Skip to content

docs: the --npu-target refusal measured in all three forms, with the ordering caveat serve adds (refs #803) #1664

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
bong-water-water-bong merged 1 commit into from
Nov 9, 2026
Merged

docs: the --npu-target refusal measured in all three forms, with the ordering caveat serve adds (refs #803) #1664

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 17 additions & 0 deletions docs/npu-p4-target-split.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -176,6 +176,23 @@ A3 is the step with the large diff and no numerical content; it goes after the g
`--npu-target hexagon: this build has no such target (compiled: xdna)` and exits, and the same message comes
out of `serve --device npu --npu-target hexagon` because it reaches the forward unchanged.

**Measured (2026-11-09, `ONEBIT_NPU=ON` against the shipping runtime, no ticket and no device — the refusal is
before both):**

| form | what it prints |
|---|---|
| `forward-serve … --npu-target hexagon` | `--npu-target hexagon: this build has no such target (compiled: xdna)` |
| `forward-serve …` (default) | proceeds to its own checks — `xdna` is the default |
| `serve --device npu --npu-target hexagon` | the same by-name message, carried through the child |

**One caveat the `serve` row earned:** it only reaches the child after `serve`'s own kernel resolution succeeds.
With a pack whose full ELFs resolve nothing, `serve` refuses first with *its* message (`… no NPU kernel (full ELF)
directory resolves for it: set ONEBIT_NPU_KERNELS …`) and the target refusal is never printed — measured on
`Qwen3-0.6B` and on `bonsai2-27b-t2s2-serve`, whose `kernels-m16/` does not satisfy that check. That is not a
defect to fix in code: the route cannot run without kernels either, and duplicating the compiled-target list into
`serve` to reorder the two refusals would break the rule that keeps the list in one place. It is a reason to set
the kernels up before testing this row, which is what the table's third line did.

Runs go through `npu-hold` tickets; a new ELF is not involved (no kernel changes), so tickets are the normal
kind. The reference binary is the `c20`-era build with the same model dir, so a difference can only be the
split.
Expand Down
Loading