Mixed-precision GGUF of tencent/Hy3 (295B total / 21B active MoE, Apache 2.0) built for SSD-streaming inference engines: routed experts live on disk and stream per token, so the model runs on GPUs that cannot hold it.

Runs on pulsar (Rust + CUDA, recommended) and the NeutronStar hy3 branch (C, a CUDA port of antirez/ds4).

Measured decode, greedy, warm cache:

Per token only 8 of 192 experts per layer are read (~3GB/token at this quant); attention, shared experts, and the router stay resident.

Both use the identical recipe and the same importance matrix; they differ only in what the quantizer saw as input.

fromBF16 (new): tencent/Hy3 BF16 (598GB) converted to a q8_0 intermediate with the AngelSlim llama.cpp patches (proper hy_v3 architecture support), then quantized with the imatrix in one step. q8_0 is effectively lossless as an intermediate, so the routed experts see exactly one lossy quantization.

original: built before the BF16 pipeline existed here, from the IQ4-UD edition of YanissAmz/Hy3-295B-A21B-GGUF. The tensors this recipe keeps at Q8_0 passed through essentially lossless, but the routed experts went through two lossy steps (IQ4_XS/IQ3_S to IQ2_XXS). At a 2-bit target the 2-bit noise dominates, so the gap is small, but the fromBF16 file removes it entirely.

The layout targets a streaming expert cache: routed experts must be uniform fixed-size slabs, and everything that makes decisions stays high precision. Same design as antirez's GLM-5.2 ds4 build, including the MTP layer riding at Q2_K because importance matrices never cover the draft layer (imatrix generation runs normal forwards, which skip it).

imatrix: the 125-chunk general-purpose matrix published with the source repo of the original build, reused for the fromBF16 build.

Zero-config multi-GPU: pulsar measures each card's PCIe bandwidth at startup, streams experts over the fastest link, and fills spare GPUs with resident hot experts. First run is cold; a .warm sidecar makes every later run start hot. Main knob: PULSAR_CACHE_GB (host expert cache, defaults to measured free RAM minus a reserve).

Useful knobs: DS4_CUDA_HOST_EXPERT_CACHE_GB=16 (host expert cache, the main speed lever) and DS4_CUDA_PARALLEL_FETCH_THREADS=16.

blk.80 (the MTP draft layer) is present in both files. pulsar wires it behind PULSAR_MTP=1 (opt-in): drafts verify correctly at 44 percent acceptance, but measurements show speculative decoding only pays when the verify pass is fully cache-resident, which a 30GB-RAM box cannot deliver. Tencent's own AngelSlim deployment guidance reaches the same conclusion, enabling MTP only from 2x H20 (192GB) upward. Keep it off for speed; it is there for bigger boxes.