---
license: apache-2.0
base_model:
- amd/Qwen3.8-27B-Quark-AWQ-MXFP4
- z-lab/Qwen3.8-27B-DFlash2
pipeline_tag: image-text-to-text
tags:
- radiance
- rocm
- amd
- rdna4
- mxfp4
- speculative-decoding
- dflash2
- vision
---

# Qwen3.8-27B MXFP4 + DFlash2 — radiance container

[amd/Qwen3.8-27B-Quark-AWQ-MXFP4](https://huggingface.co/amd/Qwen3.8-27B-Quark-AWQ-MXFP4) —
Qwen3.8-27B with AWQ-calibrated OCP-MXFP4 weights from AMD Quark — as a single `.rad` container for
the **radiance** inference engine (AMD RDNA4, ROCm), with its vision tower and the
[z-lab/Qwen3.8-27B-DFlash2](https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2) block-diffusion
drafter merged in for speculative decoding.

**Needs radiance 1.2.3 or newer.** The drafter's MXFP4 linears and the rescored 2-bit head arrive in 1.2.3; older releases refuse this file. The previous version of this file (fp8 drafter) is in this repository's history.

| | |
|---|---|
| File | `qwen3.8-27b-mxfp4.rad` — 18.05 GiB |
| Weights | AMD's MXFP4 trunk exactly as released — e2m1 codes, an E8M0 exponent per 32 — served at 4-bit weights against E4M3 activations; the lm_head made block FP8 by the recipe |
| Speculator | DFlash2, its linears OCP-MXFP4 from its bf16 release; its vocabulary head as 2-bit codes, whose top 32 candidates the engine rescores against the model's own FP8 head. Drafts are verified, so they change the speed and never the output |
| Vision | the 27-block vision tower, bf16: images and video in chat requests |
| Context | 262,144 tokens trained; 200K tested |

## Speed

Radeon AI PRO R9700 (gfx1201), one card (`--tp 1`) and two (`--tp 2`):

| | this file, TP1 | this file, TP2 | previous file, TP2 |
|---|---|---|---|
| one stream, all eight prompts (tok/s) | 122 | 204 | 182 |
| one stream, prose / code (tok/s) | 77 / 164 | 130 / 285 | 112 / 256 |
| four streams, total (tok/s) | 280 | 438 | 387 |

Greedy, 768 tokens each of eight prompts (prose, code, an explanation, JSON, Rust, German, two with reasoning), depth 7, `--max-num-seqs 8 --max-model-len 200000 --kv-cache-dtype fp8`; "previous file" is this repository's earlier version (fp8 drafter, unrescored 2-bit head) served by 1.2.2 on the same machine.

AMD reports the quality of these weights on its model card (GSM8K 5-shot, 99–102% of bf16); this
file serves the same weights.

## Serve

```sh
radiance --model qwen3.8-27b-mxfp4.rad --tp 2 --max-model-len 200000 --kv-cache-dtype fp8 \
    --max-num-seqs 8 --host 0.0.0.0 --port 8000
```

It also fits one 32 GB card: `--tp 1`, the same flags otherwise.

The server speaks the OpenAI API (`/v1/chat/completions`, `/v1/completions`), with tool calls and
structured output, and `image_url` / `video_url` parts in chat messages. The drafter's depth is
chosen automatically (`--num-speculative-tokens N` states one, `0` turns speculation off).

## How this file was made

rad-convert reads Quark's MXFP4 linears as they ship (e2m1 codes and E8M0 exponents, byte for byte),
so the recipe names only what the release left in bf16:

```sh
rad-convert amd/Qwen3.8-27B-Quark-AWQ-MXFP4 --draft-model z-lab/Qwen3.8-27B-DFlash2 \
    --recipe q38-27b-quark-mxfp4-df2.recipe -o qwen3.8-27b-mxfp4.rad
```

The recipe (`q38-27b-quark-mxfp4-df2.recipe` in this repository); everything it does not name is
the checkpoint's own:

```
blk.*.ssm_ab.weight cast dtype=bf16
output.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16
dflash.draft_head.weight rtn codes=u2 zero=u8 group=128 scale=f16 rule=search
dflash.fc.weight rtn codes=fp4_e2m1 group=32 scale=e8m0
dflash.blk.*.attn_q.weight rtn codes=fp4_e2m1 group=32 scale=e8m0
dflash.blk.*.attn_k.weight rtn codes=fp4_e2m1 group=32 scale=e8m0
dflash.blk.*.attn_v.weight rtn codes=fp4_e2m1 group=32 scale=e8m0
dflash.blk.*.attn_output.weight rtn codes=fp4_e2m1 group=32 scale=e8m0
dflash.blk.*.ffn_gate_up.weight rtn codes=fp4_e2m1 group=32 scale=e8m0
dflash.blk.*.ffn_down.weight rtn codes=fp4_e2m1 group=32 scale=e8m0
```

The delta net's a/b projection is declared bf16 and keeps AMD's values that way.