support Qwen3.5 FP8 Model on GCU (Enflame)

vllm_fl/dispatch/backends/vendor/gcu/gcu.py

@@ -73,15 +73,11 @@ def rotary_embedding(
         )
 
     def attention_backend(self, use_mla: bool = False, use_sparse: bool = False) -> str:
-        from vllm.v1.attention.backends.registry import AttentionBackendEnum
-
         if use_mla:
             if use_sparse:
                 raise NotImplementedError("GCU does not support sparse attention yet")
             raise NotImplementedError("GCU does not support MLA yet")
-
-        import flash_attn.vllm_flash_attn
-
-        sys.modules["vllm.vllm_flash_attn"] = flash_attn.vllm_flash_attn
-
-        return AttentionBackendEnum.FLASH_ATTN.get_path()
+        # GCU uses a standalone flash_attn backend (AttentionGCUBackend) that calls
+        # Enflame's native flash_attn_varlen_func directly, without depending on
+        # vllm upstream FlashAttentionBackend / FlashAttentionImpl.
+        return "vllm_fl.dispatch.backends.vendor.gcu.impl.attention.AttentionGCUBackend"

vllm_fl/dispatch/backends/vendor/gcu/impl/apply_moe_activation_gcu.py

@@ -0,0 +1,75 @@
+# Copyright (c) 2026 BAAI. All rights reserved.
+
+"""GCU fix for ``apply_moe_activation``.
+
+The upstream ``apply_moe_activation`` directly calls ``torch.ops._C.*`` custom
+ops (e.g. ``silu_and_mul``, ``gelu_and_mul``) that are CUDA-only.  On GCU
+these ops are not available.
+
+The vllm_fl variant (:func:`vllm_fl.ops.fused_moe.activation.apply_moe_activation`)
+uses ``CachedOp("silu_and_mul")`` / ``CachedOp("gelu_and_mul")``, which
+dispatches through the FlagOS system and selects a GCU-compatible
+implementation.
+"""
+
+from __future__ import annotations
+
+import logging
+
+import torch
+
+logger = logging.getLogger(__name__)
+
+_patched = False
+
+
+def apply_moe_activation_gcu(
+    activation,
+    output: torch.Tensor,
+    input: torch.Tensor,
+) -> torch.Tensor:
+    """GCU-compatible ``apply_moe_activation`` via the FlagOS dispatch system."""
+    from vllm_fl.ops.fused_moe.activation import apply_moe_activation as _fl_impl
+
+    return _fl_impl(activation, output, input)
+
+
+def apply_moe_activation_gcu_patch() -> None:
+    """Patch ``apply_moe_activation`` for GCU devices."""
+    global _patched
+    if _patched:
+        return
+
+    gcu = getattr(torch, "gcu", None)
+    if gcu is None or not gcu.is_available():
+        return
+
+    # Modules that hold a reference to apply_moe_activation via
+    # ``from ...activation import apply_moe_activation``.
+    _IMPORTERS: list[str] = [
+        "vllm.model_executor.layers.fused_moe.activation",
+        "vllm.model_executor.layers.fused_moe.modular_kernel",
+        "vllm.model_executor.layers.fused_moe.fused_moe",
+        "vllm.model_executor.layers.fused_moe.experts.cutlass_moe",
+        "vllm.model_executor.layers.fused_moe.fused_marlin_moe",
+        "vllm.model_executor.layers.quantization.gguf",
+    ]
+
+    try:
+        for module_name in _IMPORTERS:
+            try:
+                mod = __import__(module_name, fromlist=["apply_moe_activation"])
+            except ImportError:
+                continue
+            if hasattr(mod, "apply_moe_activation"):
+                mod.apply_moe_activation = apply_moe_activation_gcu
+
+        _patched = True
+        logger.info(
+            "Patched apply_moe_activation for GCU (using FlagOS dispatch)"
+        )
+    except Exception as exc:
+        logger.warning(
+            "Failed to patch apply_moe_activation for GCU: %s",
+            exc,
+        )