Block a user
qwen4_exp takes one
past_len for the whole batch, so ragged sequences cannot be batched
qwen4_exp C: host- or disk-resident PLE table with a device-side gather
qwen4_exp C: host- or disk-resident PLE table with a device-side gather
Closing — the table is resident nowhere and the gather is no longer a cost
Both halves of this issue are delivered and measured on beast.
Residency (slice 1, 1ae3ba72). The 102.4 GB…
qwen4_exp gate 6: full-model logits parity against the upstream reference — the port is unverified against anything but our own reading
qwen4_exp gate 6: full-model logits parity against the upstream reference — the port is unverified against anything but our own reading
Closing — gate 6 passed, and the harness runs on every push
crates/neuron/tests/numerical_reference.rs plus script/generate_qwen4_exp_fixture.py, with two committed fixtures of random…
In-situ quantisation is redone on every load and discarded on unload — ~26 minutes per load for qwen4_exp
In-situ quantisation is redone on every load and discarded on unload — ~26 minutes per load for qwen4_exp
Closing — landed and measured
The per-layer ISQ cache is on main and has been exercised repeatedly since, including through every qwen4_exp window this week.
**Measured, same OOM point…
qwen4_exp C: host- or disk-resident PLE table with a device-side gather
Solved by
Solved by MADV_RANDOM — 20 lines, 176× on the gather. Quantisation not needed.
eb403ef, measured on beast with three distinct prompts so every run gathers cold rows.
qwen4_exp C: host- or disk-resident PLE table with a device-side gather
Measured: the mmap gather is the dominant decode cost, and it contends with #318
Slice 1 (mmap the table instead of making it resident) is what made the model loadable at all. It also turns…