fix(metax): avoid vllm._C dependency in activation ops for empty builds

[tengqm]

Problem

When vllm is built with VLLM_TARGET_DEVICE=empty, vllm._C is not compiled and torch.ops._C.silu_and_mul / torch.ops._C.gelu_and_mul do not exist. The MetaX activation functions instantiate SiluAndMul() / GeluAndMul(), whose init accesses torch.ops._C.* — causing AttributeError at model load time.

Root cause

silu_and_mul_maca and gelu_and_mul_maca delegate to vllm's compiled CUDA kernel classes (SiluAndMul, GeluAndMul). These depend on vllm._C.abi3.so which is absent in empty builds.

Fix

Replace with pure PyTorch: F.silu / F.gelu — functionally identical and works on both empty and full vllm builds.

Credit

Originally authored by ftgreat ldwang@baai.ac.cn in commit 1357ddf (2026-07-21) on the MetaX CI node. This PR upstreams that fix.

Testing

Verified on MetaX C550 (MACA 3.7.2.0, torch 2.8.0+metax): vllm 0.20.2 empty + vllm-plugin-FL, Qwen3-4B, eager mode, flash attention, inference succeeds without _C stubs.

[tengqm]

Error without this fix

Full stack trace from MetaX C550 (MACA 3.7.2.0, vllm 0.20.2 built with VLLM_TARGET_DEVICE=empty):

vllm_fl/ops/activation.py,: cannot open vllm_fl/ops/activation.py,' (No such file or directory) line: cannot open line' (No such file or directory)
10,: cannot open 10,' (No such file or directory) in: cannot open in' (No such file or directory)
init: cannot open `init' (No such file or directory)

Why dispatch doesn't help here

The crash happens in __init__, not in forward. vllm's SiluAndMul.__init__ unconditionally accesses torch.ops._C.silu_and_mul during model construction — before any inference, before any dispatch.

The dispatch layer (forward_oot → call_op → flagos/vendor/reference) only takes effect at inference time in SiluAndMulFL.forward_oot. But we never get there — __init__ crashes first.

__init__ is called during model loading by the vllm core model builder:

The fix replaces SiluAndMul() / GeluAndMul() with F.silu / F.gelu — no class instantiation, no __init__ dependency on vllm._C.

[tengqm]

Error without this fix

Full stack trace from MetaX C550 (MACA 3.7.2.0, vllm 0.20.2 built with VLLM_TARGET_DEVICE=empty):

File "vllm_fl/ops/activation.py", line 10, in __init__
  class SiluAndMulFL(SiluAndMul):
      def __init__(self):
          super().__init__()  ← line 10

File "vllm/model_executor/layers/activation.py", line 133, in __init__
  class SiluAndMul(CustomOp):
      def __init__(self, *, compile_native: bool = True):
          super().__init__(compile_native=compile_native)
          if current_platform.is_cuda_alike() or current_platform.is_xpu():
              self.op = torch.ops._C.silu_and_mul  ← line 133, CRASH

File "torch/_ops.py", line 1353, in __getattr__
  raise AttributeError(
      f"'_OpNamespace' '_C' object has no attribute 'silu_and_mul'")

Why dispatch doesn't help here

The crash happens in __init__, not in forward. vllm's SiluAndMul.__init__ unconditionally accesses torch.ops._C.silu_and_mul during model construction — before any inference, before any dispatch.

The dispatch layer (forward_oot -> call_op -> flagos/vendor/reference) only takes effect at inference time in SiluAndMulFL.forward_oot. But we never get there — __init__ crashes first.

__init__ is called during model loading by the vllm core model builder:

qwen3.py:204 Qwen3DecoderLayer.__init__()
  -> qwen2.py:111 init via lambda
    -> vllm_fl/ops/activation.py:10 SiluAndMulFL.__init__()
      -> super().__init__()
        -> activation.py:133 self.op = torch.ops._C.silu_and_mul

The fix replaces SiluAndMul() / GeluAndMul() with F.silu / F.gelu — no class instantiation, no __init__ dependency on vllm._C.

[tengqm]

This PR is not ready. Need further investigation.

[tengqm]

This PR was generated based on an experiment using cpu version of vllm. So the conclusion and the proposed fixes are both wrong.