G2 reference embedding lookup. This intentionally materializes the checkpoint-native quantized table through the same exact dequantizer as linear weights; G3 replaces the full-table materialization with a row-gather path.
Gather only the requested embedding rows before dequantization.
The production GLM table is signed int8 and expands to ~3.8 GiB as f32. The streamed path must never create that full-table transient merely to read a handful of token rows.
Bounded signed-int8 linear used by the streamed production lm_head.
Each output-row tile is fully evaluated before the next tile is built, so MLX cannot retain a lazy graph containing the complete f32-expanded head. Concatenation retains only the small materialized logits tiles.
MLX affine-Q4 linear for the streamed resident spine.
Colibri stores already-packed nibbles but exposes them as uint8. Packing
those bytes into MLX's uint32 view is still far smaller than expanding the
complete matrix to f32, and preserves the exact (nibble - 8) * scale
affine values.
StaticopenStaticopenOpen only the shards that own tensorNames.
This is an explicit probe/tooling seam, not a partial-model constructor.
Quantized callers must include both the packed tensor and its .qs
companion. Accessing any catalogued tensor in an unopened shard fails.
Materialize the logical [output,input] tensor as f32.
This is intentionally the slow correctness path. Colibri int4 is offset-binary (
nibble - 8); int8 is signed two's-complement. Neither is inferred from config.quantization_config, which still describes the pre-conversion FP8 checkpoint.