hermitdave's picture
Upload README.md with huggingface_hub
a197bd3 verified
|
Raw History Blame Contribute Delete
3.39 kB
metadata
license: apache-2.0
tags:
  - imatrix
  - quantization
  - mlx
  - moe
  - inkling2
  - looped-transformor
pretty_name: Inkling 2 Mini imatrix

Inkling-2-Mini imatrix

Activation-importance matrix (in_sum2 / counts) for thinkingmachines/Inkling-2-Mini, captured for rco-mlx mixed-precision quantization.

File

file size contents
imatrix_inkling2.npz 2.8 MB 369 tensors × (.in_sum2, .counts)

Capture

Source checkpoint thinkingmachines/Inkling-2-Mini (native NVFP4 + MXFP8)
Calibration WikiText-2, 24 rows × 2048 tokens = 49,152 tokens
Corpus calib_wikitext_128x2048.npy
Routed expert slots 37,748,736
Capture driver rco-mlx scripts/capture_imatrix_inkling2.py

Captured on the native checkpoint as the primary source: the published experts are NVFP4/MXFP8, and the native forward pass is the function the quant has to reproduce. (The BF16 reconstruction is 52.8 GB and thrashes a 64 GB machine during capture.)

A second capture was run on the dequantised BF16 source to cross-check. It completed and produced a byte-identical file (MD5 176f39b7…), which independently confirms that the numpy NVFP4/MXFP8 dequantisation reproduces the published checkpoint's forward pass exactly.

Coverage

class tensors
Routed experts (stacked gate_up_proj / down_proj) 48
Attention projections 168
Other linears (vision tower, mergers, message gates) 153

Per-tensor routed-slot counts follow the loop structure exactly — core layers 6–17 (run 3× per token) accumulate 3× the slots of the non-core layers, and every expert tensor carries tokens × top_k × visits slots. Mean is ≈ 3,072 slots per expert per layer; no expert is unrouted at this corpus size.

Key format

Keys are MLX module paths without .weight, matching the tensor names in the RCO artifacts:

model.language_model.layers.0.mlp.experts.gate_up_proj
model.language_model.layers.0.mlp.experts.down_proj
model.language_model.layers.0.self_attn.q_proj

Routed experts are stored aggregated per stacked tensor (one row per (layer, projection)) rather than per expert index, because the RCO artifacts keep each layer's 128 experts in a single pre-stacked tensor — one group, one imatrix row. counts is the total routed-slot count for that tensor.

Verified against both published artifacts: 224 / 224 quantized tensors mapped (100 %) for hermitdave/Inkling-2-Mini-MLX-RCO-4.5bpw and hermitdave/Inkling-2-Mini-MLX-RCO-3.5bpw.

Use

python examples/moe_pipeline.py \
  --source <bf16-source> --output <out> \
  --model-type inkling2_loop --target-bpw 4.25 --group-size 64 \
  --bits 3,4,6,8 --expert-allocation per-projection \
  --imatrix imatrix_inkling2.npz --per-group-clip

--per-group-clip is where the imatrix pays: it picks a clip ratio per (row, group) by the imatrix-weighted MSE. Weighting the allocation alone was measured not to help at tight budgets; an unweighted per-group search is worse than the per-tensor clip it replaces.

Attribution

Captured by Hermes Agent with hermitdave/rco-mlx. Model by Thinking Machines Lab, Apache 2.0.