【FR】Bug: vllm-plugin-FL error on ascend (google/gemma-3-1b-it)

[Lxiparer]

Bug Report: Bug: vllm-plugin-FL error: vllm-plugin-FL on ascend (google/gemma-3-1b-it)

Description

vllm-plugin-FL error encountered on model google/gemma-3-1b-it. Related components: vllm-plugin-FL.

Environment

Item	Value
Hardware	ascend
GPU	Ascend910
PyTorch	2.8.0+cpu
FlagGems	5.0.2
vLLM	0.13.0

Steps to Reproduce

1. Set up environment on ascend
2. Install dependencies (PyTorch + Triton + FlagGems + vLLM)
3. Install vllm-plugin-FL (git clone + pip install --no-build-isolation)
4. Start vLLM inference service with plugin enabled
5. Observe plugin-related errors

Expected Behavior

vllm-plugin-FL should load and dispatch operators without errors.

Actual Behavior

vllm-plugin-FL errors:

• (see error logs)

Error Logs

... (truncated) ...
21:22 [__init__.py:217] Platform plugin fl is activated
Traceback (most recent call last):
  File "/usr/local/python3.11.14/bin/vllm", line 3, in <module>
    from vllm.entrypoints.cli.main import main
  File "/vllm-workspace/vllm/vllm/entrypoints/cli/__init__.py", line 3, in <module>
    from vllm.entrypoints.cli.benchmark.latency import BenchmarkLatencySubcommand
  File "/vllm-workspace/vllm/vllm/entrypoints/cli/benchmark/latency.py", line 5, in <module>
    from vllm.benchmarks.latency import add_cli_args, main
  File "/vllm-workspace/vllm/vllm/benchmarks/latency.py", line 16, in <module>
    from vllm.engine.arg_utils import EngineArgs
  File "/vllm-workspace/vllm/vllm/engine/arg_utils.py", line 36, in <module>
    from vllm.config import (
  File "/vllm-workspace/vllm/vllm/config/__init__.py", line 6, in <module>
    from vllm.config.compilation import (
  File "/vllm-workspace/vllm/vllm/config/compilation.py", line 23, in <module>
    from vllm.platforms import current_platform
  File "/vllm-workspace/vllm/vllm/platforms/__init__.py", line 257, in __getattr__
    _current_platform = resolve_obj_by_qualname(platform_cls_qualname)()
                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/vllm-workspace/vllm/vllm/utils/import_utils.py", line 122, in resolve_obj_by_qualname
    module = importlib.import_module(module_name)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/python3.11.14/lib/python3.11/importlib/__init__.py", line 126, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/python3.11.14/lib/python3.11/site-packages/vllm_fl/platform.py", line 23, in <module>
    from vllm.v1.attention.backends.registry import AttentionBackendEnum
ModuleNotFoundError: No module named 'vllm.v1.attention.backends.registry'
[ERROR] 2026-05-13-20:21:24 (PID:563, Device:-1, RankID:-1) ERR99999 UNKNOWN ap

FlagGems Integration Code

File: /vllm-workspace/vllm-ascend/vllm_ascend/worker/model_runner_v1.py
Enable call: flag_gems.enable(record=True, once=True, path="/root/gems.txt")

Current gems.txt (7 operators, path: /root/gems.txt):

[DEBUG] flag_gems.runtime._ascend.ops.bmm.bmm: GEMS_ASCEND BMM
[DEBUG] flag_gems.runtime._ascend.ops.amax.amax: GEMS_ASCEND AMAX
[DEBUG] flag_gems.runtime._ascend.ops.ones.ones: GEMS_ASCEND ONES
[DEBUG] flag_gems.runtime._ascend.ops.zeros.zeros: GEMS_ASCEND ZEROS
[DEBUG] flag_gems.runtime._ascend.ops.arange.arange_start: GEMS_ASCEND ARANGE
[DEBUG] flag_gems.runtime._ascend.ops.index_select.index_select: GEMS_ASCEND INDEX SELECT
[DEBUG] flag_gems.runtime._ascend.ops.index.index: GEMS_ASCEND INDEX

Additional Context

• The issue appears to be specific to ascend platform and may not reproduce on standard CUDA (NVIDIA) environments.

Possible Directions

• Verify vllm-plugin-FL version compatibility with vLLM and FlagGems
• Check OpManager dispatch configuration and operator registration
• Try reinstalling plugin with --no-build-isolation flag