DeepDavid/Qwen3.6-35B-A3B-qcache-q4k

Qwen3.6-35B-A3B — q4_k qcache (candle ignite-perf format)

Quantized serving set for Qwen/Qwen3.6-35B-A3B (hybrid GDN/attention MoE: 40 layers, 256 experts, top-8, moe_intermediate 512), produced and consumed by the ignite-perf candle fork (branch ignite-perf).

Files

file	contents
qcache-q4_k.v1.gguf	all quantized linear projections (q4_k), classic per-expert key names (…mlp.experts.{i}.gate_proj etc.) regardless of the packed source layout
aux-tensors.safetensors	every non-quantized text-model tensor: embeddings, norms, GDN conv/dt/A_log params, router gates — original checkpoint names
config.json, tokenizer*	upstream configs (VL wrapper with text_config)

qcache + aux is a complete runnable set — the original bf16 shards are not required.

Loading (ignite-catalog exp1 executor)

{
  "asset_prefix": "<this repo's local dir>",
  "model_family": "qwen3_5",
  "quantize": "q4_k",
  "model_files": ["aux-tensors.safetensors"]
}

The loader reads quantized weights from the qcache and everything else from the aux file. Packed-expert checkpoints (experts.gate_up_proj [E, 2I, H], down_proj [E, H, I]) are handled natively.

Notes

• Quantization is deterministic (byte-identical across regenerations).
• Produced on RTX PRO 4500 Blackwell, CUDA 13.3 SASS (CANDLE_CUBIN=1).
• Derived from the Apache-2.0 licensed upstream checkpoint.