The central R9700 agent serving campaign studied Qwen3.8 27B. Its primary checkpoint was ukisai/Swift-1.5-Qwen3.8-27b-INT4, an AWQ+GPTQ INT4 hybrid GDN variant. This is the model behind the single-card RAM16 milestone and the follow-on TP=2 investigation summarised in the research journal.
The follow-up single/dual comparison used the Swift-1.5 checkpoint at revision 81a28ee75eb1cd469f209813a368a79a7e3db3bc, FP8 attention KV and model-matched calibration. MTP depth was four on GPU A and eight on dual. DFlash2 and MXFP4 experiments below are separate research branches, not the configuration behind these matrix results. See the configuration record.
The broader collected Radiance work also contains tests and build/benchmark notes for the following Hugging Face checkpoints. Their presence here records model use in the research archive; it does not mean their results are included in the retained case charts or directly comparable to the primary Swift-1.5 campaign. Check the cited notes and each benchmark’s configuration before interpreting a result.
| Hugging Face repository | Role in the collected work | How to interpret it |
|---|---|---|
ukisai/Swift-1.5-Qwen3.8-27b-INT4 |
Primary research target: hybrid Qwen3.8 27B, AWQ+GPTQ INT4; used for the one-card agent campaign and TP=2 follow-up. | Central model for the historical agent-serving findings in this package. |
amd/Qwen3.8-27B-Quark-AWQ-MXFP4 |
Qwen3.8 27B Quark AWQ/MXFP4 target used in separate Radiance W4A8 and DFlash2 work. | Related model-format/kernel experiments, not a substitute for the Swift-1.5 INT4 run results. |
tcclaviger/Qwen3.8-27B-DFlash2-FP8 |
Target-matched Qwen3.8 DFlash2 draft checkpoint used in separate speculative decoding experiments. | Drafter, not the primary target model. Speculative results depend on the paired target and runtime path. |
Qwen3.8-27B-heretic-ara-fp8-magiccodingman |
Native-FP8 Qwen3.8 target recorded in Radiance build and benchmark notes. | Separate target variant. The source archive refers to it by local model name/path; no HF namespace is asserted here. |
Qwen3.8-27B-heretic-ara-DFlash2-fp8-magiccodingman |
Selective-FP8 DFlash2 draft recorded alongside the heretic-ara target. | Drafter, not a standalone target result. Source notes use local paths rather than a verified HF repository ID. |
Qwen3.8-27B-heretic-ara-heretic-org |
Additional Qwen3.8 target variant appearing in DFlash2 loading/serving notes. | Related exploratory checkpoint; not the primary Swift-1.5 campaign. Source notes do not establish a canonical HF namespace. |
Qwen3.8-27B-DFlash2-z-lab |
DFlash2 draft variant appearing in model load and serve notes. | Drafter; source notes do not establish a canonical HF namespace. |
The model snapshot, tokenizer files, draft weights, and FP8 KV calibration sidecars are not redistributed with this package. The primary launch instructions use operator-supplied model paths and IDs. See launch instructions, benchmark interpretation, and the run index.
Home · Overview · ROCm 10.1 results · Charts