batching
The scaffolded arch.py omitted memory_planner, so ported KV-cache
models silently used degraded memory estimation; render it in the
skeleton and document how batching is inherited in SKILL.md.
MODULAR_ORIG_COMMIT_REV_ID: 3d73290461dc771deabe61fc524416446e8e35d3
Removing the bad lines, if people want this functionality for their
scripts they can go back and mark it as excecutable. This won't break
anything.
MODULAR_ORIG_COMMIT_REV_ID: 529c5c5fcf3c33ba314d3c74816fc6bfefe4ed52
Update various instances of install commands to use `max[all]` or other
packages as appropriate (remove some if not necessary).
MODULAR_ORIG_COMMIT_REV_ID: 9e93b8476c329efd90953b1be1637cf3afb57d0a
The website is now just MAX docs, so this path is redundant. Also
reorganize a bit to better reflect the latest framework information
architecture:
- Move `deploy/` into `serve/`
- Move `inference/` into `serve/`
- Move `graph/` into `develop/`
This doesn’t move the source files to remove the `/max` path in the
repo, but instead changes the build configuration to strip the `/max`
prefix when generating the docs build. This makes the change a lot more
simple and primarily requires updating a lot of links.
We can come back later to delete the `/max` path in the source, but it’s
not urgent.
MODULAR_ORIG_COMMIT_REV_ID: 2d37611db386612ef73f98848caa7af8a63f1d0b
Patrick Rachford