fix(metax): fall back to Triton cache ops when vllm C extensions are missing

When vllm is built with VLLM_TARGET_DEVICE=empty, torch.ops._C_cache_ops
is not registered. The MetaX flash attention backend calls
vllm._custom_ops.reshape_and_cache_flash which delegates to that C
extension — causing AttributeError at runtime.

vllm already ships a pure-Triton implementation
(triton_reshape_and_cache_flash) for platforms without compiled
extensions. This patch routes _custom_ops through the Triton fallback
when _C_cache_ops is absent.

The patch only fires when _C_cache_ops is missing — standard vllm
wheels with compiled C extensions are unaffected.

Note: _C activation ops (silu_and_mul etc.) do NOT need stubs here.
The MetaX vendor backend (silu_and_mul_maca) already bypasses vllm's
SiluAndMul/GeluAndMul classes and computes the activations directly
with F.silu/F.gelu — no dependency on torch.ops._C.

Tested on MetaX C550 (MACA 3.7.2.0, torch 2.8.0+metax):
vllm 0.20.2 empty + vllm-plugin-FL + Qwen3-4B, eager + flash attn,
inference succeeds.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: tengqm

vllm_fl/dispatch/backends/vendor/metax/patches/__init__.py

@@ -29,3 +29,27 @@
 # TODO: remove when MetaX Triton support is available.
 import vllm.utils.import_utils as iu
 iu.has_triton_kernels = lambda: False
+
+# --------------------------------------------------
+# When vllm is built with VLLM_TARGET_DEVICE=empty, the C extension modules
+# are not compiled.  torch.ops._C_cache_ops.* are not registered, so the
+# MetaX flash attention backend's reshape_and_cache_flash call path fails.
+#
+# vllm ships a pure-Triton implementation (triton_reshape_and_cache_flash)
+# for platforms lacking the compiled extension.  Route _custom_ops through
+# the Triton fallback when the C ops are missing.
+#
+# Safe for standard vllm wheels: the patch only fires when _C_cache_ops is
+# absent, and every op in the dispatch chain is covered by the Triton kernel.
+import torch
+
+if not hasattr(torch.ops, "_C_cache_ops"):
+    import vllm._custom_ops as _custom_ops
+    from vllm.v1.attention.ops.triton_reshape_and_cache_flash import (
+        triton_reshape_and_cache_flash,
+        triton_reshape_and_cache_flash_per_token_head_quant,
+    )
+    _custom_ops.reshape_and_cache_flash = triton_reshape_and_cache_flash
+    _custom_ops.reshape_and_cache_flash_per_token_head_quant = (
+        triton_reshape_and_cache_flash_per_token_head_quant
+    )