GLM-5.3-Flash: NVFP4 / FP8 hybrid

This checkpoint combines the main-stack routed experts from dabsLabs/GLM-5.3-Flash-NVFP4 with the remaining tensors from unsloth/GLM-5.3-Flash-FP8. It was assembled for DGPP, a C++/CUDA inference engine for DGX Spark. The full checkpoint is included here, along with its tokenizer, configuration files, chat template, and composition manifest.

This is a repackaging of existing quantized weights, not a new training run or quantization pass. Each output tensor was copied from one of the two releases without arithmetic. Tensors taken from the FP8 release retain their original dtypes; they are not all FP8.

Sources and attribution

Z.ai created GLM-5.3-Flash. Credit for the NVFP4 expert quantization belongs to dabsLabs, and credit for the FP8 release used here belongs to Unsloth and the upstream model authors. HawkBearPig composed and published this hybrid checkpoint.

The second row includes attention, shared experts, dense MLPs, routers, mHC, normalization, embeddings, the output head, vision tensors, and the complete MTP draft layer (layer 45). In particular, neither DSA attention nor the MTP layer comes from the dabsLabs checkpoint.

Weight format

The configuration identifies this layout as quant_method: "dgpp_mixed". Main-stack routed experts use three tensors per projection:

• weight_packed: two E2M1 values per byte, with the even element in the low nibble.
• weight_scale: one E4M3 scale per group of 16 weights along the input dimension.
• weight_global_scale: one FP32 scale per projection tensor.

Their dequantization rule is E2M1(code) * (float32(weight_scale) / weight_global_scale). DGPP uses BF16 activations for these weight-only NVFP4 projections (NVFP4A16). FP8 weights elsewhere retain the source release's weight and weight_scale_inv tensors, with 128 × 128 block scales.

Running the checkpoint

The recorded serving configuration is world size 4: four DGX Spark (GB10) nodes, one rank per node, using tensor parallelism over RoCE. Other world sizes have not been validated for this checkpoint.

Download it into the Hugging Face cache on each node:

hf download HawkBearPig/GLM-5.3-Flash-NVFP4-FP8

In DGPP, use deploy/cluster.nvfp4.example.json as the deployment template and set its model field to HawkBearPig/GLM-5.3-Flash-NVFP4-FP8. The original local composition used the cache name dgpp/GLM-5.3-Flash-NVFP4-FP8; that name remains in the provenance manifest.

This checkpoint requires a loader that understands dgpp_mixed. Compatibility with a source checkpoint does not establish compatibility with this hybrid; it is not a drop-in Transformers or vLLM checkpoint. DGPP currently serves text only. Vision tensors are included, but their presence does not imply that DGPP supports image input.

Composition and verification

The checkpoint contains 112,396 tensors in 40 Safetensors shards, with 195,075,058,296 bytes of tensor data (about 195.1 GB, excluding headers). DGPP's tools/compose_nvfp4_hybrid.py produced it using the revisions above, with --mtp-from base, --dsa-from base, and a 5,000,000,000-byte shard target.

At composition time, every output tensor was reread and compared byte for byte with its source. MANIFEST.json records the source revisions, tensor census, composition options, and each shard's SHA-256 digest. Its revision value, 833e7cc0e86d9eb125551660dbf48cab2e224df9, identifies the composition plan; it is not a Git revision of this Hugging Face repository. The local build hostname has been omitted from the published manifest.

These checks establish that composition preserved the source tensors. They do not establish equal accuracy to the original model or either source checkpoint. This release does not report a separate perplexity or downstream benchmark evaluation of the hybrid.

License

The checkpoint is distributed under the MIT license, consistent with both source releases. The included LICENSE preserves Z.ai's copyright and permission notice. Please retain the source attribution when redistributing this composition.