feat: add RWKV-7-World model support on cuda device.

[Dragonliu2018]

Description

Adds RWKV-7-World CUDA inference support to xLLM. Official checkpoints ship as native .pth files; this PR provides a conversion script to produce HuggingFace-style model directories and wires the model into xLLM’s linear-attention KV cache and serving stack.

Changes

• Model: New rwkv7 causal LM (embedding + RWKV-7 blocks + lm_head), registered as model_type=rwkv7
• Decoder layer: RWKV-7 time-mix / channel-mix implementation, using linear-attention KV cache for W-matrix and token-shift state
• Tokenizer: Native RWKV trie tokenizer (rwkv_vocab_v20230424.txt), registered as tokenizer_type=rwkv
• Conversion tool: tools/convert_rwkv7_world.py — converts ModelScope/BlinkDL .pth to config.json + model.safetensors + tokenizer files
• Tests: rwkv_tokenizer_test, rwkv7_decoder_layer_test, hf_model_loader_test (rwkv7 args)

The test program and the generated text are as follows.

• OS: Ubuntu22.04
• device: NVIDIA A800-SXM4-40GB
• model: Blink_DL/rwkv-7-world

Convert model:

python tools/convert_rwkv7_world.py \
  --src /root/models2/rwkv-7-world \
  --dst /root/models2 \
  --sizes all

Serve:

./build/xllm/core/server/xllm --model=/mnt/data2/models2/rwkv-7-world-2.9b-xllm --port=9977 --device_id=0

Test program:

#!/usr/bin/env python3
"""Smoke test for RWKV-7 World xLLM service (converted HF-style directory).
"""

import os
import sys

import requests

BASE_URL = os.environ.get("XLLM_BASE_URL", "http://127.0.0.1:9977")
PROMPT_TEXT = os.environ.get("PROMPT_TEXT", "The capital of France is")
MAX_TOKENS = int(os.environ.get("MAX_TOKENS", "10"))
TEMPERATURE = float(os.environ.get("TEMPERATURE", "0"))
TOP_K = int(os.environ.get("TOP_K", "1"))


def main() -> int:
    model_id = requests.get(f"{BASE_URL}/v1/models", timeout=10).json()["data"][0]["id"]
    print(f"model: {model_id}")
    print(f"prompt: {PROMPT_TEXT!r}")
    print(f"generation: max_tokens={MAX_TOKENS} temperature={TEMPERATURE} top_k={TOP_K}")

    response = requests.post(
        f"{BASE_URL}/v1/completions",
        json={
            "model": model_id,
            "prompt": PROMPT_TEXT,
            "max_tokens": MAX_TOKENS,
            "temperature": TEMPERATURE,
            "top_k": TOP_K,
            "stream": False,
        },
        timeout=180,
    )
    if response.status_code != 200:
        print(f"HTTP {response.status_code}: {response.text}", file=sys.stderr)
        return 1

    payload = response.json()
    choice = payload["choices"][0]
    text = choice.get("text", "")
    print(f"completion text: {text!r}")

    usage = payload.get("usage")
    if usage:
        print(
            "usage: "
            f"prompt_tokens={usage.get('prompt_tokens')} "
            f"generated_tokens={usage.get('completion_tokens', usage.get('generated_tokens'))}"
        )

    if PROMPT_TEXT == "The capital of France is" and not text.lstrip().startswith("Paris"):
        print(
            "warning: expected completion to start with 'Paris' for the default prompt",
            file=sys.stderr,
        )
        return 1

    return 0


if __name__ == "__main__":
    raise SystemExit(main())

Output:

(xllm) ➜  xllm git:(lzl/feat/rwkv7-world-model-support) ✗ python examples/test_rwkv7_world.py
model: rwkv-7-world-2.9b-xllm
prompt: 'The capital of France is'
generation: max_tokens=10 temperature=0.0 top_k=1
completion text: ' Paris.\nThe capital of France is Paris.'
usage: prompt_tokens=5 generated_tokens=10

Related Issues

Change Type

• Bug fix
• New feature
• Performance improvement
• Refactor
• Documentation
• Test
• Build or CI

Pull Request Checklist

Thank you for contributing to xLLM. Before requesting review, please make sure the following items are complete.

PR Title and Commit Messages

• The PR title and each commit message follow the xLLM commit format: <type>: <subject>.

Allowed types: feat, bugfix, docs, test, refactor, chore, style, revert, perf, model, build, release.
The subject should use clear English, start with a verb, include at least 4 words, and end with ..

Pre-commit Checks

• I have installed pre-commit by running pip install pre-commit or an equivalent command.
• I have installed the hooks with pre-commit install.
• I have run pre-commit run --all-files and fixed any reported issues.

If you are unsure how to set up pre-commit, see the pre-commit documentation.

Self Review

• I have self-reviewed the code according to .agents/skills/code-review/references/custom-code-style.md, especially code written or assisted by AI.
• I have rebased this PR onto the latest main branch.

Build and Test Coverage

• Tests have been added or updated as needed.
• CUDA: python setup.py build test has passed on a CUDA machine.
• NPU: python setup.py build test has passed on an NPU machine.
• MLU: python setup.py build test has passed on an MLU machine.

Reviewer Notes

[Dragonliu2018]

Please review this PR when you have a chance. Thanks!
@yq33victor @liutongxuan @walsonyang @RobbieLeung @JimHsiung @DongheJin @yingxudeng @Kang-Meng @DragonFive @Clement-Wang26 @zhang-minchao @xiao-yu-chen @liujinguang0125