Expand description
Resolving the compiled target’s serving defaults — baked first, environment second.
§The defect this closes
Maintainer review of the H100 integration branch, 2026-09-11 (tbraun96):
There is no arch separation at all. H100 builds compile GB10’s kernel tree. Every Hopper/GB10 divergence is expressed as an env lever set by an H100 recipe living outside this repo — not as arch-selected code. “No interference” rests on discipline rather than structure.
Every lever here used to read the environment and fall back to a literal that described GB10. The H100 configuration therefore lived in a launch script nobody in this repository could see, review or test, and “GB10 is unaffected” was a promise about which prefixes people remembered to type.
Now the fallback is atlas_kernels::TARGET_DEFAULTS, baked by
build.rs from the ONE kernels/<hw>/HARDWARE.toml this binary compiled
([defaults]). A GB10 build cannot carry Hopper’s numbers, an H100 serve
needs no prefixes, and every value is reviewable beside the arch it
belongs to.
§The override grammar
| input | meaning |
|---|---|
| variable absent | the baked target default |
0, false, off, no (any case, trimmed) | OFF — explicit override |
| any other value, including empty | ON — explicit override |
⚠️ VAR=0 NOW MEANS OFF. These levers were PRESENCE-gated
(var_os(..).is_some()), chosen so an A/B recipe could stay a bare VAR=1
prefix with no “=0 means on” trap. Presence cannot express “off”, and
once a target’s default can be ON, an operator with no way to turn a lever
off is back to editing launch scripts. The trap the old rule avoided is
gone in the direction that matters: VAR=0 now means what it reads as.
VAR=1 is unchanged everywhere.
The legacy ATLAS_NO_* kill switches stay PRESENCE-gated and still force
their lever OFF, so no script that predates this file changes meaning.
ATLAS_NO_DECODE_SPLIT_SILU is the one in this table.
§One resolution, one log line
resolved is the SSOT: every consumer below reads it, and
spark-server prints it as target defaults (<hw>): … with the
environment-sourced values marked. A lever resolved in two places is a
lever that can disagree with the line that claims to report it.
§Adding a lever — the contract
ONE commit touches all of: the field in
atlas_kernels::TargetDefaults, the parse arm in
atlas-kernels/build_defaults.rs, the row in EVERY
kernels/<hw>/HARDWARE.toml that has a [defaults] table, the
TargetLevers field and its arm in resolve, the field in
format_levers’s line, and a test. parse_defaults panics on an
unknown key, so a half-landed lever fails the build rather than reading as
agreement with the baseline.
★ And the commit that does all that is the one landing the lever’s
CONSUMER. A row whose dispatch site does not exist yet cannot be graded,
answers nothing when an operator sets its variable, and puts an (env)
tag in the boot line against a decision that changes no code. So a kernel
PR brings its own row; this module ships only the rows whose arms are
already here.
Structs§
- Resolved
- A resolved lever and where it came from.
- Target
Levers - Every serving lever this target declares, resolved against the environment.
Enums§
- Source
- Where a resolved value came from — the whole point of the log line.
Constants§
- BASELINE_
BATCHM_ MAX - The declaration a target that says nothing gets — kept in sync with
atlas-kernels/build_defaults.rs::baselinebytarget_defaults_tests::the_baseline_band_is_the_frozen_one.
Functions§
- declared
- The baked declaration this binary carries, for the serve log’s header and
for callers that must stay pure over their own inputs (
ModelLevers). - format_
levers summary_lineover a table the caller already has — pure, so the line can be graded for ANY target from a CPU test without touching the process environment or sealing theOnceLock.- resolve
- The whole table, as a pure function of the baked declaration and a variable lookup — so the resolution is testable for ANY target from a CPU test, on any host, without touching the process environment.
- resolve_
batchm_ max - The BF16 decode head’s batched-GEMV band.
- resolve_
max_ m - Upper
Mfor the W8A8 dense-FFN prefill, per projection shape. - resolve_
toggle - The override grammar for a boolean lever, as a pure function.
- resolved
- The process-wide resolution.
- summary_
line target defaults (<hw>): …— one line naming every resolved value and which came from the environment.