For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
MAX includes an interpreter that runs operations one at a time as your graph calls them. For some of these operations, such as matrix multiplication and elementwise math, the interpreter builds an optimized, compiled version the first time it runs the operation. It saves each compiled version to an on-disk cache and reuses it on later runs.
There's one compiled version for each combination of operation, device, and
data type your machine supports, so compiling them all on first use can take
several minutes. Use max warm-interpreter-cache to compile every combination
for the current hardware up front.
Because the compiled results depend on the hardware, run this command on the same kind of machine you plan to run on. A common use is during system provisioning, such as a step in a Dockerfile after you install MAX.
MAX saves the cache next to the engine's own model cache and records your machine's hardware alongside it, so other MAX processes on the same machine reuse the compiled results with no extra setup.
The command refuses to run only when a set variable's value conflicts with
the warm: MAX_EAGER_ALLOW_LAZY_COMPILE=0 for a real warm, or
MAX_EAGER_OP_PRECOMPILE=1 for either mode; --check tolerates the
former since it compiles nothing. Unset the variable for this command, for
example env -u MAX_EAGER_ALLOW_LAZY_COMPILE max warm-interpreter-cache.
Use --check to report whether this machine is already warmed, without
compiling anything:
$ max warm-interpreter-cache --checkThis exits with status 0 if the machine is warmed, or 1 otherwise, so a provisioning script or health check can branch on the result.
Running the command again on an already-warmed machine does nothing. Pass
--force to recompile anyway, such as after a toolchain change:
$ max warm-interpreter-cache --forceCompilation runs concurrently in worker processes, one per operation family
by default, capped at the CPU count. Pass --jobs to bound the number of
workers, or --jobs 1 to compile serially in-process.
max warm-interpreter-cache
Usage
max warm-interpreter-cache [OPTIONS]Options
-
--checkReport whether this machine is already warmed, compiling nothing (exit 0 if so, 1 if not).
-
--forceRe-warm even if this machine is already warmed, e.g. after a toolchain change.
-
--jobs <jobs>Concurrent compile worker processes. Defaults to one per op family, capped at the CPU count; 1 compiles in-process, serially.