RelaxingSnorlax/nemotron35-dspark-30l-bf16
This is a 30-layer BF16 DSpark drafter derived from RedHatAI/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B.dspark at revision 4ff45128272e4e5c00ab82d2886946650229ced7.
Dummy layer construction
This checkpoint was created for layer-count experiments. It was not retrained. For every generated layer i, the tensors are copied byte-for-byte from source layer i % 5; non-layer tensors are copied unchanged. The config declares 30 layers and repeats the source layer-type pattern.
Use it as a structural control for storage, loading, memory, or latency experiments. The repeated layers do not represent a trained 30-layer model, and the source model's quality and acceptance measurements do not apply to this variant. No evaluation has been run on this variant.
Source and reproducibility
- Source drafter: RedHatAI/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B.dspark
- Source revision:
4ff45128272e4e5c00ab82d2886946650229ced7 - Source layer count: 5
- Output layer count: 30
- Layer mapping: output layer
icopies source layeri % 5 - Quantization format:
BF16
derived_from.json records the source and construction rule. drafter_checkpoint_sha256.txt contains the SHA-256 of this variant's model.safetensors. Original 5-layer training and quantization records, when present, are retained in provenance/ and are explicitly labeled as source artifacts.
Example serving command
This mirrors the source model's Nemotron 3.5 deployment settings. Serving compatibility and performance of this dummy variant have not been validated.
CUDA_VISIBLE_DEVICES=0,1,2,3 vllm serve \
nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 \
--tensor-parallel-size 4 \
--gpu-memory-utilization 0.89 \
--enable-expert-parallel \
--trust-remote-code \
--speculative_config.method dspark \
--speculative_config.model RelaxingSnorlax/nemotron35-dspark-30l-bf16 \
--speculative_config.num_speculative_tokens 8 \
--speculative_config.draft_tensor_parallel_size 1 \
--speculative_config.draft_sample_method greedy \
--mamba-backend flashinfer