[feat] Add support for Qwen3.5 and Qwen3-Next to ATOM-plugined SGLang

[wanzhenchn]

Motivation

Background: ROCm/ATOM#355 and ROCm/ATOM#359.

PR #355 integrated ATOM with upstream SGLang through the SGLANG_EXTERNAL_MODEL_PACKAGE out-of-tree mechanism, replacing a fork-based workflow and establishing atom.plugin.sglang.models as the external entry package for ATOM-backed architectures.

Building on that foundation, this PR extends the SGLang plugin path so that two major ATOM model families—Qwen3-next (Qwen3NextForCausalLM) and Qwen3.5 (Qwen3_5ForConditionalGeneration/Qwen3_5MoeForConditionalGeneration)—can run as first-class external models inside SGLang. The goal is parity with prior ATOM-in-SGLang accuracy while improving end-to-end inference performance on the supported paths (e.g. ATOM’s fused kernels, quantization, and MLA / MoE handling tuned for ROCm), without requiring a patched SGLang tree—users continue to point SGLANG_EXTERNAL_MODEL_PACKAGE at atom.plugin.sglang.models and launch with standard upstream sglang.launch_server.

Technical Details

• Qwen3-next

  • Qwen3NextForCausalLM is registered under atom.plugin.sglang.models and subclasses _AtomCausalLMBaseForSglang, reusing the same SGLang-facing contract as other OOT entry points: the wrapper calls prepare_model(..., engine="sglang") to build the ATOM weight stack, runs the language model forward with pipeline-parallel state mapped from pp_proxy_tensors, applies LogitsProcessor on the last PP rank, and loads weights via load_model_in_plugin_mode.
  • The linear-attention (GDN) path is wrapped by Qwen3NextSglangModel plus sglang_gdn_bridge so GDN layers see the SGLang forward_batch context they expect. At prepare time, apply_qwen3_next_sglang_model_patch swaps atom.models.qwen3_next.Qwen3NextModel to that bridged implementation; the shared prepare hook defaults ATOM_SGLANG_USE_NATIVE_AITER_ATTN_BACKEND for Qwen3NextForCausalLM before register_ops_to_sglang.

• Qwen3.5

  • Qwen3.5 text / MoE / multimodal stacks reuse SGLang’s in-tree Qwen3_5* container classes while the language tower is still constructed through ATOM prepare_model. apply_prepare_model_adaptations applies Qwen3.5-only config fixes (e.g. MoE text fields, quant remaps for ROCm) and _apply_qwen35_sglang_model_patch rebinds atom.models.qwen3_5.Qwen3_5Model to the SGLang-specific implementation.
  • Weights are not applied through SGLang’s default load_weights iterator; instead load_model_in_plugin_mode drives ATOM’s loader, with _QWEN35_OOT_PACKED_MODULES_MAPPING (and related mapping) kept aligned to SGLang’s packed/fused parameter layout so FP8 and fused checkpoints behave like the in-tree model.
  • The same prepare hook defaults native Aiter attention backend selection for Qwen3_5* (and Qwen3-next) unless the user has already set the env var.

How to Run

export SGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models
python3 -m sglang.launch_server --model-path <model_path> ...

The following models have been supported:

• Qwen3.5-397B-A17B
• Qwen3.5-397B-A17B-FP8
• Qwen3.5-35B-A3B
• Qwen3.5-35B-A3B-FP8
• Qwen3.5-27B
• Qwen3.5-27B-FP8
• Qwen3-Next-80B-A3B-Instruct

Accuracy

# launch server
#! /usr/bin/bash

set -euxo pipefail

export PYTHONPATH=/opt/sglang/python
export SGLANG_DISABLE_CUDNN_CHECK=1

if [ $# -lt 4 ]; then
  echo "Usage: $0 <port> <model_path> <device_id> <enable_atom>"
  exit 1
fi

port=$1
model_path=$2
device_id=$3
enable_atom=$4 # true or false

if [ "${enable_atom}" == true ]; then
  export SGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models
fi

tp=$(echo "$device_id" |grep -o "[0-9]" |grep -c "")

HIP_VISIBLE_DEVICES=${device_id} python3 -m sglang.launch_server \
  --model-path ${model_path} --port ${port} \
  --tensor-parallel-size ${tp} --mem-fraction-static 0.9 \
  --reasoning-parser qwen3 --disable-radix-cache

model_path=$1
port=$2
task=$3

models_url="http://localhost:${port}/v1/models"
echo "Waiting for OpenAI-compatible server at ${models_url} ..." >&2
until curl -sf --connect-timeout 2 --max-time 10 "${models_url}" >/dev/null; do
  sleep 2
done
echo "Server is up; starting lm_eval." >&2

lm_eval --model local-completions \
        --model_args model=${model_path},base_url=http://localhost:${port}/v1/completions,num_concurrent=64,max_retries=3,max_gen_toks=2048,tokenized_requests=False \
        --tasks ${task} \
        --batch_size auto \
        --trust_remote_code

• Qwen3.5-397B-A17B-FP8

  • ATOM + SGLang
  
  • SGLang only
  

• Qwen3-Next-80B-A3B-Instruct

  • ATOM + SGLang
  
  • SGLang only
  

Inference Perf.

• Qwen3.5-397B-A17B-FP8 on MI355X

• Qwen3.5-397B-A17B-FP8 on MI308X

[wuhuikx]

we should remove the **model_kwargs and refer to DS how to implement the attn part.

[wuhuikx]

Let's remove the **kwargs here and use wrapper in plugin/sglang to handle.