Red Lite dev9 — native runtime foundation

Dev9 freezes the Python/ctypes path as a correctness oracle and starts the migration to a DS4-style native runtime.

The new redlite-native executable is written in C and has no Python dependency. In this milestone it owns two pieces that previously lived in Python:

GGUF v2/v3 tensor-directory parsing and routed-expert mapping;
hard-bounded LRU metadata scheduling with top-k and in-flight victim protection.

The field-validated dev8 Metal kernels remain unchanged and continue to be the numerical oracle while the control plane moves native.

Build#

make native

The binary is written to:

.deps/redmetal/redlite-native

make native also runs the standalone native scheduler selftest.

Inspect the real Qwen3-Next model#

.deps/redmetal/redlite-native inspect \
  models/Qwen_Qwen3-Next-80B-A3B-Instruct-IQ2_XXS.gguf \
  --cache-mib 256 \
  --layers

For the field model, the native parser is expected to reproduce the Python oracle observations:

GGUF routed layers: 48
routed tensors: 144
expert count: 512
quant tensors: 33 IQ2_XS and 111 IQ1_M
layer patterns: 11 IQ2_XS layers and 37 IQ1_M layers
all merged routed tensors slice-safe
top-10 fits in a 256 MiB cache

The exact routed payload and slot size must also agree with the existing Python planner within normal binary-unit formatting.

Native selftest#

.deps/redmetal/redlite-native selftest

It validates:

hard capacity enforcement;
LRU hit/miss/eviction behavior;
current top-k selection protection during acquire_many;
in-flight slot protection;
zero Python dependency.

Architecture after dev9#

Development / oracle                 Runtime direction
--------------------                 -----------------
Python GGUF parser       ───────►     native/redlite_native_gguf.c
Python LRU policy        ───────►     native/redlite_native_cache.c
Python CLI planning      ───────►     redlite-native
Python parity reference  retained     tests/oracle only
ctypes Metal probes      retained     correctness oracle
Metal kernels            retained     production kernels

Dev9 does not yet execute a token entirely through the native binary. The next native milestones are to embed canonical IQ2_XS/IQ1_M codebooks, bind the already validated Metal top-k executor directly to the native cache, then port the Qwen3-Next router and full layer/token scheduling. Python should eventually be unnecessary for normal inference.