julianmb/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-ROCmFP4-GGUF

NVIDIA Nemotron 3.5 Lightning 30B-A3B — ROCmFP4 GGUF

Quantized GGUF conversions of NVIDIA Nemotron 3.5 Lightning 30B-A3B Base (BF16) using the experimental ROCmFPX quantization family from the ROCmFPX llama.cpp fork.

IMPORTANT: These files use custom Q4_0_ROCMFP4_* GGUF tensor types. They are not loadable by mainline llama.cpp. You must build/run the ROCmFPX fork (see how_to_use.md). They are tuned and validated on AMD Strix Halo / RDNA3.5 (gfx1151), with HIP/ROCm and Vulkan kernel support (plus CPU reference paths). Note on Hugging Face Hub metadata: The Hub warning "We're not able to determine the quantization variants" is expected because Q4_0_ROCMFP4_* are experimental, non-standard GGUF types. The table below is authoritative.

Files

All files are quantized from the BF16 source (recommended quality path) and advertise context_length = 262144 (the model's real max_position_embeddings).

Note: Token embeddings for 2688 hidden dimensions fall back from `Q5_K`/`Q6_K` to `Q5_1`/`Q8_0` because 2688 is not divisible by 256.

Model overview

• —Architecture: nemotron_h_moe (hybrid Mamba2 / Attention / MoE)
• —52 layers: 23 Mamba2 + 6 Attention + 23 MoE
• —128 routed experts, 6 active (A3B), 1 shared expert
• —~31.6B total params (3.5B active per token, A3B)
• —Vocabulary: 131072 · Context: 262144 (256K)
• —License: OpenMDW-1.1 (see LICENSE)

Quality & Benchmarks

Perplexity: STRIX_LEAN scores 5.9936 ± 0.0358 on wikitext-2. Measured on Framework AMD Strix Halo (128 GB unified RAM, `gfx1151`, ROCm 7.2.3).

Speculative decoding (embedded MTP)

This model carries an MTP head that works well on its hybrid MoE architecture (unlike some other hybrids). Measured on Strix Halo with the tuned profile (draft-n 6 / p-min 0.60):

Draft acceptance ≈ 88%. Serve via halofpx (halofpx load nemotron-3.5-30b applies this profile automatically).

Quick start

See how_to_use.md for full instructions (build + run).

# 1. Build the ROCmFPX fork
git clone https://github.com/charlie12345/ROCmFPX.git
cd ROCmFPX && env JOBS=16 scripts/build-strix-rocmfp4-mtp.sh   # Strix Halo

# 2. Run on Vulkan (recommended on Strix Halo)
build-strix-rocmfp4/bin/llama-completion -m NVIDIA-Nemotron-3.5-Lightning-30B-A3B-ROCmFP4-STRIX_LEAN.gguf \
  -p "What is 2+2?" -n 64 -dev Vulkan0 -ngl 999 -fa on -c 8192

# 3. Or on HIP/ROCm (unified memory enables APUs/iGPUs)
HSA_OVERRIDE_GFX_VERSION=11.5.1 GGML_HIP_ENABLE_UNIFIED_MEMORY=1 \
  build-strix-rocmfp4/bin/llama-completion -m NVIDIA-Nemotron-3.5-Lightning-30B-A3B-ROCmFP4-STRIX_LEAN.gguf \
  -p "What is 2+2?" -n 64 -dev ROCm0 -ngl 999 -fa on -c 8192

Technical Notes & Findings

• —MTP/NextN head is not included in this conversion (the NemotronH converter skips mtp.* tensors for MoE models).
• —NVFP4 path analysis: We evaluated converting the native NVFP4 checkpoint. With our converter patches, native NVFP4 GGUF loads, but scores PPL 109.79 because runtime kernels do not integrate ModelOpt's companion scale2 factor (~1.4e-4). The clean BF16 → ROCmFP4 path is used for all delivered models.
• —Converter patches, reproduction scripts, and full benchmark notes are available in the companion repo: [julianmb/nemotron-3.5-30b-a3b-rocmfp4](https://github.com/julianmb/nemotron-3.5-30b-a3b-rocmfp4) on GitHub.

Credits

• —Model: NVIDIA (weights, architecture, license: OpenMDW-1.1)
• —ROCmFPX / ROCmFP4 quantization + kernels: charlie12345/ROCmFPX
• —Conversion & quantization performed by: julianmb