short prompt touches a few
dozen, not 512.
Off unless ATLAS_QWEN4EXP_DUMP names a directory. Writes
<dir>/L{layer:02}_{tag}.bin as raw little-endian FP32, [T, hc*H].
Functions§
- tap_
bf16 - Tap a BF16 buffer (the embedding, a block output) the same way.
- tap_f32
- Tap an FP32 buffer of
n_elements(the injection vector, a gate). - tap_
highway - Tap the FP32 highway. No-op unless the dump directory is set.