Qwopus3.6-27B-v2-GPTQ-Pro-MTP-BF16

What changed

• Parent GPTQ-Pro artifact: XReyRobert/Qwopus3.6-27B-v2-GPTQ-Pro-v1
• Source for restored MTP tensors: Jackrong/Qwopus3.6-27B-v2
• Restored mtp.* tensors: 15
• Restored MTP dtype counts: {"bfloat16": 15}
• Local tensor shard size after patch: 18.22 GiB total safetensors

Practical note

This is meant for loader and vLLM speculative-decoding experiments. Previous testing on 1x RTX 3090 showed that restoring MTP made draft acceptance work, but did not improve throughput versus the non-MTP GPTQ-Pro baseline. The likely bottleneck was vLLM/GPTQ-Marlin speculative path overhead rather than MTP tensor precision.

For practical long-context 1x RTX 3090 serving, the non-MTP baseline remains the recommended artifact:

text
Copy code
XReyRobert/Qwopus3.6-27B-v2-GPTQ-Pro-v1

Validation status

Structural checks performed during patch creation:

• model.safetensors.index.json contains restored mtp.* keys.
• All indexed shard files exist locally.
• The added MTP shard is a standalone safetensors file.
• config.json records mtp_num_hidden_layers >= 1.

Runtime serving validation is still required before treating this as a working MTP deployment artifact.

References

• Source model: Jackrong/Qwopus3.6-27B-v2
• Parent GPTQ-Pro artifact: XReyRobert/Qwopus3.6-27B-v2-GPTQ-Pro-v1